@staticn0va/wigolo 0.1.1 → 0.1.2

package/dist/integrations/cloud/llm/text-adapters.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../../src/integrations/cloud/llm/text-adapters.ts"],"sourcesContent":["// Free-form text completion adapters per provider. Returns plain text\n// (markdown / prose) without a JSON schema constraint — used by research +\n// agent synthesis. The JSON-schema adapters in anthropic.ts/openai.ts/etc.\n// stay for extract's structured path.\n//\n// SDKs are imported lazily inside each adapter so module-load is cheap;\n// otherwise pulling in all four cloud SDKs at boot adds hundreds of ms to\n// MCP server startup (caught by the cold-start e2e timing test).\n\nimport type { LLMProvider } from './types.js';\n\nexport interface TextCallOpts {\n  prompt: string;\n  model: string;\n  maxTokens?: number;\n  signal?: AbortSignal;\n}\n\nexport interface TextCallResult {\n  text: string;\n  provider: LLMProvider;\n  model: string;\n  latencyMs: number;\n}\n\nconst DEFAULT_MAX_TOKENS = 2000;\n\nexport async function callAnthropicText(opts: TextCallOpts, apiKey: string): Promise<TextCallResult> {\n  const { default: Anthropic } = await import('@anthropic-ai/sdk');\n  const client = new Anthropic({ apiKey });\n  const start = Date.now();\n  const response = await client.messages.create(\n    {\n      model: opts.model,\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: [{ role: 'user', content: opts.prompt }],\n    },\n    { signal: opts.signal },\n  );\n  const block = (response.content ?? []).find((b: { type: string }) => b.type === 'text') as\n    | { type: 'text'; text: string }\n    | undefined;\n  if (!block) throw new Error('anthropic: no text block in response');\n  return {\n    text: block.text,\n    provider: 'anthropic',\n    model: response.model ?? opts.model,\n    latencyMs: Date.now() - start,\n  };\n}\n\nexport async function callOpenAIText(opts: TextCallOpts, apiKey: string): Promise<TextCallResult> {\n  const { default: OpenAI } = await import('openai');\n  const client = new OpenAI({ apiKey });\n  const start = Date.now();\n  const response = await client.chat.completions.create(\n    {\n      model: opts.model,\n      max_completion_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: [{ role: 'user', content: opts.prompt }],\n    },\n    { signal: opts.signal },\n  );\n  const text = response.choices?.[0]?.message?.content;\n  if (typeof text !== 'string' || text.trim().length === 0) {\n    throw new Error('openai: empty content in response');\n  }\n  return {\n    text,\n    provider: 'openai',\n    model: response.model ?? opts.model,\n    latencyMs: Date.now() - start,\n  };\n}\n\nexport async function callGeminiText(opts: TextCallOpts, apiKey: string): Promise<TextCallResult> {\n  const { GoogleGenAI } = await import('@google/genai');\n  const client = new GoogleGenAI({ apiKey });\n  const start = Date.now();\n  const response = await client.models.generateContent({\n    model: opts.model,\n    contents: opts.prompt,\n    config: {\n      maxOutputTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      abortSignal: opts.signal,\n    },\n  });\n  const text = response.text;\n  if (!text || text.trim().length === 0) throw new Error('gemini: empty text in response');\n  return {\n    text,\n    provider: 'gemini',\n    model: opts.model,\n    latencyMs: Date.now() - start,\n  };\n}\n\nexport async function callGroqText(opts: TextCallOpts, apiKey: string): Promise<TextCallResult> {\n  const { default: Groq } = await import('groq-sdk');\n  const client = new Groq({ apiKey });\n  const start = Date.now();\n  const response = await client.chat.completions.create(\n    {\n      model: opts.model,\n      max_completion_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      messages: [{ role: 'user', content: opts.prompt }],\n    },\n    { signal: opts.signal },\n  );\n  const text = response.choices?.[0]?.message?.content;\n  if (typeof text !== 'string' || text.trim().length === 0) {\n    throw new Error('groq: empty content in response');\n  }\n  return {\n    text,\n    provider: 'groq',\n    model: response.model ?? opts.model,\n    latencyMs: Date.now() - start,\n  };\n}\n\nexport const TEXT_ADAPTERS: Record<\n  LLMProvider,\n  (opts: TextCallOpts, apiKey: string) => Promise<TextCallResult>\n> = {\n  anthropic: callAnthropicText,\n  openai: callOpenAIText,\n  gemini: callGeminiText,\n  groq: callGroqText,\n};\n"],"mappings":"AAyBA,MAAM,qBAAqB;AAE3B,eAAsB,kBAAkB,MAAoB,QAAyC;AACnG,QAAM,EAAE,SAAS,UAAU,IAAI,MAAM,OAAO,mBAAmB;AAC/D,QAAM,SAAS,IAAI,UAAU,EAAE,OAAO,CAAC;AACvC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,WAAW,MAAM,OAAO,SAAS;AAAA,IACrC;AAAA,MACE,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK,aAAa;AAAA,MAC9B,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,KAAK,OAAO,CAAC;AAAA,IACnD;AAAA,IACA,EAAE,QAAQ,KAAK,OAAO;AAAA,EACxB;AACA,QAAM,SAAS,SAAS,WAAW,CAAC,GAAG,KAAK,CAAC,MAAwB,EAAE,SAAS,MAAM;AAGtF,MAAI,CAAC,MAAO,OAAM,IAAI,MAAM,sCAAsC;AAClE,SAAO;AAAA,IACL,MAAM,MAAM;AAAA,IACZ,UAAU;AAAA,IACV,OAAO,SAAS,SAAS,KAAK;AAAA,IAC9B,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACF;AAEA,eAAsB,eAAe,MAAoB,QAAyC;AAChG,QAAM,EAAE,SAAS,OAAO,IAAI,MAAM,OAAO,QAAQ;AACjD,QAAM,SAAS,IAAI,OAAO,EAAE,OAAO,CAAC;AACpC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,WAAW,MAAM,OAAO,KAAK,YAAY;AAAA,IAC7C;AAAA,MACE,OAAO,KAAK;AAAA,MACZ,uBAAuB,KAAK,aAAa;AAAA,MACzC,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,KAAK,OAAO,CAAC;AAAA,IACnD;AAAA,IACA,EAAE,QAAQ,KAAK,OAAO;AAAA,EACxB;AACA,QAAM,OAAO,SAAS,UAAU,CAAC,GAAG,SAAS;AAC7C,MAAI,OAAO,SAAS,YAAY,KAAK,KAAK,EAAE,WAAW,GAAG;AACxD,UAAM,IAAI,MAAM,mCAAmC;AAAA,EACrD;AACA,SAAO;AAAA,IACL;AAAA,IACA,UAAU;AAAA,IACV,OAAO,SAAS,SAAS,KAAK;AAAA,IAC9B,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACF;AAEA,eAAsB,eAAe,MAAoB,QAAyC;AAChG,QAAM,EAAE,YAAY,IAAI,MAAM,OAAO,eAAe;AACpD,QAAM,SAAS,IAAI,YAAY,EAAE,OAAO,CAAC;AACzC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,WAAW,MAAM,OAAO,OAAO,gBAAgB;AAAA,IACnD,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,IACf,QAAQ;AAAA,MACN,iBAAiB,KAAK,aAAa;AAAA,MACnC,aAAa,KAAK;AAAA,IACpB;AAAA,EACF,CAAC;AACD,QAAM,OAAO,SAAS;AACtB,MAAI,CAAC,QAAQ,KAAK,KAAK,EAAE,WAAW,EAAG,OAAM,IAAI,MAAM,gCAAgC;AACvF,SAAO;AAAA,IACL;AAAA,IACA,UAAU;AAAA,IACV,OAAO,KAAK;AAAA,IACZ,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACF;AAEA,eAAsB,aAAa,MAAoB,QAAyC;AAC9F,QAAM,EAAE,SAAS,KAAK,IAAI,MAAM,OAAO,UAAU;AACjD,QAAM,SAAS,IAAI,KAAK,EAAE,OAAO,CAAC;AAClC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,WAAW,MAAM,OAAO,KAAK,YAAY;AAAA,IAC7C;AAAA,MACE,OAAO,KAAK;AAAA,MACZ,uBAAuB,KAAK,aAAa;AAAA,MACzC,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,KAAK,OAAO,CAAC;AAAA,IACnD;AAAA,IACA,EAAE,QAAQ,KAAK,OAAO;AAAA,EACxB;AACA,QAAM,OAAO,SAAS,UAAU,CAAC,GAAG,SAAS;AAC7C,MAAI,OAAO,SAAS,YAAY,KAAK,KAAK,EAAE,WAAW,GAAG;AACxD,UAAM,IAAI,MAAM,iCAAiC;AAAA,EACnD;AACA,SAAO;AAAA,IACL;AAAA,IACA,UAAU;AAAA,IACV,OAAO,SAAS,SAAS,KAAK;AAAA,IAC9B,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACF;AAEO,MAAM,gBAGT;AAAA,EACF,WAAW;AAAA,EACX,QAAQ;AAAA,EACR,QAAQ;AAAA,EACR,MAAM;AACR;","names":[]}

package/dist/providers/rerank-provider.d.ts

@@ -13,4 +13,5 @@ export interface RerankProvider {
 }
 export declare function getRerankProvider(): Promise<RerankProvider>;
 export declare function _resetRerankProviderForTest(): void;
+export declare function disposeRerankProvider(): Promise<void>;
 //# sourceMappingURL=rerank-provider.d.ts.map

package/dist/providers/rerank-provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"rerank-provider.d.ts","sourceRoot":"","sources":["../../src/providers/rerank-provider.ts"],"names":[],"mappings":"AAWA,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,CACJ,KAAK,EAAE,MAAM,EACb,UAAU,EAAE,eAAe,EAAE,EAC7B,IAAI,CAAC,EAAE,MAAM,GACZ,OAAO,CAAC,YAAY,EAAE,CAAC,CAAC;IAC3B,8DAA8D;IAC9D,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAID,wBAAgB,iBAAiB,IAAI,OAAO,CAAC,cAAc,CAAC,CAkB3D;AAED,wBAAgB,2BAA2B,IAAI,IAAI,CAElD"}
+{"version":3,"file":"rerank-provider.d.ts","sourceRoot":"","sources":["../../src/providers/rerank-provider.ts"],"names":[],"mappings":"AAWA,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;CACd;AAED,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,cAAc;IAC7B,MAAM,CACJ,KAAK,EAAE,MAAM,EACb,UAAU,EAAE,eAAe,EAAE,EAC7B,IAAI,CAAC,EAAE,MAAM,GACZ,OAAO,CAAC,YAAY,EAAE,CAAC,CAAC;IAC3B,8DAA8D;IAC9D,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;CAC1B;AAID,wBAAgB,iBAAiB,IAAI,OAAO,CAAC,cAAc,CAAC,CAkB3D;AAED,wBAAgB,2BAA2B,IAAI,IAAI,CAElD;AAID,wBAAsB,qBAAqB,IAAI,OAAO,CAAC,IAAI,CAAC,CAW3D"}

package/dist/providers/rerank-provider.js

@@ -21,8 +21,21 @@ function getRerankProvider() {
 function _resetRerankProviderForTest() {
   cached = null;
 }
+async function disposeRerankProvider() {
+  if (!cached) return;
+  try {
+    const provider = await cached;
+    const disposable = provider;
+    if (typeof disposable.dispose === "function") await disposable.dispose();
+  } catch (err) {
+    log.debug("rerank dispose failed", { error: err instanceof Error ? err.message : String(err) });
+  } finally {
+    cached = null;
+  }
+}
 export {
   _resetRerankProviderForTest,
+  disposeRerankProvider,
   getRerankProvider
 };
 //# sourceMappingURL=rerank-provider.js.map

package/dist/providers/rerank-provider.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/providers/rerank-provider.ts"],"sourcesContent":["/**\n * Rerank provider interface — Phase 1 Task 1.3 of v1 engine overhaul.\n *\n * Phase 4 Part A switches the default factory to TransformersRerankProvider\n * (Transformers.js cross-encoder, in-process ONNX runtime). The legacy\n * Python FlashRank adapter still exists in `search/reranker/legacy-provider.ts`\n * pending Phase 4 Part B deletions, but it is no longer wired in.\n */\nimport { createLogger } from '../logger.js';\n\nconst log = createLogger('providers');\nexport interface RerankCandidate {\n  id: string;\n  text: string;\n}\n\nexport interface RerankResult {\n  id: string;\n  score: number;\n}\n\nexport interface RerankProvider {\n  rerank(\n    query: string,\n    candidates: RerankCandidate[],\n    topK?: number,\n  ): Promise<RerankResult[]>;\n  /** Model identifier (for cache invalidation / provenance). */\n  readonly modelId: string;\n}\n\nlet cached: Promise<RerankProvider> | null = null;\n\nexport function getRerankProvider(): Promise<RerankProvider> {\n  if (cached) return cached;\n  cached = import('../search/reranker/transformers-rerank-provider.js')\n    .then(async (m) => {\n      const p = new m.TransformersRerankProvider();\n      await p.warmup();\n      log.info('rerank provider ready', {\n        provider: 'rerank',\n        impl: 'transformers',\n        modelId: p.modelId,\n      });\n      return p;\n    })\n    .catch((err) => {\n      cached = null;\n      throw err;\n    });\n  return cached;\n}\n\nexport function _resetRerankProviderForTest(): void {\n  cached = null;\n}\n"],"mappings":"AAQA,SAAS,oBAAoB;AAE7B,MAAM,MAAM,aAAa,WAAW;AAqBpC,IAAI,SAAyC;AAEtC,SAAS,oBAA6C;AAC3D,MAAI,OAAQ,QAAO;AACnB,WAAS,OAAO,oDAAoD,EACjE,KAAK,OAAO,MAAM;AACjB,UAAM,IAAI,IAAI,EAAE,2BAA2B;AAC3C,UAAM,EAAE,OAAO;AACf,QAAI,KAAK,yBAAyB;AAAA,MAChC,UAAU;AAAA,MACV,MAAM;AAAA,MACN,SAAS,EAAE;AAAA,IACb,CAAC;AACD,WAAO;AAAA,EACT,CAAC,EACA,MAAM,CAAC,QAAQ;AACd,aAAS;AACT,UAAM;AAAA,EACR,CAAC;AACH,SAAO;AACT;AAEO,SAAS,8BAAoC;AAClD,WAAS;AACX;","names":[]}
+{"version":3,"sources":["../../src/providers/rerank-provider.ts"],"sourcesContent":["/**\n * Rerank provider interface — Phase 1 Task 1.3 of v1 engine overhaul.\n *\n * Phase 4 Part A switches the default factory to TransformersRerankProvider\n * (Transformers.js cross-encoder, in-process ONNX runtime). The legacy\n * Python FlashRank adapter still exists in `search/reranker/legacy-provider.ts`\n * pending Phase 4 Part B deletions, but it is no longer wired in.\n */\nimport { createLogger } from '../logger.js';\n\nconst log = createLogger('providers');\nexport interface RerankCandidate {\n  id: string;\n  text: string;\n}\n\nexport interface RerankResult {\n  id: string;\n  score: number;\n}\n\nexport interface RerankProvider {\n  rerank(\n    query: string,\n    candidates: RerankCandidate[],\n    topK?: number,\n  ): Promise<RerankResult[]>;\n  /** Model identifier (for cache invalidation / provenance). */\n  readonly modelId: string;\n}\n\nlet cached: Promise<RerankProvider> | null = null;\n\nexport function getRerankProvider(): Promise<RerankProvider> {\n  if (cached) return cached;\n  cached = import('../search/reranker/transformers-rerank-provider.js')\n    .then(async (m) => {\n      const p = new m.TransformersRerankProvider();\n      await p.warmup();\n      log.info('rerank provider ready', {\n        provider: 'rerank',\n        impl: 'transformers',\n        modelId: p.modelId,\n      });\n      return p;\n    })\n    .catch((err) => {\n      cached = null;\n      throw err;\n    });\n  return cached;\n}\n\nexport function _resetRerankProviderForTest(): void {\n  cached = null;\n}\n\n// Best-effort disposal of the cached rerank provider's native resources.\n// Called from CLI shutdown to release the ONNX session before process exit.\nexport async function disposeRerankProvider(): Promise<void> {\n  if (!cached) return;\n  try {\n    const provider = await cached;\n    const disposable = provider as unknown as { dispose?: () => Promise<void> };\n    if (typeof disposable.dispose === 'function') await disposable.dispose();\n  } catch (err) {\n    log.debug('rerank dispose failed', { error: err instanceof Error ? err.message : String(err) });\n  } finally {\n    cached = null;\n  }\n}\n"],"mappings":"AAQA,SAAS,oBAAoB;AAE7B,MAAM,MAAM,aAAa,WAAW;AAqBpC,IAAI,SAAyC;AAEtC,SAAS,oBAA6C;AAC3D,MAAI,OAAQ,QAAO;AACnB,WAAS,OAAO,oDAAoD,EACjE,KAAK,OAAO,MAAM;AACjB,UAAM,IAAI,IAAI,EAAE,2BAA2B;AAC3C,UAAM,EAAE,OAAO;AACf,QAAI,KAAK,yBAAyB;AAAA,MAChC,UAAU;AAAA,MACV,MAAM;AAAA,MACN,SAAS,EAAE;AAAA,IACb,CAAC;AACD,WAAO;AAAA,EACT,CAAC,EACA,MAAM,CAAC,QAAQ;AACd,aAAS;AACT,UAAM;AAAA,EACR,CAAC;AACH,SAAO;AACT;AAEO,SAAS,8BAAoC;AAClD,WAAS;AACX;AAIA,eAAsB,wBAAuC;AAC3D,MAAI,CAAC,OAAQ;AACb,MAAI;AACF,UAAM,WAAW,MAAM;AACvB,UAAM,aAAa;AACnB,QAAI,OAAO,WAAW,YAAY,WAAY,OAAM,WAAW,QAAQ;AAAA,EACzE,SAAS,KAAK;AACZ,QAAI,MAAM,yBAAyB,EAAE,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAE,CAAC;AAAA,EAChG,UAAE;AACA,aAAS;AAAA,EACX;AACF;","names":[]}

package/dist/research/brief.d.ts

@@ -2,4 +2,5 @@ import type { ResearchBrief, ResearchSource, CrossReference } from '../types.js'
 import type { QueryType } from './decompose.js';
 export declare function buildResearchBrief(question: string, sources: ResearchSource[], subQueries: string[], perSourceCharCap: number, totalSourcesCharCap: number, queryType?: QueryType, comparisonEntities?: string[], synthesisText?: string): Promise<ResearchBrief>;
 export declare function detectCrossReferences(sources: ResearchSource[]): CrossReference[];
+export declare function stripMarkdownLinks(text: string): string;
 //# sourceMappingURL=brief.d.ts.map

package/dist/research/brief.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"brief.d.ts","sourceRoot":"","sources":["../../src/research/brief.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,cAAc,EAAoB,cAAc,EAAE,MAAM,aAAa,CAAC;AACnG,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAahD,wBAAsB,kBAAkB,CACtC,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE,cAAc,EAAE,EACzB,UAAU,EAAE,MAAM,EAAE,EACpB,gBAAgB,EAAE,MAAM,EACxB,mBAAmB,EAAE,MAAM,EAC3B,SAAS,GAAE,SAAqB,EAChC,kBAAkB,GAAE,MAAM,EAAO,EACjC,aAAa,CAAC,EAAE,MAAM,GACrB,OAAO,CAAC,aAAa,CAAC,CAgDxB;AA6BD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,cAAc,EAAE,GAAG,cAAc,EAAE,CAuCjF"}
+{"version":3,"file":"brief.d.ts","sourceRoot":"","sources":["../../src/research/brief.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,cAAc,EAAoB,cAAc,EAAE,MAAM,aAAa,CAAC;AACnG,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,gBAAgB,CAAC;AAahD,wBAAsB,kBAAkB,CACtC,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE,cAAc,EAAE,EACzB,UAAU,EAAE,MAAM,EAAE,EACpB,gBAAgB,EAAE,MAAM,EACxB,mBAAmB,EAAE,MAAM,EAC3B,SAAS,GAAE,SAAqB,EAChC,kBAAkB,GAAE,MAAM,EAAO,EACjC,aAAa,CAAC,EAAE,MAAM,GACrB,OAAO,CAAC,aAAa,CAAC,CAgDxB;AA6BD,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,cAAc,EAAE,GAAG,cAAc,EAAE,CAuCjF;AA6FD,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAOvD"}

package/dist/research/brief.js

@@ -156,12 +156,15 @@ function firstSubstantiveParagraph(markdown) {
   for (const p of paragraphs) {
     if (p.length < 80) continue;
     if (p.startsWith("#") || p.startsWith("|") || p.startsWith("```")) continue;
-    const stripped = p.replace(/^!\[[^\]]*\]\([^)]*\)\s*/g, "").replace(/^\[!\[[^\]]*\]\([^)]*\)\]\([^)]*\)\s*/g, "").trim();
-    if (stripped.length < 80) continue;
-    return stripped.replace(/\s+/g, " ");
+    const cleaned = stripMarkdownLinks(p);
+    if (cleaned.length < 80) continue;
+    return cleaned.replace(/\s+/g, " ");
   }
   return null;
 }
+function stripMarkdownLinks(text) {
+  return text.replace(/!\[[^\]]*\]\([^)]*\)/g, "").replace(/\[!\[[^\]]*\]\([^)]*\)\]\([^)]*\)/g, "").replace(/\[([^\]]+)\]\([^)]*\)/g, "$1").replace(/<https?:\/\/[^>]+>/g, "").trim();
+}
 function dedupe(list) {
   const seen = /* @__PURE__ */ new Set();
   const out = [];
@@ -237,6 +240,7 @@ const STOP_WORDS = /* @__PURE__ */ new Set([
 ]);
 export {
   buildResearchBrief,
-  detectCrossReferences
+  detectCrossReferences,
+  stripMarkdownLinks
 };
 //# sourceMappingURL=brief.js.map

package/dist/research/brief.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/research/brief.ts"],"sourcesContent":["import type { ResearchBrief, ResearchSource, SearchResultItem, CrossReference } from '../types.js';\nimport type { QueryType } from './decompose.js';\nimport { extractHighlights } from '../search/highlights.js';\nimport { buildCitationGraph } from './citation-graph.js';\n\nconst MAX_HIGHLIGHTS = 12;\nconst MAX_KEY_FINDING_LEN = 280;\nconst MAX_TOPICS = 8;\nconst MAX_CROSS_REFS = 10;\nconst MIN_PHRASE_LEN = 4;\n\n// Build a host-LLM-friendly structured brief when internal sampling is\n// unavailable. The host model (Claude Code / Cursor / etc.) consumes this\n// shape to produce the final report without needing to re-read raw sources.\nexport async function buildResearchBrief(\n  question: string,\n  sources: ResearchSource[],\n  subQueries: string[],\n  perSourceCharCap: number,\n  totalSourcesCharCap: number,\n  queryType: QueryType = 'general',\n  comparisonEntities: string[] = [],\n  synthesisText?: string,\n): Promise<ResearchBrief> {\n  const fetched = sources.filter((s) => s.fetched && s.markdown_content.length > 0);\n\n  // Highlights reuse the ONNX-reranker-or-paragraph scorer so briefs align with\n  // whatever format='highlights' produces for single-query searches.\n  const searchItems: SearchResultItem[] = fetched.map((s) => ({\n    title: s.title,\n    url: s.url,\n    snippet: s.markdown_content.slice(0, 200),\n    markdown_content: s.markdown_content,\n    relevance_score: s.relevance_score,\n  }));\n\n  const { highlights } = await extractHighlights(question, searchItems, MAX_HIGHLIGHTS);\n\n  const topics = buildTopics(subQueries, fetched);\n  const keyFindings = buildKeyFindings(fetched);\n  const crossReferences = detectCrossReferences(fetched);\n  const gaps = detectGaps(subQueries, fetched);\n\n  const comparison = queryType === 'comparison' && comparisonEntities.length >= 2\n    ? buildComparisonSection(comparisonEntities, fetched)\n    : undefined;\n\n  const citationGraph = synthesisText && synthesisText.trim().length > 0 && fetched.length > 0\n    ? buildCitationGraph(\n        synthesisText,\n        fetched.map((s) => ({ url: s.url, title: s.title, markdown: s.markdown_content })),\n      )\n    : undefined;\n\n  return {\n    topics,\n    highlights,\n    key_findings: keyFindings,\n    per_source_char_cap: perSourceCharCap,\n    total_sources_char_cap: totalSourcesCharCap,\n    sections: {\n      overview: {\n        key_findings: keyFindings.slice(0, 5),\n        cross_references: crossReferences,\n      },\n      ...(comparison ? { comparison } : {}),\n      gaps,\n    },\n    query_type: queryType,\n    ...(citationGraph && citationGraph.length > 0 ? { citation_graph: citationGraph } : {}),\n  };\n}\n\n// Prefer sub-queries (planner's view of the topic space) when available;\n// otherwise derive compact topic labels from source titles.\nfunction buildTopics(subQueries: string[], sources: ResearchSource[]): string[] {\n  if (subQueries.length > 0) {\n    return dedupe(subQueries).slice(0, MAX_TOPICS);\n  }\n  const labels = sources\n    .map((s) => s.title.split(/[–|:·-]/)[0].trim())\n    .filter((t) => t.length >= 5 && t.length <= 100);\n  return dedupe(labels).slice(0, MAX_TOPICS);\n}\n\n// First substantive paragraph per source, trimmed to a finding-sized blurb.\n// Ordered by source relevance so the most-weighted finding is first.\nfunction buildKeyFindings(sources: ResearchSource[]): string[] {\n  const out: string[] = [];\n  for (const s of [...sources].sort((a, b) => b.relevance_score - a.relevance_score)) {\n    const first = firstSubstantiveParagraph(s.markdown_content);\n    if (!first) continue;\n    const trimmed = first.length > MAX_KEY_FINDING_LEN\n      ? first.slice(0, MAX_KEY_FINDING_LEN - 1).trimEnd() + '…'\n      : first;\n    out.push(trimmed);\n  }\n  return dedupe(out);\n}\n\nexport function detectCrossReferences(sources: ResearchSource[]): CrossReference[] {\n  if (sources.length < 2) return [];\n\n  // Extract significant phrases from each source's content\n  const phraseMap = new Map<string, Set<number>>();\n\n  for (let idx = 0; idx < sources.length; idx++) {\n    const content = sources[idx].markdown_content.toLowerCase();\n    const words = content\n      .replace(/[^a-z0-9\\s]/g, ' ')\n      .split(/\\s+/)\n      .filter((w) => w.length >= MIN_PHRASE_LEN && !STOP_WORDS.has(w));\n\n    const seenForSource = new Set<string>();\n    for (let i = 0; i < words.length - 2; i++) {\n      const phrase = words.slice(i, i + 3).join(' ');\n      if (seenForSource.has(phrase)) continue;\n      seenForSource.add(phrase);\n\n      if (!phraseMap.has(phrase)) phraseMap.set(phrase, new Set());\n      phraseMap.get(phrase)!.add(idx);\n    }\n  }\n\n  // Phrases found in 2+ sources are cross-references\n  const candidates: CrossReference[] = [];\n  for (const [phrase, sourceIndices] of phraseMap) {\n    if (sourceIndices.size >= 2) {\n      candidates.push({\n        finding: phrase,\n        source_indices: [...sourceIndices].sort(),\n        confidence: sourceIndices.size >= 3 ? 'high' : 'medium',\n      });\n    }\n  }\n\n  // Sort by number of sources (desc), then deduplicate overlapping phrases\n  candidates.sort((a, b) => b.source_indices.length - a.source_indices.length);\n  return deduplicateOverlapping(candidates).slice(0, MAX_CROSS_REFS);\n}\n\nfunction deduplicateOverlapping(refs: CrossReference[]): CrossReference[] {\n  const kept: CrossReference[] = [];\n  const usedWords = new Set<string>();\n\n  for (const ref of refs) {\n    const words = ref.finding.split(' ');\n    // Skip if most words already covered by a higher-ranked cross-reference\n    const overlapCount = words.filter((w) => usedWords.has(w)).length;\n    if (overlapCount >= words.length - 1 && kept.length > 0) continue;\n\n    kept.push(ref);\n    for (const w of words) usedWords.add(w);\n  }\n\n  return kept;\n}\n\nfunction detectGaps(subQueries: string[], sources: ResearchSource[]): string[] {\n  if (subQueries.length === 0) return [];\n\n  const gaps: string[] = [];\n  const contentLower = sources.map((s) => s.markdown_content.toLowerCase()).join(' ');\n\n  for (const query of subQueries) {\n    // Extract significant words from sub-query\n    const words = query.toLowerCase()\n      .replace(/[^a-z0-9\\s]/g, ' ')\n      .split(/\\s+/)\n      .filter((w) => w.length >= MIN_PHRASE_LEN && !STOP_WORDS.has(w));\n\n    if (words.length === 0) continue;\n\n    // Count how many significant words appear in any source\n    const found = words.filter((w) => contentLower.includes(w)).length;\n    const coverage = found / words.length;\n\n    if (coverage < 0.5) {\n      gaps.push(`Limited coverage for: \"${query}\"`);\n    }\n  }\n\n  return gaps;\n}\n\nfunction buildComparisonSection(\n  entities: string[],\n  sources: ResearchSource[],\n): { entities: string[]; comparison_points: string[] } {\n  const comparisonPoints: string[] = [];\n  const contentLower = sources.map((s) => s.markdown_content.toLowerCase()).join('\\n');\n\n  // Look for comparison keywords near entity mentions\n  const comparisonTerms = ['faster', 'slower', 'better', 'worse', 'more', 'less',\n    'easier', 'harder', 'simpler', 'complex', 'lightweight', 'heavy',\n    'performance', 'scalability', 'ecosystem', 'community', 'support'];\n\n  for (const term of comparisonTerms) {\n    if (!contentLower.includes(term)) continue;\n\n    // Check if term appears near any entity\n    const nearEntity = entities.some((e) => {\n      const entityLower = e.toLowerCase();\n      const idx = contentLower.indexOf(entityLower);\n      if (idx === -1) return false;\n      // Check within 200 chars of entity mention\n      const neighborhood = contentLower.slice(Math.max(0, idx - 200), idx + e.length + 200);\n      return neighborhood.includes(term);\n    });\n\n    if (nearEntity) {\n      comparisonPoints.push(term);\n    }\n  }\n\n  return { entities, comparison_points: [...new Set(comparisonPoints)] };\n}\n\nfunction firstSubstantiveParagraph(markdown: string): string | null {\n  const paragraphs = markdown.split(/\\n\\n+/).map((p) => p.trim());\n  for (const p of paragraphs) {\n    if (p.length < 80) continue;\n    if (p.startsWith('#') || p.startsWith('|') || p.startsWith('```')) continue;\n    // Strip leading images / links-around-images that pad alt text into the\n    // paragraph; if nothing of substance remains, skip.\n    const stripped = p\n      .replace(/^!\\[[^\\]]*\\]\\([^)]*\\)\\s*/g, '')\n      .replace(/^\\[!\\[[^\\]]*\\]\\([^)]*\\)\\]\\([^)]*\\)\\s*/g, '')\n      .trim();\n    if (stripped.length < 80) continue;\n    return stripped.replace(/\\s+/g, ' ');\n  }\n  return null;\n}\n\nfunction dedupe(list: string[]): string[] {\n  const seen = new Set<string>();\n  const out: string[] = [];\n  for (const item of list) {\n    const key = item.toLowerCase();\n    if (seen.has(key)) continue;\n    seen.add(key);\n    out.push(item);\n  }\n  return out;\n}\n\nconst STOP_WORDS = new Set([\n  'about', 'after', 'also', 'been', 'before', 'being', 'between',\n  'both', 'could', 'does', 'doing', 'done', 'each', 'even', 'every',\n  'from', 'have', 'here', 'into', 'just', 'like', 'made', 'make',\n  'many', 'more', 'most', 'much', 'must', 'need', 'only', 'other',\n  'over', 'same', 'should', 'some', 'such', 'than', 'that', 'their',\n  'them', 'then', 'there', 'these', 'they', 'this', 'those', 'through',\n  'very', 'want', 'well', 'were', 'what', 'when', 'where', 'which',\n  'while', 'will', 'with', 'would', 'your',\n]);\n"],"mappings":"AAEA,SAAS,yBAAyB;AAClC,SAAS,0BAA0B;AAEnC,MAAM,iBAAiB;AACvB,MAAM,sBAAsB;AAC5B,MAAM,aAAa;AACnB,MAAM,iBAAiB;AACvB,MAAM,iBAAiB;AAKvB,eAAsB,mBACpB,UACA,SACA,YACA,kBACA,qBACA,YAAuB,WACvB,qBAA+B,CAAC,GAChC,eACwB;AACxB,QAAM,UAAU,QAAQ,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,iBAAiB,SAAS,CAAC;AAIhF,QAAM,cAAkC,QAAQ,IAAI,CAAC,OAAO;AAAA,IAC1D,OAAO,EAAE;AAAA,IACT,KAAK,EAAE;AAAA,IACP,SAAS,EAAE,iBAAiB,MAAM,GAAG,GAAG;AAAA,IACxC,kBAAkB,EAAE;AAAA,IACpB,iBAAiB,EAAE;AAAA,EACrB,EAAE;AAEF,QAAM,EAAE,WAAW,IAAI,MAAM,kBAAkB,UAAU,aAAa,cAAc;AAEpF,QAAM,SAAS,YAAY,YAAY,OAAO;AAC9C,QAAM,cAAc,iBAAiB,OAAO;AAC5C,QAAM,kBAAkB,sBAAsB,OAAO;AACrD,QAAM,OAAO,WAAW,YAAY,OAAO;AAE3C,QAAM,aAAa,cAAc,gBAAgB,mBAAmB,UAAU,IAC1E,uBAAuB,oBAAoB,OAAO,IAClD;AAEJ,QAAM,gBAAgB,iBAAiB,cAAc,KAAK,EAAE,SAAS,KAAK,QAAQ,SAAS,IACvF;AAAA,IACE;AAAA,IACA,QAAQ,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,KAAK,OAAO,EAAE,OAAO,UAAU,EAAE,iBAAiB,EAAE;AAAA,EACnF,IACA;AAEJ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd,qBAAqB;AAAA,IACrB,wBAAwB;AAAA,IACxB,UAAU;AAAA,MACR,UAAU;AAAA,QACR,cAAc,YAAY,MAAM,GAAG,CAAC;AAAA,QACpC,kBAAkB;AAAA,MACpB;AAAA,MACA,GAAI,aAAa,EAAE,WAAW,IAAI,CAAC;AAAA,MACnC;AAAA,IACF;AAAA,IACA,YAAY;AAAA,IACZ,GAAI,iBAAiB,cAAc,SAAS,IAAI,EAAE,gBAAgB,cAAc,IAAI,CAAC;AAAA,EACvF;AACF;AAIA,SAAS,YAAY,YAAsB,SAAqC;AAC9E,MAAI,WAAW,SAAS,GAAG;AACzB,WAAO,OAAO,UAAU,EAAE,MAAM,GAAG,UAAU;AAAA,EAC/C;AACA,QAAM,SAAS,QACZ,IAAI,CAAC,MAAM,EAAE,MAAM,MAAM,SAAS,EAAE,CAAC,EAAE,KAAK,CAAC,EAC7C,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;AACjD,SAAO,OAAO,MAAM,EAAE,MAAM,GAAG,UAAU;AAC3C;AAIA,SAAS,iBAAiB,SAAqC;AAC7D,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,CAAC,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,kBAAkB,EAAE,eAAe,GAAG;AAClF,UAAM,QAAQ,0BAA0B,EAAE,gBAAgB;AAC1D,QAAI,CAAC,MAAO;AACZ,UAAM,UAAU,MAAM,SAAS,sBAC3B,MAAM,MAAM,GAAG,sBAAsB,CAAC,EAAE,QAAQ,IAAI,WACpD;AACJ,QAAI,KAAK,OAAO;AAAA,EAClB;AACA,SAAO,OAAO,GAAG;AACnB;AAEO,SAAS,sBAAsB,SAA6C;AACjF,MAAI,QAAQ,SAAS,EAAG,QAAO,CAAC;AAGhC,QAAM,YAAY,oBAAI,IAAyB;AAE/C,WAAS,MAAM,GAAG,MAAM,QAAQ,QAAQ,OAAO;AAC7C,UAAM,UAAU,QAAQ,GAAG,EAAE,iBAAiB,YAAY;AAC1D,UAAM,QAAQ,QACX,QAAQ,gBAAgB,GAAG,EAC3B,MAAM,KAAK,EACX,OAAO,CAAC,MAAM,EAAE,UAAU,kBAAkB,CAAC,WAAW,IAAI,CAAC,CAAC;AAEjE,UAAM,gBAAgB,oBAAI,IAAY;AACtC,aAAS,IAAI,GAAG,IAAI,MAAM,SAAS,GAAG,KAAK;AACzC,YAAM,SAAS,MAAM,MAAM,GAAG,IAAI,CAAC,EAAE,KAAK,GAAG;AAC7C,UAAI,cAAc,IAAI,MAAM,EAAG;AAC/B,oBAAc,IAAI,MAAM;AAExB,UAAI,CAAC,UAAU,IAAI,MAAM,EAAG,WAAU,IAAI,QAAQ,oBAAI,IAAI,CAAC;AAC3D,gBAAU,IAAI,MAAM,EAAG,IAAI,GAAG;AAAA,IAChC;AAAA,EACF;AAGA,QAAM,aAA+B,CAAC;AACtC,aAAW,CAAC,QAAQ,aAAa,KAAK,WAAW;AAC/C,QAAI,cAAc,QAAQ,GAAG;AAC3B,iBAAW,KAAK;AAAA,QACd,SAAS;AAAA,QACT,gBAAgB,CAAC,GAAG,aAAa,EAAE,KAAK;AAAA,QACxC,YAAY,cAAc,QAAQ,IAAI,SAAS;AAAA,MACjD,CAAC;AAAA,IACH;AAAA,EACF;AAGA,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,eAAe,SAAS,EAAE,eAAe,MAAM;AAC3E,SAAO,uBAAuB,UAAU,EAAE,MAAM,GAAG,cAAc;AACnE;AAEA,SAAS,uBAAuB,MAA0C;AACxE,QAAM,OAAyB,CAAC;AAChC,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,MAAM,GAAG;AAEnC,UAAM,eAAe,MAAM,OAAO,CAAC,MAAM,UAAU,IAAI,CAAC,CAAC,EAAE;AAC3D,QAAI,gBAAgB,MAAM,SAAS,KAAK,KAAK,SAAS,EAAG;AAEzD,SAAK,KAAK,GAAG;AACb,eAAW,KAAK,MAAO,WAAU,IAAI,CAAC;AAAA,EACxC;AAEA,SAAO;AACT;AAEA,SAAS,WAAW,YAAsB,SAAqC;AAC7E,MAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AAErC,QAAM,OAAiB,CAAC;AACxB,QAAM,eAAe,QAAQ,IAAI,CAAC,MAAM,EAAE,iBAAiB,YAAY,CAAC,EAAE,KAAK,GAAG;AAElF,aAAW,SAAS,YAAY;AAE9B,UAAM,QAAQ,MAAM,YAAY,EAC7B,QAAQ,gBAAgB,GAAG,EAC3B,MAAM,KAAK,EACX,OAAO,CAAC,MAAM,EAAE,UAAU,kBAAkB,CAAC,WAAW,IAAI,CAAC,CAAC;AAEjE,QAAI,MAAM,WAAW,EAAG;AAGxB,UAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,aAAa,SAAS,CAAC,CAAC,EAAE;AAC5D,UAAM,WAAW,QAAQ,MAAM;AAE/B,QAAI,WAAW,KAAK;AAClB,WAAK,KAAK,0BAA0B,KAAK,GAAG;AAAA,IAC9C;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,uBACP,UACA,SACqD;AACrD,QAAM,mBAA6B,CAAC;AACpC,QAAM,eAAe,QAAQ,IAAI,CAAC,MAAM,EAAE,iBAAiB,YAAY,CAAC,EAAE,KAAK,IAAI;AAGnF,QAAM,kBAAkB;AAAA,IAAC;AAAA,IAAU;AAAA,IAAU;AAAA,IAAU;AAAA,IAAS;AAAA,IAAQ;AAAA,IACtE;AAAA,IAAU;AAAA,IAAU;AAAA,IAAW;AAAA,IAAW;AAAA,IAAe;AAAA,IACzD;AAAA,IAAe;AAAA,IAAe;AAAA,IAAa;AAAA,IAAa;AAAA,EAAS;AAEnE,aAAW,QAAQ,iBAAiB;AAClC,QAAI,CAAC,aAAa,SAAS,IAAI,EAAG;AAGlC,UAAM,aAAa,SAAS,KAAK,CAAC,MAAM;AACtC,YAAM,cAAc,EAAE,YAAY;AAClC,YAAM,MAAM,aAAa,QAAQ,WAAW;AAC5C,UAAI,QAAQ,GAAI,QAAO;AAEvB,YAAM,eAAe,aAAa,MAAM,KAAK,IAAI,GAAG,MAAM,GAAG,GAAG,MAAM,EAAE,SAAS,GAAG;AACpF,aAAO,aAAa,SAAS,IAAI;AAAA,IACnC,CAAC;AAED,QAAI,YAAY;AACd,uBAAiB,KAAK,IAAI;AAAA,IAC5B;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,mBAAmB,CAAC,GAAG,IAAI,IAAI,gBAAgB,CAAC,EAAE;AACvE;AAEA,SAAS,0BAA0B,UAAiC;AAClE,QAAM,aAAa,SAAS,MAAM,OAAO,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAC9D,aAAW,KAAK,YAAY;AAC1B,QAAI,EAAE,SAAS,GAAI;AACnB,QAAI,EAAE,WAAW,GAAG,KAAK,EAAE,WAAW,GAAG,KAAK,EAAE,WAAW,KAAK,EAAG;AAGnE,UAAM,WAAW,EACd,QAAQ,6BAA6B,EAAE,EACvC,QAAQ,0CAA0C,EAAE,EACpD,KAAK;AACR,QAAI,SAAS,SAAS,GAAI;AAC1B,WAAO,SAAS,QAAQ,QAAQ,GAAG;AAAA,EACrC;AACA,SAAO;AACT;AAEA,SAAS,OAAO,MAA0B;AACxC,QAAM,OAAO,oBAAI,IAAY;AAC7B,QAAM,MAAgB,CAAC;AACvB,aAAW,QAAQ,MAAM;AACvB,UAAM,MAAM,KAAK,YAAY;AAC7B,QAAI,KAAK,IAAI,GAAG,EAAG;AACnB,SAAK,IAAI,GAAG;AACZ,QAAI,KAAK,IAAI;AAAA,EACf;AACA,SAAO;AACT;AAEA,MAAM,aAAa,oBAAI,IAAI;AAAA,EACzB;AAAA,EAAS;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAU;AAAA,EAAS;AAAA,EACrD;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAC1D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EACxD;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EACxD;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAU;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAC1D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EAC3D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EACzD;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AACpC,CAAC;","names":[]}
+{"version":3,"sources":["../../src/research/brief.ts"],"sourcesContent":["import type { ResearchBrief, ResearchSource, SearchResultItem, CrossReference } from '../types.js';\nimport type { QueryType } from './decompose.js';\nimport { extractHighlights } from '../search/highlights.js';\nimport { buildCitationGraph } from './citation-graph.js';\n\nconst MAX_HIGHLIGHTS = 12;\nconst MAX_KEY_FINDING_LEN = 280;\nconst MAX_TOPICS = 8;\nconst MAX_CROSS_REFS = 10;\nconst MIN_PHRASE_LEN = 4;\n\n// Build a host-LLM-friendly structured brief when internal sampling is\n// unavailable. The host model (Claude Code / Cursor / etc.) consumes this\n// shape to produce the final report without needing to re-read raw sources.\nexport async function buildResearchBrief(\n  question: string,\n  sources: ResearchSource[],\n  subQueries: string[],\n  perSourceCharCap: number,\n  totalSourcesCharCap: number,\n  queryType: QueryType = 'general',\n  comparisonEntities: string[] = [],\n  synthesisText?: string,\n): Promise<ResearchBrief> {\n  const fetched = sources.filter((s) => s.fetched && s.markdown_content.length > 0);\n\n  // Highlights reuse the ONNX-reranker-or-paragraph scorer so briefs align with\n  // whatever format='highlights' produces for single-query searches.\n  const searchItems: SearchResultItem[] = fetched.map((s) => ({\n    title: s.title,\n    url: s.url,\n    snippet: s.markdown_content.slice(0, 200),\n    markdown_content: s.markdown_content,\n    relevance_score: s.relevance_score,\n  }));\n\n  const { highlights } = await extractHighlights(question, searchItems, MAX_HIGHLIGHTS);\n\n  const topics = buildTopics(subQueries, fetched);\n  const keyFindings = buildKeyFindings(fetched);\n  const crossReferences = detectCrossReferences(fetched);\n  const gaps = detectGaps(subQueries, fetched);\n\n  const comparison = queryType === 'comparison' && comparisonEntities.length >= 2\n    ? buildComparisonSection(comparisonEntities, fetched)\n    : undefined;\n\n  const citationGraph = synthesisText && synthesisText.trim().length > 0 && fetched.length > 0\n    ? buildCitationGraph(\n        synthesisText,\n        fetched.map((s) => ({ url: s.url, title: s.title, markdown: s.markdown_content })),\n      )\n    : undefined;\n\n  return {\n    topics,\n    highlights,\n    key_findings: keyFindings,\n    per_source_char_cap: perSourceCharCap,\n    total_sources_char_cap: totalSourcesCharCap,\n    sections: {\n      overview: {\n        key_findings: keyFindings.slice(0, 5),\n        cross_references: crossReferences,\n      },\n      ...(comparison ? { comparison } : {}),\n      gaps,\n    },\n    query_type: queryType,\n    ...(citationGraph && citationGraph.length > 0 ? { citation_graph: citationGraph } : {}),\n  };\n}\n\n// Prefer sub-queries (planner's view of the topic space) when available;\n// otherwise derive compact topic labels from source titles.\nfunction buildTopics(subQueries: string[], sources: ResearchSource[]): string[] {\n  if (subQueries.length > 0) {\n    return dedupe(subQueries).slice(0, MAX_TOPICS);\n  }\n  const labels = sources\n    .map((s) => s.title.split(/[–|:·-]/)[0].trim())\n    .filter((t) => t.length >= 5 && t.length <= 100);\n  return dedupe(labels).slice(0, MAX_TOPICS);\n}\n\n// First substantive paragraph per source, trimmed to a finding-sized blurb.\n// Ordered by source relevance so the most-weighted finding is first.\nfunction buildKeyFindings(sources: ResearchSource[]): string[] {\n  const out: string[] = [];\n  for (const s of [...sources].sort((a, b) => b.relevance_score - a.relevance_score)) {\n    const first = firstSubstantiveParagraph(s.markdown_content);\n    if (!first) continue;\n    const trimmed = first.length > MAX_KEY_FINDING_LEN\n      ? first.slice(0, MAX_KEY_FINDING_LEN - 1).trimEnd() + '…'\n      : first;\n    out.push(trimmed);\n  }\n  return dedupe(out);\n}\n\nexport function detectCrossReferences(sources: ResearchSource[]): CrossReference[] {\n  if (sources.length < 2) return [];\n\n  // Extract significant phrases from each source's content\n  const phraseMap = new Map<string, Set<number>>();\n\n  for (let idx = 0; idx < sources.length; idx++) {\n    const content = sources[idx].markdown_content.toLowerCase();\n    const words = content\n      .replace(/[^a-z0-9\\s]/g, ' ')\n      .split(/\\s+/)\n      .filter((w) => w.length >= MIN_PHRASE_LEN && !STOP_WORDS.has(w));\n\n    const seenForSource = new Set<string>();\n    for (let i = 0; i < words.length - 2; i++) {\n      const phrase = words.slice(i, i + 3).join(' ');\n      if (seenForSource.has(phrase)) continue;\n      seenForSource.add(phrase);\n\n      if (!phraseMap.has(phrase)) phraseMap.set(phrase, new Set());\n      phraseMap.get(phrase)!.add(idx);\n    }\n  }\n\n  // Phrases found in 2+ sources are cross-references\n  const candidates: CrossReference[] = [];\n  for (const [phrase, sourceIndices] of phraseMap) {\n    if (sourceIndices.size >= 2) {\n      candidates.push({\n        finding: phrase,\n        source_indices: [...sourceIndices].sort(),\n        confidence: sourceIndices.size >= 3 ? 'high' : 'medium',\n      });\n    }\n  }\n\n  // Sort by number of sources (desc), then deduplicate overlapping phrases\n  candidates.sort((a, b) => b.source_indices.length - a.source_indices.length);\n  return deduplicateOverlapping(candidates).slice(0, MAX_CROSS_REFS);\n}\n\nfunction deduplicateOverlapping(refs: CrossReference[]): CrossReference[] {\n  const kept: CrossReference[] = [];\n  const usedWords = new Set<string>();\n\n  for (const ref of refs) {\n    const words = ref.finding.split(' ');\n    // Skip if most words already covered by a higher-ranked cross-reference\n    const overlapCount = words.filter((w) => usedWords.has(w)).length;\n    if (overlapCount >= words.length - 1 && kept.length > 0) continue;\n\n    kept.push(ref);\n    for (const w of words) usedWords.add(w);\n  }\n\n  return kept;\n}\n\nfunction detectGaps(subQueries: string[], sources: ResearchSource[]): string[] {\n  if (subQueries.length === 0) return [];\n\n  const gaps: string[] = [];\n  const contentLower = sources.map((s) => s.markdown_content.toLowerCase()).join(' ');\n\n  for (const query of subQueries) {\n    // Extract significant words from sub-query\n    const words = query.toLowerCase()\n      .replace(/[^a-z0-9\\s]/g, ' ')\n      .split(/\\s+/)\n      .filter((w) => w.length >= MIN_PHRASE_LEN && !STOP_WORDS.has(w));\n\n    if (words.length === 0) continue;\n\n    // Count how many significant words appear in any source\n    const found = words.filter((w) => contentLower.includes(w)).length;\n    const coverage = found / words.length;\n\n    if (coverage < 0.5) {\n      gaps.push(`Limited coverage for: \"${query}\"`);\n    }\n  }\n\n  return gaps;\n}\n\nfunction buildComparisonSection(\n  entities: string[],\n  sources: ResearchSource[],\n): { entities: string[]; comparison_points: string[] } {\n  const comparisonPoints: string[] = [];\n  const contentLower = sources.map((s) => s.markdown_content.toLowerCase()).join('\\n');\n\n  // Look for comparison keywords near entity mentions\n  const comparisonTerms = ['faster', 'slower', 'better', 'worse', 'more', 'less',\n    'easier', 'harder', 'simpler', 'complex', 'lightweight', 'heavy',\n    'performance', 'scalability', 'ecosystem', 'community', 'support'];\n\n  for (const term of comparisonTerms) {\n    if (!contentLower.includes(term)) continue;\n\n    // Check if term appears near any entity\n    const nearEntity = entities.some((e) => {\n      const entityLower = e.toLowerCase();\n      const idx = contentLower.indexOf(entityLower);\n      if (idx === -1) return false;\n      // Check within 200 chars of entity mention\n      const neighborhood = contentLower.slice(Math.max(0, idx - 200), idx + e.length + 200);\n      return neighborhood.includes(term);\n    });\n\n    if (nearEntity) {\n      comparisonPoints.push(term);\n    }\n  }\n\n  return { entities, comparison_points: [...new Set(comparisonPoints)] };\n}\n\nfunction firstSubstantiveParagraph(markdown: string): string | null {\n  const paragraphs = markdown.split(/\\n\\n+/).map((p) => p.trim());\n  for (const p of paragraphs) {\n    if (p.length < 80) continue;\n    if (p.startsWith('#') || p.startsWith('|') || p.startsWith('```')) continue;\n    const cleaned = stripMarkdownLinks(p);\n    if (cleaned.length < 80) continue;\n    return cleaned.replace(/\\s+/g, ' ');\n  }\n  return null;\n}\n\n// Flatten markdown link/image syntax to plain text so a downstream char-slice\n// can't chop mid-link and leak `](/?source=post_page...` into key_findings.\nexport function stripMarkdownLinks(text: string): string {\n  return text\n    .replace(/!\\[[^\\]]*\\]\\([^)]*\\)/g, '')\n    .replace(/\\[!\\[[^\\]]*\\]\\([^)]*\\)\\]\\([^)]*\\)/g, '')\n    .replace(/\\[([^\\]]+)\\]\\([^)]*\\)/g, '$1')\n    .replace(/<https?:\\/\\/[^>]+>/g, '')\n    .trim();\n}\n\nfunction dedupe(list: string[]): string[] {\n  const seen = new Set<string>();\n  const out: string[] = [];\n  for (const item of list) {\n    const key = item.toLowerCase();\n    if (seen.has(key)) continue;\n    seen.add(key);\n    out.push(item);\n  }\n  return out;\n}\n\nconst STOP_WORDS = new Set([\n  'about', 'after', 'also', 'been', 'before', 'being', 'between',\n  'both', 'could', 'does', 'doing', 'done', 'each', 'even', 'every',\n  'from', 'have', 'here', 'into', 'just', 'like', 'made', 'make',\n  'many', 'more', 'most', 'much', 'must', 'need', 'only', 'other',\n  'over', 'same', 'should', 'some', 'such', 'than', 'that', 'their',\n  'them', 'then', 'there', 'these', 'they', 'this', 'those', 'through',\n  'very', 'want', 'well', 'were', 'what', 'when', 'where', 'which',\n  'while', 'will', 'with', 'would', 'your',\n]);\n"],"mappings":"AAEA,SAAS,yBAAyB;AAClC,SAAS,0BAA0B;AAEnC,MAAM,iBAAiB;AACvB,MAAM,sBAAsB;AAC5B,MAAM,aAAa;AACnB,MAAM,iBAAiB;AACvB,MAAM,iBAAiB;AAKvB,eAAsB,mBACpB,UACA,SACA,YACA,kBACA,qBACA,YAAuB,WACvB,qBAA+B,CAAC,GAChC,eACwB;AACxB,QAAM,UAAU,QAAQ,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,iBAAiB,SAAS,CAAC;AAIhF,QAAM,cAAkC,QAAQ,IAAI,CAAC,OAAO;AAAA,IAC1D,OAAO,EAAE;AAAA,IACT,KAAK,EAAE;AAAA,IACP,SAAS,EAAE,iBAAiB,MAAM,GAAG,GAAG;AAAA,IACxC,kBAAkB,EAAE;AAAA,IACpB,iBAAiB,EAAE;AAAA,EACrB,EAAE;AAEF,QAAM,EAAE,WAAW,IAAI,MAAM,kBAAkB,UAAU,aAAa,cAAc;AAEpF,QAAM,SAAS,YAAY,YAAY,OAAO;AAC9C,QAAM,cAAc,iBAAiB,OAAO;AAC5C,QAAM,kBAAkB,sBAAsB,OAAO;AACrD,QAAM,OAAO,WAAW,YAAY,OAAO;AAE3C,QAAM,aAAa,cAAc,gBAAgB,mBAAmB,UAAU,IAC1E,uBAAuB,oBAAoB,OAAO,IAClD;AAEJ,QAAM,gBAAgB,iBAAiB,cAAc,KAAK,EAAE,SAAS,KAAK,QAAQ,SAAS,IACvF;AAAA,IACE;AAAA,IACA,QAAQ,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,KAAK,OAAO,EAAE,OAAO,UAAU,EAAE,iBAAiB,EAAE;AAAA,EACnF,IACA;AAEJ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd,qBAAqB;AAAA,IACrB,wBAAwB;AAAA,IACxB,UAAU;AAAA,MACR,UAAU;AAAA,QACR,cAAc,YAAY,MAAM,GAAG,CAAC;AAAA,QACpC,kBAAkB;AAAA,MACpB;AAAA,MACA,GAAI,aAAa,EAAE,WAAW,IAAI,CAAC;AAAA,MACnC;AAAA,IACF;AAAA,IACA,YAAY;AAAA,IACZ,GAAI,iBAAiB,cAAc,SAAS,IAAI,EAAE,gBAAgB,cAAc,IAAI,CAAC;AAAA,EACvF;AACF;AAIA,SAAS,YAAY,YAAsB,SAAqC;AAC9E,MAAI,WAAW,SAAS,GAAG;AACzB,WAAO,OAAO,UAAU,EAAE,MAAM,GAAG,UAAU;AAAA,EAC/C;AACA,QAAM,SAAS,QACZ,IAAI,CAAC,MAAM,EAAE,MAAM,MAAM,SAAS,EAAE,CAAC,EAAE,KAAK,CAAC,EAC7C,OAAO,CAAC,MAAM,EAAE,UAAU,KAAK,EAAE,UAAU,GAAG;AACjD,SAAO,OAAO,MAAM,EAAE,MAAM,GAAG,UAAU;AAC3C;AAIA,SAAS,iBAAiB,SAAqC;AAC7D,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,CAAC,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,kBAAkB,EAAE,eAAe,GAAG;AAClF,UAAM,QAAQ,0BAA0B,EAAE,gBAAgB;AAC1D,QAAI,CAAC,MAAO;AACZ,UAAM,UAAU,MAAM,SAAS,sBAC3B,MAAM,MAAM,GAAG,sBAAsB,CAAC,EAAE,QAAQ,IAAI,WACpD;AACJ,QAAI,KAAK,OAAO;AAAA,EAClB;AACA,SAAO,OAAO,GAAG;AACnB;AAEO,SAAS,sBAAsB,SAA6C;AACjF,MAAI,QAAQ,SAAS,EAAG,QAAO,CAAC;AAGhC,QAAM,YAAY,oBAAI,IAAyB;AAE/C,WAAS,MAAM,GAAG,MAAM,QAAQ,QAAQ,OAAO;AAC7C,UAAM,UAAU,QAAQ,GAAG,EAAE,iBAAiB,YAAY;AAC1D,UAAM,QAAQ,QACX,QAAQ,gBAAgB,GAAG,EAC3B,MAAM,KAAK,EACX,OAAO,CAAC,MAAM,EAAE,UAAU,kBAAkB,CAAC,WAAW,IAAI,CAAC,CAAC;AAEjE,UAAM,gBAAgB,oBAAI,IAAY;AACtC,aAAS,IAAI,GAAG,IAAI,MAAM,SAAS,GAAG,KAAK;AACzC,YAAM,SAAS,MAAM,MAAM,GAAG,IAAI,CAAC,EAAE,KAAK,GAAG;AAC7C,UAAI,cAAc,IAAI,MAAM,EAAG;AAC/B,oBAAc,IAAI,MAAM;AAExB,UAAI,CAAC,UAAU,IAAI,MAAM,EAAG,WAAU,IAAI,QAAQ,oBAAI,IAAI,CAAC;AAC3D,gBAAU,IAAI,MAAM,EAAG,IAAI,GAAG;AAAA,IAChC;AAAA,EACF;AAGA,QAAM,aAA+B,CAAC;AACtC,aAAW,CAAC,QAAQ,aAAa,KAAK,WAAW;AAC/C,QAAI,cAAc,QAAQ,GAAG;AAC3B,iBAAW,KAAK;AAAA,QACd,SAAS;AAAA,QACT,gBAAgB,CAAC,GAAG,aAAa,EAAE,KAAK;AAAA,QACxC,YAAY,cAAc,QAAQ,IAAI,SAAS;AAAA,MACjD,CAAC;AAAA,IACH;AAAA,EACF;AAGA,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,eAAe,SAAS,EAAE,eAAe,MAAM;AAC3E,SAAO,uBAAuB,UAAU,EAAE,MAAM,GAAG,cAAc;AACnE;AAEA,SAAS,uBAAuB,MAA0C;AACxE,QAAM,OAAyB,CAAC;AAChC,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,QAAQ,MAAM,GAAG;AAEnC,UAAM,eAAe,MAAM,OAAO,CAAC,MAAM,UAAU,IAAI,CAAC,CAAC,EAAE;AAC3D,QAAI,gBAAgB,MAAM,SAAS,KAAK,KAAK,SAAS,EAAG;AAEzD,SAAK,KAAK,GAAG;AACb,eAAW,KAAK,MAAO,WAAU,IAAI,CAAC;AAAA,EACxC;AAEA,SAAO;AACT;AAEA,SAAS,WAAW,YAAsB,SAAqC;AAC7E,MAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AAErC,QAAM,OAAiB,CAAC;AACxB,QAAM,eAAe,QAAQ,IAAI,CAAC,MAAM,EAAE,iBAAiB,YAAY,CAAC,EAAE,KAAK,GAAG;AAElF,aAAW,SAAS,YAAY;AAE9B,UAAM,QAAQ,MAAM,YAAY,EAC7B,QAAQ,gBAAgB,GAAG,EAC3B,MAAM,KAAK,EACX,OAAO,CAAC,MAAM,EAAE,UAAU,kBAAkB,CAAC,WAAW,IAAI,CAAC,CAAC;AAEjE,QAAI,MAAM,WAAW,EAAG;AAGxB,UAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,aAAa,SAAS,CAAC,CAAC,EAAE;AAC5D,UAAM,WAAW,QAAQ,MAAM;AAE/B,QAAI,WAAW,KAAK;AAClB,WAAK,KAAK,0BAA0B,KAAK,GAAG;AAAA,IAC9C;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,uBACP,UACA,SACqD;AACrD,QAAM,mBAA6B,CAAC;AACpC,QAAM,eAAe,QAAQ,IAAI,CAAC,MAAM,EAAE,iBAAiB,YAAY,CAAC,EAAE,KAAK,IAAI;AAGnF,QAAM,kBAAkB;AAAA,IAAC;AAAA,IAAU;AAAA,IAAU;AAAA,IAAU;AAAA,IAAS;AAAA,IAAQ;AAAA,IACtE;AAAA,IAAU;AAAA,IAAU;AAAA,IAAW;AAAA,IAAW;AAAA,IAAe;AAAA,IACzD;AAAA,IAAe;AAAA,IAAe;AAAA,IAAa;AAAA,IAAa;AAAA,EAAS;AAEnE,aAAW,QAAQ,iBAAiB;AAClC,QAAI,CAAC,aAAa,SAAS,IAAI,EAAG;AAGlC,UAAM,aAAa,SAAS,KAAK,CAAC,MAAM;AACtC,YAAM,cAAc,EAAE,YAAY;AAClC,YAAM,MAAM,aAAa,QAAQ,WAAW;AAC5C,UAAI,QAAQ,GAAI,QAAO;AAEvB,YAAM,eAAe,aAAa,MAAM,KAAK,IAAI,GAAG,MAAM,GAAG,GAAG,MAAM,EAAE,SAAS,GAAG;AACpF,aAAO,aAAa,SAAS,IAAI;AAAA,IACnC,CAAC;AAED,QAAI,YAAY;AACd,uBAAiB,KAAK,IAAI;AAAA,IAC5B;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,mBAAmB,CAAC,GAAG,IAAI,IAAI,gBAAgB,CAAC,EAAE;AACvE;AAEA,SAAS,0BAA0B,UAAiC;AAClE,QAAM,aAAa,SAAS,MAAM,OAAO,EAAE,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAC9D,aAAW,KAAK,YAAY;AAC1B,QAAI,EAAE,SAAS,GAAI;AACnB,QAAI,EAAE,WAAW,GAAG,KAAK,EAAE,WAAW,GAAG,KAAK,EAAE,WAAW,KAAK,EAAG;AACnE,UAAM,UAAU,mBAAmB,CAAC;AACpC,QAAI,QAAQ,SAAS,GAAI;AACzB,WAAO,QAAQ,QAAQ,QAAQ,GAAG;AAAA,EACpC;AACA,SAAO;AACT;AAIO,SAAS,mBAAmB,MAAsB;AACvD,SAAO,KACJ,QAAQ,yBAAyB,EAAE,EACnC,QAAQ,sCAAsC,EAAE,EAChD,QAAQ,0BAA0B,IAAI,EACtC,QAAQ,uBAAuB,EAAE,EACjC,KAAK;AACV;AAEA,SAAS,OAAO,MAA0B;AACxC,QAAM,OAAO,oBAAI,IAAY;AAC7B,QAAM,MAAgB,CAAC;AACvB,aAAW,QAAQ,MAAM;AACvB,UAAM,MAAM,KAAK,YAAY;AAC7B,QAAI,KAAK,IAAI,GAAG,EAAG;AACnB,SAAK,IAAI,GAAG;AACZ,QAAI,KAAK,IAAI;AAAA,EACf;AACA,SAAO;AACT;AAEA,MAAM,aAAa,oBAAI,IAAI;AAAA,EACzB;AAAA,EAAS;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAU;AAAA,EAAS;AAAA,EACrD;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAC1D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EACxD;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EACxD;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAU;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAC1D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EAC3D;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AAAA,EACzD;AAAA,EAAS;AAAA,EAAQ;AAAA,EAAQ;AAAA,EAAS;AACpC,CAAC;","names":[]}

package/dist/research/pipeline.js

@@ -12,7 +12,7 @@ import { truncateSmartly } from "../search/truncate.js";
 import { cacheContent } from "../cache/store.js";
 import { getEmbeddingService } from "../embedding/embed.js";
 import { checkSamplingSupport } from "../search/sampling.js";
-import { isLocalLlmEnabled } from "../extraction/v1/local-llm.js";
+import { isLlmConfigured as isLocalLlmEnabled } from "../integrations/cloud/llm/run.js";
 const log = createLogger("research");
 const DEPTH_CONFIG = {
   quick: { subQueries: 2, minSources: 5, maxSources: 8 },

package/dist/research/pipeline.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/research/pipeline.ts"],"sourcesContent":["import { createLogger } from '../logger.js';\nimport { decomposeQuestion, detectQueryType, extractComparisonEntities, type QueryType } from './decompose.js';\nimport { synthesizeReport } from './synthesize.js';\nimport { synthesizeLocal } from './synthesis-local.js';\nimport { buildResearchBrief } from './brief.js';\nimport { deduplicateResults } from '../search/dedup.js';\nimport { rerankResults } from '../search/rerank.js';\nimport { applyAllFilters } from '../search/filters.js';\nimport { exploreInParallel } from './branch-exploration.js';\nimport type { RawSearchResult, SearchEngineOptions } from '../types.js';\nimport { getExtractProvider } from '../providers/extract-provider.js';\nimport { truncateSmartly } from '../search/truncate.js';\nimport { cacheContent } from '../cache/store.js';\nimport { getEmbeddingService } from '../embedding/embed.js';\nimport { checkSamplingSupport, type SamplingCapableServer } from '../search/sampling.js';\nimport { isLocalLlmEnabled } from '../extraction/v1/local-llm.js';\nimport type {\n  ResearchInput,\n  ResearchOutput,\n  ResearchSource,\n  SearchEngine,\n  Citation,\n} from '../types.js';\nimport type { SmartRouter } from '../fetch/router.js';\n\nconst log = createLogger('research');\n\nconst DEPTH_CONFIG: Record<string, { subQueries: number; minSources: number; maxSources: number }> = {\n  quick: { subQueries: 2, minSources: 5, maxSources: 8 },\n  standard: { subQueries: 4, minSources: 10, maxSources: 15 },\n  comprehensive: { subQueries: 7, minSources: 20, maxSources: 25 },\n};\n\n// Per-depth budgets for the sub-query fan-out. exploreInParallel guarantees\n// a single slow sub-query can't burn the whole research budget — comprehensive\n// runs cap at ~60s total and 15s per sub-query.\nconst SEARCH_TOTAL_BUDGET_MS: Record<string, number> = {\n  quick: 15_000,\n  standard: 30_000,\n  comprehensive: 60_000,\n};\nconst SEARCH_PER_QUERY_BUDGET_MS: Record<string, number> = {\n  quick: 8_000,\n  standard: 10_000,\n  comprehensive: 15_000,\n};\n\nconst PER_SOURCE_CHAR_CAP = 3000;\nconst TOTAL_SOURCES_CHAR_CAP = 40000;\n\nexport async function runResearchPipeline(\n  input: ResearchInput,\n  engines: SearchEngine[],\n  router: SmartRouter,\n  server?: SamplingCapableServer,\n): Promise<ResearchOutput> {\n  const start = Date.now();\n  const depth = input.depth ?? 'standard';\n  const config = DEPTH_CONFIG[depth] ?? DEPTH_CONFIG.standard;\n  const maxSources = input.max_sources ?? config.maxSources;\n\n  try {\n    // Phase 1: Decompose question into sub-queries\n    log.info('research pipeline started', { question: input.question, depth });\n    const decomposeResult = await decomposeQuestion(\n      input.question,\n      depth as 'quick' | 'standard' | 'comprehensive',\n      server,\n    );\n    const subQueries = decomposeResult.subQueries;\n    const queryType = decomposeResult.queryType;\n    log.info('decomposition complete', { subQueryCount: subQueries.length, samplingUsed: decomposeResult.samplingUsed, queryType });\n\n    // Phase 2: Parallel search across sub-queries with per-query + total\n    // budget enforcement via exploreInParallel. A single hung engine no\n    // longer wedges the whole research call — the per-query timer aborts\n    // it and the rest of the fan-out keeps going. Engine cap when\n    // sub-queries are many preserves the multi-query.ts invariant.\n    const effEngines = subQueries.length >= 3 && engines.length > 2 ? engines.slice(0, 2) : engines;\n    const perEngineMaxResults = Math.ceil(maxSources / subQueries.length) * 2;\n\n    const branchResults = await exploreInParallel(\n      subQueries,\n      async (subQuery, signal) => {\n        const results: RawSearchResult[] = [];\n        const usedHere = new Set<string>();\n        const engineOpts: SearchEngineOptions = {\n          maxResults: perEngineMaxResults,\n          includeDomains: input.include_domains,\n          excludeDomains: input.exclude_domains,\n        };\n\n        await Promise.allSettled(\n          effEngines.map(async (engine) => {\n            if (signal.aborted) return;\n            try {\n              const rs = await engine.search(subQuery, engineOpts);\n              for (const r of rs) results.push(r);\n              usedHere.add(engine.name);\n            } catch (err) {\n              log.warn('research engine search failed', {\n                engine: engine.name,\n                query: subQuery,\n                error: err instanceof Error ? err.message : String(err),\n              });\n            }\n          }),\n        );\n\n        return { results, enginesUsed: [...usedHere] };\n      },\n      {\n        maxConcurrent: 3,\n        totalBudgetMs: SEARCH_TOTAL_BUDGET_MS[depth] ?? SEARCH_TOTAL_BUDGET_MS.standard,\n        perQueryBudgetMs: SEARCH_PER_QUERY_BUDGET_MS[depth] ?? SEARCH_PER_QUERY_BUDGET_MS.standard,\n      },\n    );\n\n    const allRaw: RawSearchResult[] = [];\n    const enginesUsed = new Set<string>();\n    const searchErrors: string[] = [];\n    for (const br of branchResults) {\n      if (br.ok && br.result) {\n        allRaw.push(...br.result.results);\n        for (const e of br.result.enginesUsed) enginesUsed.add(e);\n      } else if (br.error) {\n        searchErrors.push(`${br.query}: ${br.error}`);\n      }\n    }\n    if (searchErrors.length > 0) {\n      log.warn('some search sub-queries failed', { errors: searchErrors });\n    }\n\n    log.info('search phase complete', { totalRaw: allRaw.length, engines: [...enginesUsed] });\n\n    // Phase 3: Deduplicate, filter, rerank\n    let merged = deduplicateResults(allRaw);\n\n    merged = applyAllFilters(merged, {\n      includeDomains: input.include_domains,\n      excludeDomains: input.exclude_domains,\n    });\n\n    merged = await rerankResults(input.question, merged);\n    merged = merged.slice(0, maxSources);\n\n    if (merged.length === 0) {\n      return {\n        report: `## Research: ${input.question}\\n\\nNo sources could be found for this query.`,\n        citations: [],\n        sources: [],\n        sub_queries: subQueries,\n        depth,\n        total_time_ms: Date.now() - start,\n        sampling_supported: !!server && checkSamplingSupport(server),\n      };\n    }\n\n    // Phase 4: Fetch top sources in parallel\n    const sources: ResearchSource[] = await fetchSources(merged, router, maxSources);\n    applySourceBudget(sources, PER_SOURCE_CHAR_CAP, TOTAL_SOURCES_CHAR_CAP);\n    log.info('fetch phase complete', {\n      fetched: sources.filter((s) => s.fetched).length,\n      failed: sources.filter((s) => !s.fetched).length,\n    });\n\n    // Phase 5: Synthesize report\n    const synthesisResult = await synthesizeReport(\n      input.question,\n      sources,\n      depth as 'quick' | 'standard' | 'comprehensive',\n      server,\n    );\n    log.info('synthesis complete', { samplingUsed: synthesisResult.samplingUsed, reportLength: synthesisResult.report.length });\n\n    // Phase 5b: Local-LLM synthesis fallback — only when host LLM did not\n    // produce output AND a local provider is configured. Failures fall through\n    // to the existing heuristic report in synthesisResult.\n    let finalReport = synthesisResult.report;\n    let finalCitations: Citation[] = synthesisResult.citations;\n    let localSynthesisText: string | undefined;\n    if (!synthesisResult.samplingUsed && isLocalLlmEnabled()) {\n      try {\n        const localSources = sources\n          .filter((s) => s.fetched && s.markdown_content.length > 0)\n          .map((s) => ({ url: s.url, title: s.title, markdown: s.markdown_content }));\n        if (localSources.length > 0) {\n          const local = await synthesizeLocal(input.question, localSources);\n          finalReport = local.text;\n          localSynthesisText = local.text;\n          finalCitations = local.citations\n            .filter((idx) => idx >= 0 && idx < localSources.length)\n            .map((idx) => {\n              const s = localSources[idx];\n              return {\n                index: idx + 1,\n                url: s.url,\n                title: s.title,\n                snippet: s.markdown.slice(0, 200),\n              };\n            });\n          log.info('local synthesis succeeded', { reportLength: finalReport.length });\n        }\n      } catch (err) {\n        log.warn('local LLM synthesis failed; using heuristic fallback', {\n          error: err instanceof Error ? err.message : String(err),\n        });\n      }\n    }\n\n    // Phase 6: Structured brief — populated when internal sampling is\n    // unavailable so the host LLM has well-shaped data to write the report\n    // from without re-reading raw markdown.\n    const comparisonEntities = queryType === 'comparison'\n      ? extractComparisonEntities(input.question).entities\n      : [];\n    const brief = !synthesisResult.samplingUsed\n      ? await buildResearchBrief(\n          input.question,\n          sources,\n          subQueries,\n          PER_SOURCE_CHAR_CAP,\n          TOTAL_SOURCES_CHAR_CAP,\n          queryType,\n          comparisonEntities,\n          localSynthesisText,\n        )\n      : undefined;\n\n    return {\n      report: finalReport,\n      citations: finalCitations,\n      sources,\n      sub_queries: subQueries,\n      depth,\n      total_time_ms: Date.now() - start,\n      sampling_supported: !!server && checkSamplingSupport(server),\n      ...(brief ? { brief } : {}),\n    };\n  } catch (err) {\n    log.error('research pipeline failed', {\n      question: input.question,\n      error: err instanceof Error ? err.message : String(err),\n    });\n    return {\n      report: '',\n      citations: [],\n      sources: [],\n      sub_queries: [],\n      depth,\n      total_time_ms: Date.now() - start,\n      sampling_supported: !!server && checkSamplingSupport(server),\n      error: err instanceof Error ? err.message : String(err),\n    };\n  }\n}\n\ninterface MergedResult {\n  title: string;\n  url: string;\n  snippet: string;\n  relevance_score: number;\n  engines: string[];\n}\n\nasync function fetchSources(\n  merged: MergedResult[],\n  router: SmartRouter,\n  maxSources: number,\n): Promise<ResearchSource[]> {\n  const fetchPromises = merged.slice(0, maxSources).map(async (result): Promise<ResearchSource> => {\n    try {\n      const raw = await Promise.race([\n        router.fetch(result.url, { renderJs: 'auto' }),\n        new Promise<never>((_, reject) =>\n          setTimeout(() => reject(new Error('fetch timeout')), 15000),\n        ),\n      ]);\n\n      const extractor = await getExtractProvider();\n      const extraction = await extractor.extract(raw.html, raw.finalUrl, {\n        maxChars: 30000,\n        contentType: raw.contentType,\n      });\n      const truncated = truncateSmartly(extraction.markdown, PER_SOURCE_CHAR_CAP);\n\n      try {\n        cacheContent(raw, extraction);\n      } catch (err) {\n        log.warn('failed to cache research source', { url: result.url, error: String(err) });\n      }\n\n      try {\n        const embeddingService = getEmbeddingService();\n        if (embeddingService.isAvailable()) {\n          embeddingService.embedAsync(raw.finalUrl, extraction.markdown);\n        }\n      } catch (err) {\n        log.debug('embedding hook skipped for research source', { error: String(err) });\n      }\n\n      return {\n        url: result.url,\n        title: extraction.title || result.title,\n        markdown_content: truncated,\n        relevance_score: result.relevance_score,\n        fetched: true,\n      };\n    } catch (err) {\n      log.debug('failed to fetch research source', {\n        url: result.url,\n        error: err instanceof Error ? err.message : String(err),\n      });\n      return {\n        url: result.url,\n        title: result.title,\n        markdown_content: result.snippet,\n        relevance_score: result.relevance_score,\n        fetched: false,\n        fetch_error: err instanceof Error ? err.message : String(err),\n      };\n    }\n  });\n\n  return Promise.all(fetchPromises);\n}\n\n// Cap total returned markdown_content across sources in relevance order.\n// Later (lower-relevance) sources get trimmed further when budget runs low;\n// any source past the cap is set to empty content (caller still sees url/title).\nfunction applySourceBudget(\n  sources: ResearchSource[],\n  perSourceCap: number,\n  totalCap: number,\n): void {\n  let used = 0;\n  for (const s of sources) {\n    if (!s.markdown_content) continue;\n    if (used >= totalCap) {\n      s.markdown_content = '';\n      continue;\n    }\n    const remaining = totalCap - used;\n    const cap = Math.min(perSourceCap, remaining);\n    if (s.markdown_content.length > cap) {\n      s.markdown_content = truncateSmartly(s.markdown_content, cap);\n    }\n    used += s.markdown_content.length;\n  }\n}\n"],"mappings":"AAAA,SAAS,oBAAoB;AAC7B,SAAS,mBAAoC,iCAAiD;AAC9F,SAAS,wBAAwB;AACjC,SAAS,uBAAuB;AAChC,SAAS,0BAA0B;AACnC,SAAS,0BAA0B;AACnC,SAAS,qBAAqB;AAC9B,SAAS,uBAAuB;AAChC,SAAS,yBAAyB;AAElC,SAAS,0BAA0B;AACnC,SAAS,uBAAuB;AAChC,SAAS,oBAAoB;AAC7B,SAAS,2BAA2B;AACpC,SAAS,4BAAwD;AACjE,SAAS,yBAAyB;AAUlC,MAAM,MAAM,aAAa,UAAU;AAEnC,MAAM,eAA+F;AAAA,EACnG,OAAO,EAAE,YAAY,GAAG,YAAY,GAAG,YAAY,EAAE;AAAA,EACrD,UAAU,EAAE,YAAY,GAAG,YAAY,IAAI,YAAY,GAAG;AAAA,EAC1D,eAAe,EAAE,YAAY,GAAG,YAAY,IAAI,YAAY,GAAG;AACjE;AAKA,MAAM,yBAAiD;AAAA,EACrD,OAAO;AAAA,EACP,UAAU;AAAA,EACV,eAAe;AACjB;AACA,MAAM,6BAAqD;AAAA,EACzD,OAAO;AAAA,EACP,UAAU;AAAA,EACV,eAAe;AACjB;AAEA,MAAM,sBAAsB;AAC5B,MAAM,yBAAyB;AAE/B,eAAsB,oBACpB,OACA,SACA,QACA,QACyB;AACzB,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,QAAQ,MAAM,SAAS;AAC7B,QAAM,SAAS,aAAa,KAAK,KAAK,aAAa;AACnD,QAAM,aAAa,MAAM,eAAe,OAAO;AAE/C,MAAI;AAEF,QAAI,KAAK,6BAA6B,EAAE,UAAU,MAAM,UAAU,MAAM,CAAC;AACzE,UAAM,kBAAkB,MAAM;AAAA,MAC5B,MAAM;AAAA,MACN;AAAA,MACA;AAAA,IACF;AACA,UAAM,aAAa,gBAAgB;AACnC,UAAM,YAAY,gBAAgB;AAClC,QAAI,KAAK,0BAA0B,EAAE,eAAe,WAAW,QAAQ,cAAc,gBAAgB,cAAc,UAAU,CAAC;AAO9H,UAAM,aAAa,WAAW,UAAU,KAAK,QAAQ,SAAS,IAAI,QAAQ,MAAM,GAAG,CAAC,IAAI;AACxF,UAAM,sBAAsB,KAAK,KAAK,aAAa,WAAW,MAAM,IAAI;AAExE,UAAM,gBAAgB,MAAM;AAAA,MAC1B;AAAA,MACA,OAAO,UAAU,WAAW;AAC1B,cAAM,UAA6B,CAAC;AACpC,cAAM,WAAW,oBAAI,IAAY;AACjC,cAAM,aAAkC;AAAA,UACtC,YAAY;AAAA,UACZ,gBAAgB,MAAM;AAAA,UACtB,gBAAgB,MAAM;AAAA,QACxB;AAEA,cAAM,QAAQ;AAAA,UACZ,WAAW,IAAI,OAAO,WAAW;AAC/B,gBAAI,OAAO,QAAS;AACpB,gBAAI;AACF,oBAAM,KAAK,MAAM,OAAO,OAAO,UAAU,UAAU;AACnD,yBAAW,KAAK,GAAI,SAAQ,KAAK,CAAC;AAClC,uBAAS,IAAI,OAAO,IAAI;AAAA,YAC1B,SAAS,KAAK;AACZ,kBAAI,KAAK,iCAAiC;AAAA,gBACxC,QAAQ,OAAO;AAAA,gBACf,OAAO;AAAA,gBACP,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,cACxD,CAAC;AAAA,YACH;AAAA,UACF,CAAC;AAAA,QACH;AAEA,eAAO,EAAE,SAAS,aAAa,CAAC,GAAG,QAAQ,EAAE;AAAA,MAC/C;AAAA,MACA;AAAA,QACE,eAAe;AAAA,QACf,eAAe,uBAAuB,KAAK,KAAK,uBAAuB;AAAA,QACvE,kBAAkB,2BAA2B,KAAK,KAAK,2BAA2B;AAAA,MACpF;AAAA,IACF;AAEA,UAAM,SAA4B,CAAC;AACnC,UAAM,cAAc,oBAAI,IAAY;AACpC,UAAM,eAAyB,CAAC;AAChC,eAAW,MAAM,eAAe;AAC9B,UAAI,GAAG,MAAM,GAAG,QAAQ;AACtB,eAAO,KAAK,GAAG,GAAG,OAAO,OAAO;AAChC,mBAAW,KAAK,GAAG,OAAO,YAAa,aAAY,IAAI,CAAC;AAAA,MAC1D,WAAW,GAAG,OAAO;AACnB,qBAAa,KAAK,GAAG,GAAG,KAAK,KAAK,GAAG,KAAK,EAAE;AAAA,MAC9C;AAAA,IACF;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,UAAI,KAAK,kCAAkC,EAAE,QAAQ,aAAa,CAAC;AAAA,IACrE;AAEA,QAAI,KAAK,yBAAyB,EAAE,UAAU,OAAO,QAAQ,SAAS,CAAC,GAAG,WAAW,EAAE,CAAC;AAGxF,QAAI,SAAS,mBAAmB,MAAM;AAEtC,aAAS,gBAAgB,QAAQ;AAAA,MAC/B,gBAAgB,MAAM;AAAA,MACtB,gBAAgB,MAAM;AAAA,IACxB,CAAC;AAED,aAAS,MAAM,cAAc,MAAM,UAAU,MAAM;AACnD,aAAS,OAAO,MAAM,GAAG,UAAU;AAEnC,QAAI,OAAO,WAAW,GAAG;AACvB,aAAO;AAAA,QACL,QAAQ,gBAAgB,MAAM,QAAQ;AAAA;AAAA;AAAA,QACtC,WAAW,CAAC;AAAA,QACZ,SAAS,CAAC;AAAA,QACV,aAAa;AAAA,QACb;AAAA,QACA,eAAe,KAAK,IAAI,IAAI;AAAA,QAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC7D;AAAA,IACF;AAGA,UAAM,UAA4B,MAAM,aAAa,QAAQ,QAAQ,UAAU;AAC/E,sBAAkB,SAAS,qBAAqB,sBAAsB;AACtE,QAAI,KAAK,wBAAwB;AAAA,MAC/B,SAAS,QAAQ,OAAO,CAAC,MAAM,EAAE,OAAO,EAAE;AAAA,MAC1C,QAAQ,QAAQ,OAAO,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE;AAAA,IAC5C,CAAC;AAGD,UAAM,kBAAkB,MAAM;AAAA,MAC5B,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,QAAI,KAAK,sBAAsB,EAAE,cAAc,gBAAgB,cAAc,cAAc,gBAAgB,OAAO,OAAO,CAAC;AAK1H,QAAI,cAAc,gBAAgB;AAClC,QAAI,iBAA6B,gBAAgB;AACjD,QAAI;AACJ,QAAI,CAAC,gBAAgB,gBAAgB,kBAAkB,GAAG;AACxD,UAAI;AACF,cAAM,eAAe,QAClB,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,iBAAiB,SAAS,CAAC,EACxD,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,KAAK,OAAO,EAAE,OAAO,UAAU,EAAE,iBAAiB,EAAE;AAC5E,YAAI,aAAa,SAAS,GAAG;AAC3B,gBAAM,QAAQ,MAAM,gBAAgB,MAAM,UAAU,YAAY;AAChE,wBAAc,MAAM;AACpB,+BAAqB,MAAM;AAC3B,2BAAiB,MAAM,UACpB,OAAO,CAAC,QAAQ,OAAO,KAAK,MAAM,aAAa,MAAM,EACrD,IAAI,CAAC,QAAQ;AACZ,kBAAM,IAAI,aAAa,GAAG;AAC1B,mBAAO;AAAA,cACL,OAAO,MAAM;AAAA,cACb,KAAK,EAAE;AAAA,cACP,OAAO,EAAE;AAAA,cACT,SAAS,EAAE,SAAS,MAAM,GAAG,GAAG;AAAA,YAClC;AAAA,UACF,CAAC;AACH,cAAI,KAAK,6BAA6B,EAAE,cAAc,YAAY,OAAO,CAAC;AAAA,QAC5E;AAAA,MACF,SAAS,KAAK;AACZ,YAAI,KAAK,wDAAwD;AAAA,UAC/D,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,QACxD,CAAC;AAAA,MACH;AAAA,IACF;AAKA,UAAM,qBAAqB,cAAc,eACrC,0BAA0B,MAAM,QAAQ,EAAE,WAC1C,CAAC;AACL,UAAM,QAAQ,CAAC,gBAAgB,eAC3B,MAAM;AAAA,MACJ,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,IACA;AAEJ,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,WAAW;AAAA,MACX;AAAA,MACA,aAAa;AAAA,MACb;AAAA,MACA,eAAe,KAAK,IAAI,IAAI;AAAA,MAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC3D,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,IAC3B;AAAA,EACF,SAAS,KAAK;AACZ,QAAI,MAAM,4BAA4B;AAAA,MACpC,UAAU,MAAM;AAAA,MAChB,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD,CAAC;AACD,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,WAAW,CAAC;AAAA,MACZ,SAAS,CAAC;AAAA,MACV,aAAa,CAAC;AAAA,MACd;AAAA,MACA,eAAe,KAAK,IAAI,IAAI;AAAA,MAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC3D,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD;AAAA,EACF;AACF;AAUA,eAAe,aACb,QACA,QACA,YAC2B;AAC3B,QAAM,gBAAgB,OAAO,MAAM,GAAG,UAAU,EAAE,IAAI,OAAO,WAAoC;AAC/F,QAAI;AACF,YAAM,MAAM,MAAM,QAAQ,KAAK;AAAA,QAC7B,OAAO,MAAM,OAAO,KAAK,EAAE,UAAU,OAAO,CAAC;AAAA,QAC7C,IAAI;AAAA,UAAe,CAAC,GAAG,WACrB,WAAW,MAAM,OAAO,IAAI,MAAM,eAAe,CAAC,GAAG,IAAK;AAAA,QAC5D;AAAA,MACF,CAAC;AAED,YAAM,YAAY,MAAM,mBAAmB;AAC3C,YAAM,aAAa,MAAM,UAAU,QAAQ,IAAI,MAAM,IAAI,UAAU;AAAA,QACjE,UAAU;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AACD,YAAM,YAAY,gBAAgB,WAAW,UAAU,mBAAmB;AAE1E,UAAI;AACF,qBAAa,KAAK,UAAU;AAAA,MAC9B,SAAS,KAAK;AACZ,YAAI,KAAK,mCAAmC,EAAE,KAAK,OAAO,KAAK,OAAO,OAAO,GAAG,EAAE,CAAC;AAAA,MACrF;AAEA,UAAI;AACF,cAAM,mBAAmB,oBAAoB;AAC7C,YAAI,iBAAiB,YAAY,GAAG;AAClC,2BAAiB,WAAW,IAAI,UAAU,WAAW,QAAQ;AAAA,QAC/D;AAAA,MACF,SAAS,KAAK;AACZ,YAAI,MAAM,8CAA8C,EAAE,OAAO,OAAO,GAAG,EAAE,CAAC;AAAA,MAChF;AAEA,aAAO;AAAA,QACL,KAAK,OAAO;AAAA,QACZ,OAAO,WAAW,SAAS,OAAO;AAAA,QAClC,kBAAkB;AAAA,QAClB,iBAAiB,OAAO;AAAA,QACxB,SAAS;AAAA,MACX;AAAA,IACF,SAAS,KAAK;AACZ,UAAI,MAAM,mCAAmC;AAAA,QAC3C,KAAK,OAAO;AAAA,QACZ,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AACD,aAAO;AAAA,QACL,KAAK,OAAO;AAAA,QACZ,OAAO,OAAO;AAAA,QACd,kBAAkB,OAAO;AAAA,QACzB,iBAAiB,OAAO;AAAA,QACxB,SAAS;AAAA,QACT,aAAa,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MAC9D;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO,QAAQ,IAAI,aAAa;AAClC;AAKA,SAAS,kBACP,SACA,cACA,UACM;AACN,MAAI,OAAO;AACX,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,EAAE,iBAAkB;AACzB,QAAI,QAAQ,UAAU;AACpB,QAAE,mBAAmB;AACrB;AAAA,IACF;AACA,UAAM,YAAY,WAAW;AAC7B,UAAM,MAAM,KAAK,IAAI,cAAc,SAAS;AAC5C,QAAI,EAAE,iBAAiB,SAAS,KAAK;AACnC,QAAE,mBAAmB,gBAAgB,EAAE,kBAAkB,GAAG;AAAA,IAC9D;AACA,YAAQ,EAAE,iBAAiB;AAAA,EAC7B;AACF;","names":[]}
+{"version":3,"sources":["../../src/research/pipeline.ts"],"sourcesContent":["import { createLogger } from '../logger.js';\nimport { decomposeQuestion, detectQueryType, extractComparisonEntities, type QueryType } from './decompose.js';\nimport { synthesizeReport } from './synthesize.js';\nimport { synthesizeLocal } from './synthesis-local.js';\nimport { buildResearchBrief } from './brief.js';\nimport { deduplicateResults } from '../search/dedup.js';\nimport { rerankResults } from '../search/rerank.js';\nimport { applyAllFilters } from '../search/filters.js';\nimport { exploreInParallel } from './branch-exploration.js';\nimport type { RawSearchResult, SearchEngineOptions } from '../types.js';\nimport { getExtractProvider } from '../providers/extract-provider.js';\nimport { truncateSmartly } from '../search/truncate.js';\nimport { cacheContent } from '../cache/store.js';\nimport { getEmbeddingService } from '../embedding/embed.js';\nimport { checkSamplingSupport, type SamplingCapableServer } from '../search/sampling.js';\nimport { isLlmConfigured as isLocalLlmEnabled } from '../integrations/cloud/llm/run.js';\nimport type {\n  ResearchInput,\n  ResearchOutput,\n  ResearchSource,\n  SearchEngine,\n  Citation,\n} from '../types.js';\nimport type { SmartRouter } from '../fetch/router.js';\n\nconst log = createLogger('research');\n\nconst DEPTH_CONFIG: Record<string, { subQueries: number; minSources: number; maxSources: number }> = {\n  quick: { subQueries: 2, minSources: 5, maxSources: 8 },\n  standard: { subQueries: 4, minSources: 10, maxSources: 15 },\n  comprehensive: { subQueries: 7, minSources: 20, maxSources: 25 },\n};\n\n// Per-depth budgets for the sub-query fan-out. exploreInParallel guarantees\n// a single slow sub-query can't burn the whole research budget — comprehensive\n// runs cap at ~60s total and 15s per sub-query.\nconst SEARCH_TOTAL_BUDGET_MS: Record<string, number> = {\n  quick: 15_000,\n  standard: 30_000,\n  comprehensive: 60_000,\n};\nconst SEARCH_PER_QUERY_BUDGET_MS: Record<string, number> = {\n  quick: 8_000,\n  standard: 10_000,\n  comprehensive: 15_000,\n};\n\nconst PER_SOURCE_CHAR_CAP = 3000;\nconst TOTAL_SOURCES_CHAR_CAP = 40000;\n\nexport async function runResearchPipeline(\n  input: ResearchInput,\n  engines: SearchEngine[],\n  router: SmartRouter,\n  server?: SamplingCapableServer,\n): Promise<ResearchOutput> {\n  const start = Date.now();\n  const depth = input.depth ?? 'standard';\n  const config = DEPTH_CONFIG[depth] ?? DEPTH_CONFIG.standard;\n  const maxSources = input.max_sources ?? config.maxSources;\n\n  try {\n    // Phase 1: Decompose question into sub-queries\n    log.info('research pipeline started', { question: input.question, depth });\n    const decomposeResult = await decomposeQuestion(\n      input.question,\n      depth as 'quick' | 'standard' | 'comprehensive',\n      server,\n    );\n    const subQueries = decomposeResult.subQueries;\n    const queryType = decomposeResult.queryType;\n    log.info('decomposition complete', { subQueryCount: subQueries.length, samplingUsed: decomposeResult.samplingUsed, queryType });\n\n    // Phase 2: Parallel search across sub-queries with per-query + total\n    // budget enforcement via exploreInParallel. A single hung engine no\n    // longer wedges the whole research call — the per-query timer aborts\n    // it and the rest of the fan-out keeps going. Engine cap when\n    // sub-queries are many preserves the multi-query.ts invariant.\n    const effEngines = subQueries.length >= 3 && engines.length > 2 ? engines.slice(0, 2) : engines;\n    const perEngineMaxResults = Math.ceil(maxSources / subQueries.length) * 2;\n\n    const branchResults = await exploreInParallel(\n      subQueries,\n      async (subQuery, signal) => {\n        const results: RawSearchResult[] = [];\n        const usedHere = new Set<string>();\n        const engineOpts: SearchEngineOptions = {\n          maxResults: perEngineMaxResults,\n          includeDomains: input.include_domains,\n          excludeDomains: input.exclude_domains,\n        };\n\n        await Promise.allSettled(\n          effEngines.map(async (engine) => {\n            if (signal.aborted) return;\n            try {\n              const rs = await engine.search(subQuery, engineOpts);\n              for (const r of rs) results.push(r);\n              usedHere.add(engine.name);\n            } catch (err) {\n              log.warn('research engine search failed', {\n                engine: engine.name,\n                query: subQuery,\n                error: err instanceof Error ? err.message : String(err),\n              });\n            }\n          }),\n        );\n\n        return { results, enginesUsed: [...usedHere] };\n      },\n      {\n        maxConcurrent: 3,\n        totalBudgetMs: SEARCH_TOTAL_BUDGET_MS[depth] ?? SEARCH_TOTAL_BUDGET_MS.standard,\n        perQueryBudgetMs: SEARCH_PER_QUERY_BUDGET_MS[depth] ?? SEARCH_PER_QUERY_BUDGET_MS.standard,\n      },\n    );\n\n    const allRaw: RawSearchResult[] = [];\n    const enginesUsed = new Set<string>();\n    const searchErrors: string[] = [];\n    for (const br of branchResults) {\n      if (br.ok && br.result) {\n        allRaw.push(...br.result.results);\n        for (const e of br.result.enginesUsed) enginesUsed.add(e);\n      } else if (br.error) {\n        searchErrors.push(`${br.query}: ${br.error}`);\n      }\n    }\n    if (searchErrors.length > 0) {\n      log.warn('some search sub-queries failed', { errors: searchErrors });\n    }\n\n    log.info('search phase complete', { totalRaw: allRaw.length, engines: [...enginesUsed] });\n\n    // Phase 3: Deduplicate, filter, rerank\n    let merged = deduplicateResults(allRaw);\n\n    merged = applyAllFilters(merged, {\n      includeDomains: input.include_domains,\n      excludeDomains: input.exclude_domains,\n    });\n\n    merged = await rerankResults(input.question, merged);\n    merged = merged.slice(0, maxSources);\n\n    if (merged.length === 0) {\n      return {\n        report: `## Research: ${input.question}\\n\\nNo sources could be found for this query.`,\n        citations: [],\n        sources: [],\n        sub_queries: subQueries,\n        depth,\n        total_time_ms: Date.now() - start,\n        sampling_supported: !!server && checkSamplingSupport(server),\n      };\n    }\n\n    // Phase 4: Fetch top sources in parallel\n    const sources: ResearchSource[] = await fetchSources(merged, router, maxSources);\n    applySourceBudget(sources, PER_SOURCE_CHAR_CAP, TOTAL_SOURCES_CHAR_CAP);\n    log.info('fetch phase complete', {\n      fetched: sources.filter((s) => s.fetched).length,\n      failed: sources.filter((s) => !s.fetched).length,\n    });\n\n    // Phase 5: Synthesize report\n    const synthesisResult = await synthesizeReport(\n      input.question,\n      sources,\n      depth as 'quick' | 'standard' | 'comprehensive',\n      server,\n    );\n    log.info('synthesis complete', { samplingUsed: synthesisResult.samplingUsed, reportLength: synthesisResult.report.length });\n\n    // Phase 5b: Local-LLM synthesis fallback — only when host LLM did not\n    // produce output AND a local provider is configured. Failures fall through\n    // to the existing heuristic report in synthesisResult.\n    let finalReport = synthesisResult.report;\n    let finalCitations: Citation[] = synthesisResult.citations;\n    let localSynthesisText: string | undefined;\n    if (!synthesisResult.samplingUsed && isLocalLlmEnabled()) {\n      try {\n        const localSources = sources\n          .filter((s) => s.fetched && s.markdown_content.length > 0)\n          .map((s) => ({ url: s.url, title: s.title, markdown: s.markdown_content }));\n        if (localSources.length > 0) {\n          const local = await synthesizeLocal(input.question, localSources);\n          finalReport = local.text;\n          localSynthesisText = local.text;\n          finalCitations = local.citations\n            .filter((idx) => idx >= 0 && idx < localSources.length)\n            .map((idx) => {\n              const s = localSources[idx];\n              return {\n                index: idx + 1,\n                url: s.url,\n                title: s.title,\n                snippet: s.markdown.slice(0, 200),\n              };\n            });\n          log.info('local synthesis succeeded', { reportLength: finalReport.length });\n        }\n      } catch (err) {\n        log.warn('local LLM synthesis failed; using heuristic fallback', {\n          error: err instanceof Error ? err.message : String(err),\n        });\n      }\n    }\n\n    // Phase 6: Structured brief — populated when internal sampling is\n    // unavailable so the host LLM has well-shaped data to write the report\n    // from without re-reading raw markdown.\n    const comparisonEntities = queryType === 'comparison'\n      ? extractComparisonEntities(input.question).entities\n      : [];\n    const brief = !synthesisResult.samplingUsed\n      ? await buildResearchBrief(\n          input.question,\n          sources,\n          subQueries,\n          PER_SOURCE_CHAR_CAP,\n          TOTAL_SOURCES_CHAR_CAP,\n          queryType,\n          comparisonEntities,\n          localSynthesisText,\n        )\n      : undefined;\n\n    return {\n      report: finalReport,\n      citations: finalCitations,\n      sources,\n      sub_queries: subQueries,\n      depth,\n      total_time_ms: Date.now() - start,\n      sampling_supported: !!server && checkSamplingSupport(server),\n      ...(brief ? { brief } : {}),\n    };\n  } catch (err) {\n    log.error('research pipeline failed', {\n      question: input.question,\n      error: err instanceof Error ? err.message : String(err),\n    });\n    return {\n      report: '',\n      citations: [],\n      sources: [],\n      sub_queries: [],\n      depth,\n      total_time_ms: Date.now() - start,\n      sampling_supported: !!server && checkSamplingSupport(server),\n      error: err instanceof Error ? err.message : String(err),\n    };\n  }\n}\n\ninterface MergedResult {\n  title: string;\n  url: string;\n  snippet: string;\n  relevance_score: number;\n  engines: string[];\n}\n\nasync function fetchSources(\n  merged: MergedResult[],\n  router: SmartRouter,\n  maxSources: number,\n): Promise<ResearchSource[]> {\n  const fetchPromises = merged.slice(0, maxSources).map(async (result): Promise<ResearchSource> => {\n    try {\n      const raw = await Promise.race([\n        router.fetch(result.url, { renderJs: 'auto' }),\n        new Promise<never>((_, reject) =>\n          setTimeout(() => reject(new Error('fetch timeout')), 15000),\n        ),\n      ]);\n\n      const extractor = await getExtractProvider();\n      const extraction = await extractor.extract(raw.html, raw.finalUrl, {\n        maxChars: 30000,\n        contentType: raw.contentType,\n      });\n      const truncated = truncateSmartly(extraction.markdown, PER_SOURCE_CHAR_CAP);\n\n      try {\n        cacheContent(raw, extraction);\n      } catch (err) {\n        log.warn('failed to cache research source', { url: result.url, error: String(err) });\n      }\n\n      try {\n        const embeddingService = getEmbeddingService();\n        if (embeddingService.isAvailable()) {\n          embeddingService.embedAsync(raw.finalUrl, extraction.markdown);\n        }\n      } catch (err) {\n        log.debug('embedding hook skipped for research source', { error: String(err) });\n      }\n\n      return {\n        url: result.url,\n        title: extraction.title || result.title,\n        markdown_content: truncated,\n        relevance_score: result.relevance_score,\n        fetched: true,\n      };\n    } catch (err) {\n      log.debug('failed to fetch research source', {\n        url: result.url,\n        error: err instanceof Error ? err.message : String(err),\n      });\n      return {\n        url: result.url,\n        title: result.title,\n        markdown_content: result.snippet,\n        relevance_score: result.relevance_score,\n        fetched: false,\n        fetch_error: err instanceof Error ? err.message : String(err),\n      };\n    }\n  });\n\n  return Promise.all(fetchPromises);\n}\n\n// Cap total returned markdown_content across sources in relevance order.\n// Later (lower-relevance) sources get trimmed further when budget runs low;\n// any source past the cap is set to empty content (caller still sees url/title).\nfunction applySourceBudget(\n  sources: ResearchSource[],\n  perSourceCap: number,\n  totalCap: number,\n): void {\n  let used = 0;\n  for (const s of sources) {\n    if (!s.markdown_content) continue;\n    if (used >= totalCap) {\n      s.markdown_content = '';\n      continue;\n    }\n    const remaining = totalCap - used;\n    const cap = Math.min(perSourceCap, remaining);\n    if (s.markdown_content.length > cap) {\n      s.markdown_content = truncateSmartly(s.markdown_content, cap);\n    }\n    used += s.markdown_content.length;\n  }\n}\n"],"mappings":"AAAA,SAAS,oBAAoB;AAC7B,SAAS,mBAAoC,iCAAiD;AAC9F,SAAS,wBAAwB;AACjC,SAAS,uBAAuB;AAChC,SAAS,0BAA0B;AACnC,SAAS,0BAA0B;AACnC,SAAS,qBAAqB;AAC9B,SAAS,uBAAuB;AAChC,SAAS,yBAAyB;AAElC,SAAS,0BAA0B;AACnC,SAAS,uBAAuB;AAChC,SAAS,oBAAoB;AAC7B,SAAS,2BAA2B;AACpC,SAAS,4BAAwD;AACjE,SAAS,mBAAmB,yBAAyB;AAUrD,MAAM,MAAM,aAAa,UAAU;AAEnC,MAAM,eAA+F;AAAA,EACnG,OAAO,EAAE,YAAY,GAAG,YAAY,GAAG,YAAY,EAAE;AAAA,EACrD,UAAU,EAAE,YAAY,GAAG,YAAY,IAAI,YAAY,GAAG;AAAA,EAC1D,eAAe,EAAE,YAAY,GAAG,YAAY,IAAI,YAAY,GAAG;AACjE;AAKA,MAAM,yBAAiD;AAAA,EACrD,OAAO;AAAA,EACP,UAAU;AAAA,EACV,eAAe;AACjB;AACA,MAAM,6BAAqD;AAAA,EACzD,OAAO;AAAA,EACP,UAAU;AAAA,EACV,eAAe;AACjB;AAEA,MAAM,sBAAsB;AAC5B,MAAM,yBAAyB;AAE/B,eAAsB,oBACpB,OACA,SACA,QACA,QACyB;AACzB,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,QAAQ,MAAM,SAAS;AAC7B,QAAM,SAAS,aAAa,KAAK,KAAK,aAAa;AACnD,QAAM,aAAa,MAAM,eAAe,OAAO;AAE/C,MAAI;AAEF,QAAI,KAAK,6BAA6B,EAAE,UAAU,MAAM,UAAU,MAAM,CAAC;AACzE,UAAM,kBAAkB,MAAM;AAAA,MAC5B,MAAM;AAAA,MACN;AAAA,MACA;AAAA,IACF;AACA,UAAM,aAAa,gBAAgB;AACnC,UAAM,YAAY,gBAAgB;AAClC,QAAI,KAAK,0BAA0B,EAAE,eAAe,WAAW,QAAQ,cAAc,gBAAgB,cAAc,UAAU,CAAC;AAO9H,UAAM,aAAa,WAAW,UAAU,KAAK,QAAQ,SAAS,IAAI,QAAQ,MAAM,GAAG,CAAC,IAAI;AACxF,UAAM,sBAAsB,KAAK,KAAK,aAAa,WAAW,MAAM,IAAI;AAExE,UAAM,gBAAgB,MAAM;AAAA,MAC1B;AAAA,MACA,OAAO,UAAU,WAAW;AAC1B,cAAM,UAA6B,CAAC;AACpC,cAAM,WAAW,oBAAI,IAAY;AACjC,cAAM,aAAkC;AAAA,UACtC,YAAY;AAAA,UACZ,gBAAgB,MAAM;AAAA,UACtB,gBAAgB,MAAM;AAAA,QACxB;AAEA,cAAM,QAAQ;AAAA,UACZ,WAAW,IAAI,OAAO,WAAW;AAC/B,gBAAI,OAAO,QAAS;AACpB,gBAAI;AACF,oBAAM,KAAK,MAAM,OAAO,OAAO,UAAU,UAAU;AACnD,yBAAW,KAAK,GAAI,SAAQ,KAAK,CAAC;AAClC,uBAAS,IAAI,OAAO,IAAI;AAAA,YAC1B,SAAS,KAAK;AACZ,kBAAI,KAAK,iCAAiC;AAAA,gBACxC,QAAQ,OAAO;AAAA,gBACf,OAAO;AAAA,gBACP,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,cACxD,CAAC;AAAA,YACH;AAAA,UACF,CAAC;AAAA,QACH;AAEA,eAAO,EAAE,SAAS,aAAa,CAAC,GAAG,QAAQ,EAAE;AAAA,MAC/C;AAAA,MACA;AAAA,QACE,eAAe;AAAA,QACf,eAAe,uBAAuB,KAAK,KAAK,uBAAuB;AAAA,QACvE,kBAAkB,2BAA2B,KAAK,KAAK,2BAA2B;AAAA,MACpF;AAAA,IACF;AAEA,UAAM,SAA4B,CAAC;AACnC,UAAM,cAAc,oBAAI,IAAY;AACpC,UAAM,eAAyB,CAAC;AAChC,eAAW,MAAM,eAAe;AAC9B,UAAI,GAAG,MAAM,GAAG,QAAQ;AACtB,eAAO,KAAK,GAAG,GAAG,OAAO,OAAO;AAChC,mBAAW,KAAK,GAAG,OAAO,YAAa,aAAY,IAAI,CAAC;AAAA,MAC1D,WAAW,GAAG,OAAO;AACnB,qBAAa,KAAK,GAAG,GAAG,KAAK,KAAK,GAAG,KAAK,EAAE;AAAA,MAC9C;AAAA,IACF;AACA,QAAI,aAAa,SAAS,GAAG;AAC3B,UAAI,KAAK,kCAAkC,EAAE,QAAQ,aAAa,CAAC;AAAA,IACrE;AAEA,QAAI,KAAK,yBAAyB,EAAE,UAAU,OAAO,QAAQ,SAAS,CAAC,GAAG,WAAW,EAAE,CAAC;AAGxF,QAAI,SAAS,mBAAmB,MAAM;AAEtC,aAAS,gBAAgB,QAAQ;AAAA,MAC/B,gBAAgB,MAAM;AAAA,MACtB,gBAAgB,MAAM;AAAA,IACxB,CAAC;AAED,aAAS,MAAM,cAAc,MAAM,UAAU,MAAM;AACnD,aAAS,OAAO,MAAM,GAAG,UAAU;AAEnC,QAAI,OAAO,WAAW,GAAG;AACvB,aAAO;AAAA,QACL,QAAQ,gBAAgB,MAAM,QAAQ;AAAA;AAAA;AAAA,QACtC,WAAW,CAAC;AAAA,QACZ,SAAS,CAAC;AAAA,QACV,aAAa;AAAA,QACb;AAAA,QACA,eAAe,KAAK,IAAI,IAAI;AAAA,QAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC7D;AAAA,IACF;AAGA,UAAM,UAA4B,MAAM,aAAa,QAAQ,QAAQ,UAAU;AAC/E,sBAAkB,SAAS,qBAAqB,sBAAsB;AACtE,QAAI,KAAK,wBAAwB;AAAA,MAC/B,SAAS,QAAQ,OAAO,CAAC,MAAM,EAAE,OAAO,EAAE;AAAA,MAC1C,QAAQ,QAAQ,OAAO,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE;AAAA,IAC5C,CAAC;AAGD,UAAM,kBAAkB,MAAM;AAAA,MAC5B,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,QAAI,KAAK,sBAAsB,EAAE,cAAc,gBAAgB,cAAc,cAAc,gBAAgB,OAAO,OAAO,CAAC;AAK1H,QAAI,cAAc,gBAAgB;AAClC,QAAI,iBAA6B,gBAAgB;AACjD,QAAI;AACJ,QAAI,CAAC,gBAAgB,gBAAgB,kBAAkB,GAAG;AACxD,UAAI;AACF,cAAM,eAAe,QAClB,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,iBAAiB,SAAS,CAAC,EACxD,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,KAAK,OAAO,EAAE,OAAO,UAAU,EAAE,iBAAiB,EAAE;AAC5E,YAAI,aAAa,SAAS,GAAG;AAC3B,gBAAM,QAAQ,MAAM,gBAAgB,MAAM,UAAU,YAAY;AAChE,wBAAc,MAAM;AACpB,+BAAqB,MAAM;AAC3B,2BAAiB,MAAM,UACpB,OAAO,CAAC,QAAQ,OAAO,KAAK,MAAM,aAAa,MAAM,EACrD,IAAI,CAAC,QAAQ;AACZ,kBAAM,IAAI,aAAa,GAAG;AAC1B,mBAAO;AAAA,cACL,OAAO,MAAM;AAAA,cACb,KAAK,EAAE;AAAA,cACP,OAAO,EAAE;AAAA,cACT,SAAS,EAAE,SAAS,MAAM,GAAG,GAAG;AAAA,YAClC;AAAA,UACF,CAAC;AACH,cAAI,KAAK,6BAA6B,EAAE,cAAc,YAAY,OAAO,CAAC;AAAA,QAC5E;AAAA,MACF,SAAS,KAAK;AACZ,YAAI,KAAK,wDAAwD;AAAA,UAC/D,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,QACxD,CAAC;AAAA,MACH;AAAA,IACF;AAKA,UAAM,qBAAqB,cAAc,eACrC,0BAA0B,MAAM,QAAQ,EAAE,WAC1C,CAAC;AACL,UAAM,QAAQ,CAAC,gBAAgB,eAC3B,MAAM;AAAA,MACJ,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,IACA;AAEJ,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,WAAW;AAAA,MACX;AAAA,MACA,aAAa;AAAA,MACb;AAAA,MACA,eAAe,KAAK,IAAI,IAAI;AAAA,MAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC3D,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,IAC3B;AAAA,EACF,SAAS,KAAK;AACZ,QAAI,MAAM,4BAA4B;AAAA,MACpC,UAAU,MAAM;AAAA,MAChB,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD,CAAC;AACD,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,WAAW,CAAC;AAAA,MACZ,SAAS,CAAC;AAAA,MACV,aAAa,CAAC;AAAA,MACd;AAAA,MACA,eAAe,KAAK,IAAI,IAAI;AAAA,MAC5B,oBAAoB,CAAC,CAAC,UAAU,qBAAqB,MAAM;AAAA,MAC3D,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD;AAAA,EACF;AACF;AAUA,eAAe,aACb,QACA,QACA,YAC2B;AAC3B,QAAM,gBAAgB,OAAO,MAAM,GAAG,UAAU,EAAE,IAAI,OAAO,WAAoC;AAC/F,QAAI;AACF,YAAM,MAAM,MAAM,QAAQ,KAAK;AAAA,QAC7B,OAAO,MAAM,OAAO,KAAK,EAAE,UAAU,OAAO,CAAC;AAAA,QAC7C,IAAI;AAAA,UAAe,CAAC,GAAG,WACrB,WAAW,MAAM,OAAO,IAAI,MAAM,eAAe,CAAC,GAAG,IAAK;AAAA,QAC5D;AAAA,MACF,CAAC;AAED,YAAM,YAAY,MAAM,mBAAmB;AAC3C,YAAM,aAAa,MAAM,UAAU,QAAQ,IAAI,MAAM,IAAI,UAAU;AAAA,QACjE,UAAU;AAAA,QACV,aAAa,IAAI;AAAA,MACnB,CAAC;AACD,YAAM,YAAY,gBAAgB,WAAW,UAAU,mBAAmB;AAE1E,UAAI;AACF,qBAAa,KAAK,UAAU;AAAA,MAC9B,SAAS,KAAK;AACZ,YAAI,KAAK,mCAAmC,EAAE,KAAK,OAAO,KAAK,OAAO,OAAO,GAAG,EAAE,CAAC;AAAA,MACrF;AAEA,UAAI;AACF,cAAM,mBAAmB,oBAAoB;AAC7C,YAAI,iBAAiB,YAAY,GAAG;AAClC,2BAAiB,WAAW,IAAI,UAAU,WAAW,QAAQ;AAAA,QAC/D;AAAA,MACF,SAAS,KAAK;AACZ,YAAI,MAAM,8CAA8C,EAAE,OAAO,OAAO,GAAG,EAAE,CAAC;AAAA,MAChF;AAEA,aAAO;AAAA,QACL,KAAK,OAAO;AAAA,QACZ,OAAO,WAAW,SAAS,OAAO;AAAA,QAClC,kBAAkB;AAAA,QAClB,iBAAiB,OAAO;AAAA,QACxB,SAAS;AAAA,MACX;AAAA,IACF,SAAS,KAAK;AACZ,UAAI,MAAM,mCAAmC;AAAA,QAC3C,KAAK,OAAO;AAAA,QACZ,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AACD,aAAO;AAAA,QACL,KAAK,OAAO;AAAA,QACZ,OAAO,OAAO;AAAA,QACd,kBAAkB,OAAO;AAAA,QACzB,iBAAiB,OAAO;AAAA,QACxB,SAAS;AAAA,QACT,aAAa,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MAC9D;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO,QAAQ,IAAI,aAAa;AAClC;AAKA,SAAS,kBACP,SACA,cACA,UACM;AACN,MAAI,OAAO;AACX,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,EAAE,iBAAkB;AACzB,QAAI,QAAQ,UAAU;AACpB,QAAE,mBAAmB;AACrB;AAAA,IACF;AACA,UAAM,YAAY,WAAW;AAC7B,UAAM,MAAM,KAAK,IAAI,cAAc,SAAS;AAC5C,QAAI,EAAE,iBAAiB,SAAS,KAAK;AACnC,QAAE,mBAAmB,gBAAgB,EAAE,kBAAkB,GAAG;AAAA,IAC9D;AACA,YAAQ,EAAE,iBAAiB;AAAA,EAC7B;AACF;","names":[]}

package/dist/research/synthesis-local.d.ts

@@ -2,6 +2,8 @@ export interface LocalSynthesisOptions {
     maxSources?: number;
     maxCharsPerSource?: number;
     timeoutMs?: number;
+    maxTokens?: number;
+    modelOverride?: string;
 }
 export interface LocalSynthesisSource {
     url: string;
@@ -13,4 +15,5 @@ export interface LocalSynthesisResult {
     citations: number[];
 }
 export declare function synthesizeLocal(question: string, sources: LocalSynthesisSource[], opts?: LocalSynthesisOptions): Promise<LocalSynthesisResult>;
+export declare function isLocalLlmEnabled(): boolean;
 //# sourceMappingURL=synthesis-local.d.ts.map

package/dist/research/synthesis-local.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"synthesis-local.d.ts","sourceRoot":"","sources":["../../src/research/synthesis-local.ts"],"names":[],"mappings":"AASA,MAAM,WAAW,qBAAqB;IACpC,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,SAAS,CAAC,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,oBAAoB;IACnC,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,EAAE,CAAC;CACrB;AAED,wBAAsB,eAAe,CACnC,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE,oBAAoB,EAAE,EAC/B,IAAI,GAAE,qBAA0B,GAC/B,OAAO,CAAC,oBAAoB,CAAC,CA2D/B"}
+{"version":3,"file":"synthesis-local.d.ts","sourceRoot":"","sources":["../../src/research/synthesis-local.ts"],"names":[],"mappings":"AAUA,MAAM,WAAW,qBAAqB;IACpC,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,aAAa,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,MAAM,WAAW,oBAAoB;IACnC,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,EAAE,CAAC;CACrB;AAED,wBAAsB,eAAe,CACnC,QAAQ,EAAE,MAAM,EAChB,OAAO,EAAE,oBAAoB,EAAE,EAC/B,IAAI,GAAE,qBAA0B,GAC/B,OAAO,CAAC,oBAAoB,CAAC,CAkC/B;AAID,wBAAgB,iBAAiB,IAAI,OAAO,CAE3C"}

package/dist/research/synthesis-local.js

@@ -1,24 +1,21 @@
 import { createLogger } from "../logger.js";
-import { isLocalLlmEnabled } from "../extraction/v1/local-llm.js";
+import { isLlmConfigured, runLlmText } from "../integrations/cloud/llm/run.js";
 const log = createLogger("research");
 const DEFAULT_MAX_SOURCES = 8;
 const DEFAULT_MAX_CHARS_PER_SOURCE = 4e3;
 const DEFAULT_TIMEOUT_MS = 6e4;
+const DEFAULT_MAX_TOKENS = 3e3;
 async function synthesizeLocal(question, sources, opts = {}) {
-  if (!isLocalLlmEnabled()) {
-    throw new Error("Local LLM not configured. Set WIGOLO_LLM_PROVIDER.");
+  if (!isLlmConfigured()) {
+    throw new Error("LLM not configured. Set WIGOLO_LLM_PROVIDER or a provider API key.");
   }
   const maxSources = opts.maxSources ?? DEFAULT_MAX_SOURCES;
   const maxCharsPerSource = opts.maxCharsPerSource ?? DEFAULT_MAX_CHARS_PER_SOURCE;
-  const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;
-  const provider = process.env["WIGOLO_LLM_PROVIDER"];
-  const endpoint = provider.includes("/chat/completions") ? provider : provider.replace(/\/+$/, "") + "/v1/chat/completions";
-  const model = process.env["WIGOLO_LLM_MODEL"] ?? "local";
   const sliced = sources.slice(0, maxSources);
   const sourceBlocks = sliced.map((s, i) => {
-    const body2 = s.markdown.length > maxCharsPerSource ? s.markdown.slice(0, maxCharsPerSource) : s.markdown;
+    const body = s.markdown.length > maxCharsPerSource ? s.markdown.slice(0, maxCharsPerSource) : s.markdown;
     return `[${i + 1}] ${s.title}
-${body2}`;
+${body}`;
   });
   const prompt = `You answer questions using ONLY the provided sources. Cite each fact with [N] where N is the source number.
 
@@ -26,31 +23,22 @@ Question: ${question}
 
 Sources:
 ${sourceBlocks.join("\n\n")}`;
-  const body = {
-    model,
-    messages: [{ role: "user", content: prompt }]
-  };
-  let response;
   try {
-    response = await fetch(endpoint, {
-      method: "POST",
-      headers: { "content-type": "application/json" },
-      body: JSON.stringify(body),
-      signal: AbortSignal.timeout(timeoutMs)
+    const result = await runLlmText({
+      prompt,
+      maxTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,
+      modelOverride: opts.modelOverride,
+      timeoutMs: opts.timeoutMs ?? DEFAULT_TIMEOUT_MS
     });
+    log.info("local synthesis ok", { provider: result.provider, model: result.model, latencyMs: result.latencyMs });
+    return { text: result.text, citations: extractCitations(result.text) };
   } catch (err) {
-    log.error("local synthesis request failed", { error: String(err) });
+    log.error("local synthesis request failed", { error: err instanceof Error ? err.message : String(err) });
     throw err;
   }
-  if (!response.ok) {
-    throw new Error(`Local LLM endpoint returned ${response.status}`);
-  }
-  const payload = await response.json();
-  const content = payload.choices?.[0]?.message?.content;
-  if (typeof content !== "string") {
-    throw new Error("Local LLM response missing message content");
-  }
-  return { text: content, citations: extractCitations(content) };
+}
+function isLocalLlmEnabled() {
+  return isLlmConfigured();
 }
 function extractCitations(text) {
   const matches = text.match(/\[(\d+)\]/g);
@@ -68,6 +56,7 @@ function extractCitations(text) {
   return out;
 }
 export {
+  isLocalLlmEnabled,
   synthesizeLocal
 };
 //# sourceMappingURL=synthesis-local.js.map

package/dist/research/synthesis-local.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/research/synthesis-local.ts"],"sourcesContent":["import { createLogger } from '../logger.js';\nimport { isLocalLlmEnabled } from '../extraction/v1/local-llm.js';\n\nconst log = createLogger('research');\n\nconst DEFAULT_MAX_SOURCES = 8;\nconst DEFAULT_MAX_CHARS_PER_SOURCE = 4000;\nconst DEFAULT_TIMEOUT_MS = 60_000;\n\nexport interface LocalSynthesisOptions {\n  maxSources?: number;\n  maxCharsPerSource?: number;\n  timeoutMs?: number;\n}\n\nexport interface LocalSynthesisSource {\n  url: string;\n  title: string;\n  markdown: string;\n}\n\nexport interface LocalSynthesisResult {\n  text: string;\n  citations: number[];\n}\n\nexport async function synthesizeLocal(\n  question: string,\n  sources: LocalSynthesisSource[],\n  opts: LocalSynthesisOptions = {},\n): Promise<LocalSynthesisResult> {\n  if (!isLocalLlmEnabled()) {\n    throw new Error('Local LLM not configured. Set WIGOLO_LLM_PROVIDER.');\n  }\n\n  const maxSources = opts.maxSources ?? DEFAULT_MAX_SOURCES;\n  const maxCharsPerSource = opts.maxCharsPerSource ?? DEFAULT_MAX_CHARS_PER_SOURCE;\n  const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;\n\n  const provider = process.env['WIGOLO_LLM_PROVIDER']!;\n  const endpoint = provider.includes('/chat/completions')\n    ? provider\n    : provider.replace(/\\/+$/, '') + '/v1/chat/completions';\n  const model = process.env['WIGOLO_LLM_MODEL'] ?? 'local';\n\n  const sliced = sources.slice(0, maxSources);\n  const sourceBlocks = sliced.map((s, i) => {\n    const body = s.markdown.length > maxCharsPerSource\n      ? s.markdown.slice(0, maxCharsPerSource)\n      : s.markdown;\n    return `[${i + 1}] ${s.title}\\n${body}`;\n  });\n\n  const prompt =\n    'You answer questions using ONLY the provided sources. Cite each fact with [N] where N is the source number.\\n\\n' +\n    `Question: ${question}\\n\\n` +\n    `Sources:\\n${sourceBlocks.join('\\n\\n')}`;\n\n  const body = {\n    model,\n    messages: [{ role: 'user', content: prompt }],\n  };\n\n  let response: Response;\n  try {\n    response = await fetch(endpoint, {\n      method: 'POST',\n      headers: { 'content-type': 'application/json' },\n      body: JSON.stringify(body),\n      signal: AbortSignal.timeout(timeoutMs),\n    });\n  } catch (err) {\n    log.error('local synthesis request failed', { error: String(err) });\n    throw err;\n  }\n\n  if (!response.ok) {\n    throw new Error(`Local LLM endpoint returned ${response.status}`);\n  }\n\n  const payload = (await response.json()) as {\n    choices?: Array<{ message?: { content?: string } }>;\n  };\n  const content = payload.choices?.[0]?.message?.content;\n  if (typeof content !== 'string') {\n    throw new Error('Local LLM response missing message content');\n  }\n\n  return { text: content, citations: extractCitations(content) };\n}\n\nfunction extractCitations(text: string): number[] {\n  const matches = text.match(/\\[(\\d+)\\]/g);\n  if (!matches) return [];\n  const seen = new Set<number>();\n  const out: number[] = [];\n  for (const m of matches) {\n    const n = Number(m.slice(1, -1));\n    if (!Number.isFinite(n) || n < 1) continue;\n    const idx = n - 1;\n    if (seen.has(idx)) continue;\n    seen.add(idx);\n    out.push(idx);\n  }\n  return out;\n}\n"],"mappings":"AAAA,SAAS,oBAAoB;AAC7B,SAAS,yBAAyB;AAElC,MAAM,MAAM,aAAa,UAAU;AAEnC,MAAM,sBAAsB;AAC5B,MAAM,+BAA+B;AACrC,MAAM,qBAAqB;AAmB3B,eAAsB,gBACpB,UACA,SACA,OAA8B,CAAC,GACA;AAC/B,MAAI,CAAC,kBAAkB,GAAG;AACxB,UAAM,IAAI,MAAM,oDAAoD;AAAA,EACtE;AAEA,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,oBAAoB,KAAK,qBAAqB;AACpD,QAAM,YAAY,KAAK,aAAa;AAEpC,QAAM,WAAW,QAAQ,IAAI,qBAAqB;AAClD,QAAM,WAAW,SAAS,SAAS,mBAAmB,IAClD,WACA,SAAS,QAAQ,QAAQ,EAAE,IAAI;AACnC,QAAM,QAAQ,QAAQ,IAAI,kBAAkB,KAAK;AAEjD,QAAM,SAAS,QAAQ,MAAM,GAAG,UAAU;AAC1C,QAAM,eAAe,OAAO,IAAI,CAAC,GAAG,MAAM;AACxC,UAAMA,QAAO,EAAE,SAAS,SAAS,oBAC7B,EAAE,SAAS,MAAM,GAAG,iBAAiB,IACrC,EAAE;AACN,WAAO,IAAI,IAAI,CAAC,KAAK,EAAE,KAAK;AAAA,EAAKA,KAAI;AAAA,EACvC,CAAC;AAED,QAAM,SACJ;AAAA;AAAA,YACa,QAAQ;AAAA;AAAA;AAAA,EACR,aAAa,KAAK,MAAM,CAAC;AAExC,QAAM,OAAO;AAAA,IACX;AAAA,IACA,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,OAAO,CAAC;AAAA,EAC9C;AAEA,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,MAAM,UAAU;AAAA,MAC/B,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,MAC9C,MAAM,KAAK,UAAU,IAAI;AAAA,MACzB,QAAQ,YAAY,QAAQ,SAAS;AAAA,IACvC,CAAC;AAAA,EACH,SAAS,KAAK;AACZ,QAAI,MAAM,kCAAkC,EAAE,OAAO,OAAO,GAAG,EAAE,CAAC;AAClE,UAAM;AAAA,EACR;AAEA,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,IAAI,MAAM,+BAA+B,SAAS,MAAM,EAAE;AAAA,EAClE;AAEA,QAAM,UAAW,MAAM,SAAS,KAAK;AAGrC,QAAM,UAAU,QAAQ,UAAU,CAAC,GAAG,SAAS;AAC/C,MAAI,OAAO,YAAY,UAAU;AAC/B,UAAM,IAAI,MAAM,4CAA4C;AAAA,EAC9D;AAEA,SAAO,EAAE,MAAM,SAAS,WAAW,iBAAiB,OAAO,EAAE;AAC/D;AAEA,SAAS,iBAAiB,MAAwB;AAChD,QAAM,UAAU,KAAK,MAAM,YAAY;AACvC,MAAI,CAAC,QAAS,QAAO,CAAC;AACtB,QAAM,OAAO,oBAAI,IAAY;AAC7B,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,SAAS;AACvB,UAAM,IAAI,OAAO,EAAE,MAAM,GAAG,EAAE,CAAC;AAC/B,QAAI,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,EAAG;AAClC,UAAM,MAAM,IAAI;AAChB,QAAI,KAAK,IAAI,GAAG,EAAG;AACnB,SAAK,IAAI,GAAG;AACZ,QAAI,KAAK,GAAG;AAAA,EACd;AACA,SAAO;AACT;","names":["body"]}
+{"version":3,"sources":["../../src/research/synthesis-local.ts"],"sourcesContent":["import { createLogger } from '../logger.js';\nimport { isLlmConfigured, runLlmText } from '../integrations/cloud/llm/run.js';\n\nconst log = createLogger('research');\n\nconst DEFAULT_MAX_SOURCES = 8;\nconst DEFAULT_MAX_CHARS_PER_SOURCE = 4000;\nconst DEFAULT_TIMEOUT_MS = 60_000;\nconst DEFAULT_MAX_TOKENS = 3000;\n\nexport interface LocalSynthesisOptions {\n  maxSources?: number;\n  maxCharsPerSource?: number;\n  timeoutMs?: number;\n  maxTokens?: number;\n  modelOverride?: string;\n}\n\nexport interface LocalSynthesisSource {\n  url: string;\n  title: string;\n  markdown: string;\n}\n\nexport interface LocalSynthesisResult {\n  text: string;\n  citations: number[];\n}\n\nexport async function synthesizeLocal(\n  question: string,\n  sources: LocalSynthesisSource[],\n  opts: LocalSynthesisOptions = {},\n): Promise<LocalSynthesisResult> {\n  if (!isLlmConfigured()) {\n    throw new Error('LLM not configured. Set WIGOLO_LLM_PROVIDER or a provider API key.');\n  }\n\n  const maxSources = opts.maxSources ?? DEFAULT_MAX_SOURCES;\n  const maxCharsPerSource = opts.maxCharsPerSource ?? DEFAULT_MAX_CHARS_PER_SOURCE;\n\n  const sliced = sources.slice(0, maxSources);\n  const sourceBlocks = sliced.map((s, i) => {\n    const body = s.markdown.length > maxCharsPerSource\n      ? s.markdown.slice(0, maxCharsPerSource)\n      : s.markdown;\n    return `[${i + 1}] ${s.title}\\n${body}`;\n  });\n\n  const prompt =\n    'You answer questions using ONLY the provided sources. Cite each fact with [N] where N is the source number.\\n\\n' +\n    `Question: ${question}\\n\\n` +\n    `Sources:\\n${sourceBlocks.join('\\n\\n')}`;\n\n  try {\n    const result = await runLlmText({\n      prompt,\n      maxTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      modelOverride: opts.modelOverride,\n      timeoutMs: opts.timeoutMs ?? DEFAULT_TIMEOUT_MS,\n    });\n    log.info('local synthesis ok', { provider: result.provider, model: result.model, latencyMs: result.latencyMs });\n    return { text: result.text, citations: extractCitations(result.text) };\n  } catch (err) {\n    log.error('local synthesis request failed', { error: err instanceof Error ? err.message : String(err) });\n    throw err;\n  }\n}\n\n// Backwards-compat shim — callers (research/pipeline.ts) used isLocalLlmEnabled()\n// to gate this fallback. Keep the same gate name pointing at the unified runner.\nexport function isLocalLlmEnabled(): boolean {\n  return isLlmConfigured();\n}\n\nfunction extractCitations(text: string): number[] {\n  const matches = text.match(/\\[(\\d+)\\]/g);\n  if (!matches) return [];\n  const seen = new Set<number>();\n  const out: number[] = [];\n  for (const m of matches) {\n    const n = Number(m.slice(1, -1));\n    if (!Number.isFinite(n) || n < 1) continue;\n    const idx = n - 1;\n    if (seen.has(idx)) continue;\n    seen.add(idx);\n    out.push(idx);\n  }\n  return out;\n}\n"],"mappings":"AAAA,SAAS,oBAAoB;AAC7B,SAAS,iBAAiB,kBAAkB;AAE5C,MAAM,MAAM,aAAa,UAAU;AAEnC,MAAM,sBAAsB;AAC5B,MAAM,+BAA+B;AACrC,MAAM,qBAAqB;AAC3B,MAAM,qBAAqB;AAqB3B,eAAsB,gBACpB,UACA,SACA,OAA8B,CAAC,GACA;AAC/B,MAAI,CAAC,gBAAgB,GAAG;AACtB,UAAM,IAAI,MAAM,oEAAoE;AAAA,EACtF;AAEA,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,oBAAoB,KAAK,qBAAqB;AAEpD,QAAM,SAAS,QAAQ,MAAM,GAAG,UAAU;AAC1C,QAAM,eAAe,OAAO,IAAI,CAAC,GAAG,MAAM;AACxC,UAAM,OAAO,EAAE,SAAS,SAAS,oBAC7B,EAAE,SAAS,MAAM,GAAG,iBAAiB,IACrC,EAAE;AACN,WAAO,IAAI,IAAI,CAAC,KAAK,EAAE,KAAK;AAAA,EAAK,IAAI;AAAA,EACvC,CAAC;AAED,QAAM,SACJ;AAAA;AAAA,YACa,QAAQ;AAAA;AAAA;AAAA,EACR,aAAa,KAAK,MAAM,CAAC;AAExC,MAAI;AACF,UAAM,SAAS,MAAM,WAAW;AAAA,MAC9B;AAAA,MACA,WAAW,KAAK,aAAa;AAAA,MAC7B,eAAe,KAAK;AAAA,MACpB,WAAW,KAAK,aAAa;AAAA,IAC/B,CAAC;AACD,QAAI,KAAK,sBAAsB,EAAE,UAAU,OAAO,UAAU,OAAO,OAAO,OAAO,WAAW,OAAO,UAAU,CAAC;AAC9G,WAAO,EAAE,MAAM,OAAO,MAAM,WAAW,iBAAiB,OAAO,IAAI,EAAE;AAAA,EACvE,SAAS,KAAK;AACZ,QAAI,MAAM,kCAAkC,EAAE,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAE,CAAC;AACvG,UAAM;AAAA,EACR;AACF;AAIO,SAAS,oBAA6B;AAC3C,SAAO,gBAAgB;AACzB;AAEA,SAAS,iBAAiB,MAAwB;AAChD,QAAM,UAAU,KAAK,MAAM,YAAY;AACvC,MAAI,CAAC,QAAS,QAAO,CAAC;AACtB,QAAM,OAAO,oBAAI,IAAY;AAC7B,QAAM,MAAgB,CAAC;AACvB,aAAW,KAAK,SAAS;AACvB,UAAM,IAAI,OAAO,EAAE,MAAM,GAAG,EAAE,CAAC;AAC/B,QAAI,CAAC,OAAO,SAAS,CAAC,KAAK,IAAI,EAAG;AAClC,UAAM,MAAM,IAAI;AAChB,QAAI,KAAK,IAAI,GAAG,EAAG;AACnB,SAAK,IAAI,GAAG;AACZ,QAAI,KAAK,GAAG;AAAA,EACd;AACA,SAAO;AACT;","names":[]}

package/dist/search/filters.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"filters.d.ts","sourceRoot":"","sources":["../../src/search/filters.ts"],"names":[],"mappings":"AAiBA,wBAAgB,eAAe,CAAC,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,EACvD,OAAO,EAAE,CAAC,EAAE,EACZ,cAAc,CAAC,EAAE,MAAM,EAAE,EACzB,cAAc,CAAC,EAAE,MAAM,EAAE,GACxB,CAAC,EAAE,CAgBL;AAOD,wBAAgB,iBAAiB,CAAC,CAAC,EACjC,OAAO,EAAE,CAAC,EAAE,EACZ,QAAQ,CAAC,EAAE,MAAM,EACjB,MAAM,CAAC,EAAE,MAAM,GACd,CAAC,EAAE,CAYL;AAED,wBAAgB,gBAAgB,CAAC,CAAC,EAChC,OAAO,EAAE,CAAC,EAAE,EACZ,SAAS,CAAC,EAAE,MAAM,GACjB,CAAC,EAAE,CAGL;AAED,MAAM,WAAW,aAAa;IAC5B,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,wBAAgB,eAAe,CAAC,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,EACvD,OAAO,EAAE,CAAC,EAAE,EACZ,OAAO,EAAE,aAAa,GACrB,CAAC,EAAE,CAKL"}
+{"version":3,"file":"filters.d.ts","sourceRoot":"","sources":["../../src/search/filters.ts"],"names":[],"mappings":"AAiBA,wBAAgB,eAAe,CAAC,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,EACvD,OAAO,EAAE,CAAC,EAAE,EACZ,cAAc,CAAC,EAAE,MAAM,EAAE,EACzB,cAAc,CAAC,EAAE,MAAM,EAAE,GACxB,CAAC,EAAE,CAgBL;AAOD,wBAAgB,iBAAiB,CAAC,CAAC,EACjC,OAAO,EAAE,CAAC,EAAE,EACZ,QAAQ,CAAC,EAAE,MAAM,EACjB,MAAM,CAAC,EAAE,MAAM,GACd,CAAC,EAAE,CAyBL;AAED,wBAAgB,gBAAgB,CAAC,CAAC,EAChC,OAAO,EAAE,CAAC,EAAE,EACZ,SAAS,CAAC,EAAE,MAAM,GACjB,CAAC,EAAE,CAGL;AAED,MAAM,WAAW,aAAa;IAC5B,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,wBAAgB,eAAe,CAAC,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,CAAA;CAAE,EACvD,OAAO,EAAE,CAAC,EAAE,EACZ,OAAO,EAAE,aAAa,GACrB,CAAC,EAAE,CAKL"}

package/dist/search/filters.js

@@ -37,7 +37,17 @@ function filterByDateRange(results, fromDate, toDate) {
   if (fromDate && !isValidIsoDate(fromDate)) return results;
   if (toDate && !isValidIsoDate(toDate)) return results;
   if (fromDate && toDate && new Date(fromDate) > new Date(toDate)) return results;
-  return results;
+  const fromMs = fromDate ? new Date(fromDate).getTime() : null;
+  const toMs = toDate ? new Date(toDate).getTime() + 24 * 3600 * 1e3 - 1 : null;
+  return results.filter((r) => {
+    const published = r.published_date;
+    if (typeof published !== "string" || !published) return true;
+    const t = Date.parse(published);
+    if (isNaN(t)) return true;
+    if (fromMs !== null && t < fromMs) return false;
+    if (toMs !== null && t > toMs) return false;
+    return true;
+  });
 }
 function filterByCategory(results, _category) {
   return results;

package/dist/search/filters.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/search/filters.ts"],"sourcesContent":["function getDomain(url: string): string {\n  try {\n    return new URL(url).hostname.toLowerCase();\n  } catch {\n    return '';\n  }\n}\n\nfunction normalizeDomain(domain: string): string {\n  return domain.replace(/\\/+$/, '').toLowerCase();\n}\n\nfunction domainMatches(hostname: string, domain: string): boolean {\n  const normalized = normalizeDomain(domain);\n  return hostname === normalized || hostname.endsWith('.' + normalized);\n}\n\nexport function filterByDomains<T extends { url: string }>(\n  results: T[],\n  includeDomains?: string[],\n  excludeDomains?: string[],\n): T[] {\n  if (!includeDomains?.length && !excludeDomains?.length) return results;\n\n  return results.filter((r) => {\n    const hostname = getDomain(r.url);\n    if (!hostname) {\n      return !includeDomains?.length;\n    }\n    if (includeDomains?.length) {\n      if (!includeDomains.some((d) => domainMatches(hostname, d))) return false;\n    }\n    if (excludeDomains?.length) {\n      if (excludeDomains.some((d) => domainMatches(hostname, d))) return false;\n    }\n    return true;\n  });\n}\n\nfunction isValidIsoDate(dateStr: string): boolean {\n  const parsed = new Date(dateStr);\n  return !isNaN(parsed.getTime()) && /^\\d{4}-\\d{2}-\\d{2}$/.test(dateStr);\n}\n\nexport function filterByDateRange<T>(\n  results: T[],\n  fromDate?: string,\n  toDate?: string,\n): T[] {\n  if (!fromDate && !toDate) return results;\n\n  if (fromDate && !isValidIsoDate(fromDate)) return results;\n  if (toDate && !isValidIsoDate(toDate)) return results;\n\n  if (fromDate && toDate && new Date(fromDate) > new Date(toDate)) return results;\n\n  // Date filtering is best-effort on direct scraping engines.\n  // SearXNG handles dates natively via time_range. For fallback engines,\n  // snippet text doesn't reliably contain dates, so we keep all results.\n  return results;\n}\n\nexport function filterByCategory<T>(\n  results: T[],\n  _category?: string,\n): T[] {\n  // Category filtering is handled by SearXNG natively.\n  return results;\n}\n\nexport interface FilterOptions {\n  includeDomains?: string[];\n  excludeDomains?: string[];\n  fromDate?: string;\n  toDate?: string;\n  category?: string;\n}\n\nexport function applyAllFilters<T extends { url: string }>(\n  results: T[],\n  options: FilterOptions,\n): T[] {\n  let filtered = filterByDomains(results, options.includeDomains, options.excludeDomains);\n  filtered = filterByDateRange(filtered, options.fromDate, options.toDate);\n  filtered = filterByCategory(filtered, options.category);\n  return filtered;\n}\n"],"mappings":"AAAA,SAAS,UAAU,KAAqB;AACtC,MAAI;AACF,WAAO,IAAI,IAAI,GAAG,EAAE,SAAS,YAAY;AAAA,EAC3C,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,gBAAgB,QAAwB;AAC/C,SAAO,OAAO,QAAQ,QAAQ,EAAE,EAAE,YAAY;AAChD;AAEA,SAAS,cAAc,UAAkB,QAAyB;AAChE,QAAM,aAAa,gBAAgB,MAAM;AACzC,SAAO,aAAa,cAAc,SAAS,SAAS,MAAM,UAAU;AACtE;AAEO,SAAS,gBACd,SACA,gBACA,gBACK;AACL,MAAI,CAAC,gBAAgB,UAAU,CAAC,gBAAgB,OAAQ,QAAO;AAE/D,SAAO,QAAQ,OAAO,CAAC,MAAM;AAC3B,UAAM,WAAW,UAAU,EAAE,GAAG;AAChC,QAAI,CAAC,UAAU;AACb,aAAO,CAAC,gBAAgB;AAAA,IAC1B;AACA,QAAI,gBAAgB,QAAQ;AAC1B,UAAI,CAAC,eAAe,KAAK,CAAC,MAAM,cAAc,UAAU,CAAC,CAAC,EAAG,QAAO;AAAA,IACtE;AACA,QAAI,gBAAgB,QAAQ;AAC1B,UAAI,eAAe,KAAK,CAAC,MAAM,cAAc,UAAU,CAAC,CAAC,EAAG,QAAO;AAAA,IACrE;AACA,WAAO;AAAA,EACT,CAAC;AACH;AAEA,SAAS,eAAe,SAA0B;AAChD,QAAM,SAAS,IAAI,KAAK,OAAO;AAC/B,SAAO,CAAC,MAAM,OAAO,QAAQ,CAAC,KAAK,sBAAsB,KAAK,OAAO;AACvE;AAEO,SAAS,kBACd,SACA,UACA,QACK;AACL,MAAI,CAAC,YAAY,CAAC,OAAQ,QAAO;AAEjC,MAAI,YAAY,CAAC,eAAe,QAAQ,EAAG,QAAO;AAClD,MAAI,UAAU,CAAC,eAAe,MAAM,EAAG,QAAO;AAE9C,MAAI,YAAY,UAAU,IAAI,KAAK,QAAQ,IAAI,IAAI,KAAK,MAAM,EAAG,QAAO;AAKxE,SAAO;AACT;AAEO,SAAS,iBACd,SACA,WACK;AAEL,SAAO;AACT;AAUO,SAAS,gBACd,SACA,SACK;AACL,MAAI,WAAW,gBAAgB,SAAS,QAAQ,gBAAgB,QAAQ,cAAc;AACtF,aAAW,kBAAkB,UAAU,QAAQ,UAAU,QAAQ,MAAM;AACvE,aAAW,iBAAiB,UAAU,QAAQ,QAAQ;AACtD,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../../src/search/filters.ts"],"sourcesContent":["function getDomain(url: string): string {\n  try {\n    return new URL(url).hostname.toLowerCase();\n  } catch {\n    return '';\n  }\n}\n\nfunction normalizeDomain(domain: string): string {\n  return domain.replace(/\\/+$/, '').toLowerCase();\n}\n\nfunction domainMatches(hostname: string, domain: string): boolean {\n  const normalized = normalizeDomain(domain);\n  return hostname === normalized || hostname.endsWith('.' + normalized);\n}\n\nexport function filterByDomains<T extends { url: string }>(\n  results: T[],\n  includeDomains?: string[],\n  excludeDomains?: string[],\n): T[] {\n  if (!includeDomains?.length && !excludeDomains?.length) return results;\n\n  return results.filter((r) => {\n    const hostname = getDomain(r.url);\n    if (!hostname) {\n      return !includeDomains?.length;\n    }\n    if (includeDomains?.length) {\n      if (!includeDomains.some((d) => domainMatches(hostname, d))) return false;\n    }\n    if (excludeDomains?.length) {\n      if (excludeDomains.some((d) => domainMatches(hostname, d))) return false;\n    }\n    return true;\n  });\n}\n\nfunction isValidIsoDate(dateStr: string): boolean {\n  const parsed = new Date(dateStr);\n  return !isNaN(parsed.getTime()) && /^\\d{4}-\\d{2}-\\d{2}$/.test(dateStr);\n}\n\nexport function filterByDateRange<T>(\n  results: T[],\n  fromDate?: string,\n  toDate?: string,\n): T[] {\n  if (!fromDate && !toDate) return results;\n\n  if (fromDate && !isValidIsoDate(fromDate)) return results;\n  if (toDate && !isValidIsoDate(toDate)) return results;\n\n  if (fromDate && toDate && new Date(fromDate) > new Date(toDate)) return results;\n\n  const fromMs = fromDate ? new Date(fromDate).getTime() : null;\n  // toDate inclusive: treat as end-of-day so 'to=2026-01-31' keeps anything stamped Jan 31.\n  const toMs = toDate ? new Date(toDate).getTime() + 24 * 3600 * 1000 - 1 : null;\n\n  // Drop results with a published_date outside the window. Results without a\n  // published_date pass through — SearXNG and most fallback engines do not\n  // expose reliable dates per result, and the user's request for recency is\n  // already biased via time_range on the upstream call.\n  return results.filter((r) => {\n    const published = (r as { published_date?: unknown }).published_date;\n    if (typeof published !== 'string' || !published) return true;\n    const t = Date.parse(published);\n    if (isNaN(t)) return true;\n    if (fromMs !== null && t < fromMs) return false;\n    if (toMs !== null && t > toMs) return false;\n    return true;\n  });\n}\n\nexport function filterByCategory<T>(\n  results: T[],\n  _category?: string,\n): T[] {\n  // Category filtering is handled by SearXNG natively.\n  return results;\n}\n\nexport interface FilterOptions {\n  includeDomains?: string[];\n  excludeDomains?: string[];\n  fromDate?: string;\n  toDate?: string;\n  category?: string;\n}\n\nexport function applyAllFilters<T extends { url: string }>(\n  results: T[],\n  options: FilterOptions,\n): T[] {\n  let filtered = filterByDomains(results, options.includeDomains, options.excludeDomains);\n  filtered = filterByDateRange(filtered, options.fromDate, options.toDate);\n  filtered = filterByCategory(filtered, options.category);\n  return filtered;\n}\n"],"mappings":"AAAA,SAAS,UAAU,KAAqB;AACtC,MAAI;AACF,WAAO,IAAI,IAAI,GAAG,EAAE,SAAS,YAAY;AAAA,EAC3C,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,gBAAgB,QAAwB;AAC/C,SAAO,OAAO,QAAQ,QAAQ,EAAE,EAAE,YAAY;AAChD;AAEA,SAAS,cAAc,UAAkB,QAAyB;AAChE,QAAM,aAAa,gBAAgB,MAAM;AACzC,SAAO,aAAa,cAAc,SAAS,SAAS,MAAM,UAAU;AACtE;AAEO,SAAS,gBACd,SACA,gBACA,gBACK;AACL,MAAI,CAAC,gBAAgB,UAAU,CAAC,gBAAgB,OAAQ,QAAO;AAE/D,SAAO,QAAQ,OAAO,CAAC,MAAM;AAC3B,UAAM,WAAW,UAAU,EAAE,GAAG;AAChC,QAAI,CAAC,UAAU;AACb,aAAO,CAAC,gBAAgB;AAAA,IAC1B;AACA,QAAI,gBAAgB,QAAQ;AAC1B,UAAI,CAAC,eAAe,KAAK,CAAC,MAAM,cAAc,UAAU,CAAC,CAAC,EAAG,QAAO;AAAA,IACtE;AACA,QAAI,gBAAgB,QAAQ;AAC1B,UAAI,eAAe,KAAK,CAAC,MAAM,cAAc,UAAU,CAAC,CAAC,EAAG,QAAO;AAAA,IACrE;AACA,WAAO;AAAA,EACT,CAAC;AACH;AAEA,SAAS,eAAe,SAA0B;AAChD,QAAM,SAAS,IAAI,KAAK,OAAO;AAC/B,SAAO,CAAC,MAAM,OAAO,QAAQ,CAAC,KAAK,sBAAsB,KAAK,OAAO;AACvE;AAEO,SAAS,kBACd,SACA,UACA,QACK;AACL,MAAI,CAAC,YAAY,CAAC,OAAQ,QAAO;AAEjC,MAAI,YAAY,CAAC,eAAe,QAAQ,EAAG,QAAO;AAClD,MAAI,UAAU,CAAC,eAAe,MAAM,EAAG,QAAO;AAE9C,MAAI,YAAY,UAAU,IAAI,KAAK,QAAQ,IAAI,IAAI,KAAK,MAAM,EAAG,QAAO;AAExE,QAAM,SAAS,WAAW,IAAI,KAAK,QAAQ,EAAE,QAAQ,IAAI;AAEzD,QAAM,OAAO,SAAS,IAAI,KAAK,MAAM,EAAE,QAAQ,IAAI,KAAK,OAAO,MAAO,IAAI;AAM1E,SAAO,QAAQ,OAAO,CAAC,MAAM;AAC3B,UAAM,YAAa,EAAmC;AACtD,QAAI,OAAO,cAAc,YAAY,CAAC,UAAW,QAAO;AACxD,UAAM,IAAI,KAAK,MAAM,SAAS;AAC9B,QAAI,MAAM,CAAC,EAAG,QAAO;AACrB,QAAI,WAAW,QAAQ,IAAI,OAAQ,QAAO;AAC1C,QAAI,SAAS,QAAQ,IAAI,KAAM,QAAO;AACtC,WAAO;AAAA,EACT,CAAC;AACH;AAEO,SAAS,iBACd,SACA,WACK;AAEL,SAAO;AACT;AAUO,SAAS,gBACd,SACA,SACK;AACL,MAAI,WAAW,gBAAgB,SAAS,QAAQ,gBAAgB,QAAQ,cAAc;AACtF,aAAW,kBAAkB,UAAU,QAAQ,UAAU,QAAQ,MAAM;AACvE,aAAW,iBAAiB,UAAU,QAAQ,QAAQ;AACtD,SAAO;AACT;","names":[]}

package/dist/search/reranker/transformers-rerank-provider.d.ts

@@ -8,5 +8,6 @@ export declare class TransformersRerankProvider implements RerankProvider {
     warmup(): Promise<void>;
     private load;
     rerank(query: string, candidates: RerankCandidate[], topK?: number): Promise<RerankResult[]>;
+    dispose(): Promise<void>;
 }
 //# sourceMappingURL=transformers-rerank-provider.d.ts.map

package/dist/search/reranker/transformers-rerank-provider.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"transformers-rerank-provider.d.ts","sourceRoot":"","sources":["../../../src/search/reranker/transformers-rerank-provider.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EACV,cAAc,EACd,eAAe,EACf,YAAY,EACb,MAAM,oCAAoC,CAAC;AAwC5C,qBAAa,0BAA2B,YAAW,cAAc;IAC/D,OAAO,CAAC,SAAS,CAA0B;IAC3C,OAAO,CAAC,KAAK,CAAsB;IACnC,OAAO,CAAC,WAAW,CAAgE;IACnF,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;;IAMnB,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAI7B,OAAO,CAAC,IAAI;IA6BN,MAAM,CACV,KAAK,EAAE,MAAM,EACb,UAAU,EAAE,eAAe,EAAE,EAC7B,IAAI,SAAoB,GACvB,OAAO,CAAC,YAAY,EAAE,CAAC;CA8B3B"}
+{"version":3,"file":"transformers-rerank-provider.d.ts","sourceRoot":"","sources":["../../../src/search/reranker/transformers-rerank-provider.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EACV,cAAc,EACd,eAAe,EACf,YAAY,EACb,MAAM,oCAAoC,CAAC;AAwC5C,qBAAa,0BAA2B,YAAW,cAAc;IAC/D,OAAO,CAAC,SAAS,CAA0B;IAC3C,OAAO,CAAC,KAAK,CAAsB;IACnC,OAAO,CAAC,WAAW,CAAgE;IACnF,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;;IAMnB,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAI7B,OAAO,CAAC,IAAI;IA6BN,MAAM,CACV,KAAK,EAAE,MAAM,EACb,UAAU,EAAE,eAAe,EAAE,EAC7B,IAAI,SAAoB,GACvB,OAAO,CAAC,YAAY,EAAE,CAAC;IAkCpB,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAa/B"}

package/dist/search/reranker/transformers-rerank-provider.js

@@ -71,6 +71,22 @@ class TransformersRerankProvider {
     }));
     return scored.sort((a, b) => b.score - a.score).slice(0, topK);
   }
+  // Release the underlying ONNX session before process exit. Without this,
+  // the runtime's worker threads race during C++ destructor teardown and
+  // surface as `mutex lock failed: Invalid argument` on macOS.
+  async dispose() {
+    const model = this.model;
+    this.model = null;
+    this.tokenizer = null;
+    this.loadPromise = null;
+    if (!model) return;
+    try {
+      const m = model;
+      if (typeof m.dispose === "function") await m.dispose();
+    } catch (err) {
+      log.debug("reranker dispose failed", { error: err instanceof Error ? err.message : String(err) });
+    }
+  }
 }
 export {
   TransformersRerankProvider

package/dist/search/reranker/transformers-rerank-provider.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../src/search/reranker/transformers-rerank-provider.ts"],"sourcesContent":["import { join } from 'node:path';\nimport {\n  AutoTokenizer,\n  AutoModelForSequenceClassification,\n  env,\n} from '@huggingface/transformers';\nimport type {\n  RerankProvider,\n  RerankCandidate,\n  RerankResult,\n} from '../../providers/rerank-provider.js';\nimport { createLogger } from '../../logger.js';\nimport { getConfig } from '../../config.js';\n\nconst log = createLogger('reranker');\n\n// Cross-encoder reranker via Transformers.js.\n//\n// The high-level `pipeline('text-classification', ...)` API does not pass\n// `text_pair`, so it can't drive a cross-encoder properly. We therefore\n// load the tokenizer + sequence-classification model directly: feed\n// (query, document) pairs to the tokenizer and read raw logits from the\n// model. ms-marco-MiniLM-L-6-v2 is a single-output regressor (num_labels=1)\n// where higher logit = more relevant, so the logit is used as the rerank\n// score with no further transform.\ntype Tokenizer = Awaited<ReturnType<typeof AutoTokenizer.from_pretrained>>;\ntype Model = Awaited<ReturnType<typeof AutoModelForSequenceClassification.from_pretrained>>;\n\ninterface LogitsTensor {\n  data: ArrayLike<number>;\n  dims: number[];\n}\n\n// Recognize the noisy huggingface fetch failure signature and replace it\n// with an actionable instruction. Transformers.js parses a config that\n// failed to download, then dereferences `tokenizer_class` on undefined.\nfunction wrapLoadError(err: unknown): Error {\n  const message = err instanceof Error ? err.message : String(err);\n  const looksLikeMissingModel =\n    /tokenizer_class|tokenizer_config|preprocessor_config|fetch failed|ENOTFOUND|ECONNREFUSED|ETIMEDOUT|ENETUNREACH/i.test(\n      message,\n    );\n  if (looksLikeMissingModel) {\n    return new Error(\n      `Reranker model not downloaded — run \\`wigolo warmup\\` (cause: ${message})`,\n    );\n  }\n  return new Error(`Failed to load reranker model: ${message}`);\n}\n\nexport class TransformersRerankProvider implements RerankProvider {\n  private tokenizer: Tokenizer | null = null;\n  private model: Model | null = null;\n  private loadPromise: Promise<{ tokenizer: Tokenizer; model: Model }> | null = null;\n  readonly modelId: string;\n\n  constructor() {\n    this.modelId = 'Xenova/ms-marco-MiniLM-L-6-v2';\n  }\n\n  async warmup(): Promise<void> {\n    await this.load();\n  }\n\n  private load(): Promise<{ tokenizer: Tokenizer; model: Model }> {\n    if (this.tokenizer && this.model) {\n      return Promise.resolve({ tokenizer: this.tokenizer, model: this.model });\n    }\n    if (this.loadPromise) return this.loadPromise;\n\n    log.info('Loading rerank model', { modelId: this.modelId });\n    const cacheDir = join(getConfig().dataDir, 'transformers');\n    // Direct the library at a writable cache under the wigolo data dir so\n    // models don't end up in a user home cache the daemon can't manage.\n    env.cacheDir = cacheDir;\n\n    this.loadPromise = Promise.all([\n      AutoTokenizer.from_pretrained(this.modelId),\n      AutoModelForSequenceClassification.from_pretrained(this.modelId),\n    ])\n      .then(([tokenizer, model]) => {\n        this.tokenizer = tokenizer;\n        this.model = model;\n        return { tokenizer, model };\n      })\n      .catch((err: unknown) => {\n        this.loadPromise = null;\n        throw wrapLoadError(err);\n      });\n\n    return this.loadPromise;\n  }\n\n  async rerank(\n    query: string,\n    candidates: RerankCandidate[],\n    topK = candidates.length,\n  ): Promise<RerankResult[]> {\n    if (candidates.length === 0) return [];\n\n    const { tokenizer, model } = await this.load();\n\n    // Build batch: query repeated against each document.\n    const queries = candidates.map(() => query);\n    const docs = candidates.map((c) => c.text);\n\n    const inputs = tokenizer(queries, {\n      text_pair: docs,\n      padding: true,\n      truncation: true,\n    });\n\n    const outputs = (await model(inputs)) as { logits: LogitsTensor };\n    const logits = outputs.logits;\n    // logits shape is [batch, 1] for single-label regression rerankers.\n    // For multi-label heads (rare for rerankers) we still take the first\n    // value as the relevance score.\n    const stride = logits.dims.length >= 2 ? logits.dims[1] : 1;\n    const data = logits.data;\n\n    const scored: RerankResult[] = candidates.map((c, i) => ({\n      id: c.id,\n      score: Number(data[i * stride]),\n    }));\n\n    return scored.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n}\n"],"mappings":"AAAA,SAAS,YAAY;AACrB;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAMP,SAAS,oBAAoB;AAC7B,SAAS,iBAAiB;AAE1B,MAAM,MAAM,aAAa,UAAU;AAsBnC,SAAS,cAAc,KAAqB;AAC1C,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,QAAM,wBACJ,kHAAkH;AAAA,IAChH;AAAA,EACF;AACF,MAAI,uBAAuB;AACzB,WAAO,IAAI;AAAA,MACT,sEAAiE,OAAO;AAAA,IAC1E;AAAA,EACF;AACA,SAAO,IAAI,MAAM,kCAAkC,OAAO,EAAE;AAC9D;AAEO,MAAM,2BAAqD;AAAA,EACxD,YAA8B;AAAA,EAC9B,QAAsB;AAAA,EACtB,cAAsE;AAAA,EACrE;AAAA,EAET,cAAc;AACZ,SAAK,UAAU;AAAA,EACjB;AAAA,EAEA,MAAM,SAAwB;AAC5B,UAAM,KAAK,KAAK;AAAA,EAClB;AAAA,EAEQ,OAAwD;AAC9D,QAAI,KAAK,aAAa,KAAK,OAAO;AAChC,aAAO,QAAQ,QAAQ,EAAE,WAAW,KAAK,WAAW,OAAO,KAAK,MAAM,CAAC;AAAA,IACzE;AACA,QAAI,KAAK,YAAa,QAAO,KAAK;AAElC,QAAI,KAAK,wBAAwB,EAAE,SAAS,KAAK,QAAQ,CAAC;AAC1D,UAAM,WAAW,KAAK,UAAU,EAAE,SAAS,cAAc;AAGzD,QAAI,WAAW;AAEf,SAAK,cAAc,QAAQ,IAAI;AAAA,MAC7B,cAAc,gBAAgB,KAAK,OAAO;AAAA,MAC1C,mCAAmC,gBAAgB,KAAK,OAAO;AAAA,IACjE,CAAC,EACE,KAAK,CAAC,CAAC,WAAW,KAAK,MAAM;AAC5B,WAAK,YAAY;AACjB,WAAK,QAAQ;AACb,aAAO,EAAE,WAAW,MAAM;AAAA,IAC5B,CAAC,EACA,MAAM,CAAC,QAAiB;AACvB,WAAK,cAAc;AACnB,YAAM,cAAc,GAAG;AAAA,IACzB,CAAC;AAEH,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,OACJ,OACA,YACA,OAAO,WAAW,QACO;AACzB,QAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AAErC,UAAM,EAAE,WAAW,MAAM,IAAI,MAAM,KAAK,KAAK;AAG7C,UAAM,UAAU,WAAW,IAAI,MAAM,KAAK;AAC1C,UAAM,OAAO,WAAW,IAAI,CAAC,MAAM,EAAE,IAAI;AAEzC,UAAM,SAAS,UAAU,SAAS;AAAA,MAChC,WAAW;AAAA,MACX,SAAS;AAAA,MACT,YAAY;AAAA,IACd,CAAC;AAED,UAAM,UAAW,MAAM,MAAM,MAAM;AACnC,UAAM,SAAS,QAAQ;AAIvB,UAAM,SAAS,OAAO,KAAK,UAAU,IAAI,OAAO,KAAK,CAAC,IAAI;AAC1D,UAAM,OAAO,OAAO;AAEpB,UAAM,SAAyB,WAAW,IAAI,CAAC,GAAG,OAAO;AAAA,MACvD,IAAI,EAAE;AAAA,MACN,OAAO,OAAO,KAAK,IAAI,MAAM,CAAC;AAAA,IAChC,EAAE;AAEF,WAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI;AAAA,EAC/D;AACF;","names":[]}
+{"version":3,"sources":["../../../src/search/reranker/transformers-rerank-provider.ts"],"sourcesContent":["import { join } from 'node:path';\nimport {\n  AutoTokenizer,\n  AutoModelForSequenceClassification,\n  env,\n} from '@huggingface/transformers';\nimport type {\n  RerankProvider,\n  RerankCandidate,\n  RerankResult,\n} from '../../providers/rerank-provider.js';\nimport { createLogger } from '../../logger.js';\nimport { getConfig } from '../../config.js';\n\nconst log = createLogger('reranker');\n\n// Cross-encoder reranker via Transformers.js.\n//\n// The high-level `pipeline('text-classification', ...)` API does not pass\n// `text_pair`, so it can't drive a cross-encoder properly. We therefore\n// load the tokenizer + sequence-classification model directly: feed\n// (query, document) pairs to the tokenizer and read raw logits from the\n// model. ms-marco-MiniLM-L-6-v2 is a single-output regressor (num_labels=1)\n// where higher logit = more relevant, so the logit is used as the rerank\n// score with no further transform.\ntype Tokenizer = Awaited<ReturnType<typeof AutoTokenizer.from_pretrained>>;\ntype Model = Awaited<ReturnType<typeof AutoModelForSequenceClassification.from_pretrained>>;\n\ninterface LogitsTensor {\n  data: ArrayLike<number>;\n  dims: number[];\n}\n\n// Recognize the noisy huggingface fetch failure signature and replace it\n// with an actionable instruction. Transformers.js parses a config that\n// failed to download, then dereferences `tokenizer_class` on undefined.\nfunction wrapLoadError(err: unknown): Error {\n  const message = err instanceof Error ? err.message : String(err);\n  const looksLikeMissingModel =\n    /tokenizer_class|tokenizer_config|preprocessor_config|fetch failed|ENOTFOUND|ECONNREFUSED|ETIMEDOUT|ENETUNREACH/i.test(\n      message,\n    );\n  if (looksLikeMissingModel) {\n    return new Error(\n      `Reranker model not downloaded — run \\`wigolo warmup\\` (cause: ${message})`,\n    );\n  }\n  return new Error(`Failed to load reranker model: ${message}`);\n}\n\nexport class TransformersRerankProvider implements RerankProvider {\n  private tokenizer: Tokenizer | null = null;\n  private model: Model | null = null;\n  private loadPromise: Promise<{ tokenizer: Tokenizer; model: Model }> | null = null;\n  readonly modelId: string;\n\n  constructor() {\n    this.modelId = 'Xenova/ms-marco-MiniLM-L-6-v2';\n  }\n\n  async warmup(): Promise<void> {\n    await this.load();\n  }\n\n  private load(): Promise<{ tokenizer: Tokenizer; model: Model }> {\n    if (this.tokenizer && this.model) {\n      return Promise.resolve({ tokenizer: this.tokenizer, model: this.model });\n    }\n    if (this.loadPromise) return this.loadPromise;\n\n    log.info('Loading rerank model', { modelId: this.modelId });\n    const cacheDir = join(getConfig().dataDir, 'transformers');\n    // Direct the library at a writable cache under the wigolo data dir so\n    // models don't end up in a user home cache the daemon can't manage.\n    env.cacheDir = cacheDir;\n\n    this.loadPromise = Promise.all([\n      AutoTokenizer.from_pretrained(this.modelId),\n      AutoModelForSequenceClassification.from_pretrained(this.modelId),\n    ])\n      .then(([tokenizer, model]) => {\n        this.tokenizer = tokenizer;\n        this.model = model;\n        return { tokenizer, model };\n      })\n      .catch((err: unknown) => {\n        this.loadPromise = null;\n        throw wrapLoadError(err);\n      });\n\n    return this.loadPromise;\n  }\n\n  async rerank(\n    query: string,\n    candidates: RerankCandidate[],\n    topK = candidates.length,\n  ): Promise<RerankResult[]> {\n    if (candidates.length === 0) return [];\n\n    const { tokenizer, model } = await this.load();\n\n    // Build batch: query repeated against each document.\n    const queries = candidates.map(() => query);\n    const docs = candidates.map((c) => c.text);\n\n    const inputs = tokenizer(queries, {\n      text_pair: docs,\n      padding: true,\n      truncation: true,\n    });\n\n    const outputs = (await model(inputs)) as { logits: LogitsTensor };\n    const logits = outputs.logits;\n    // logits shape is [batch, 1] for single-label regression rerankers.\n    // For multi-label heads (rare for rerankers) we still take the first\n    // value as the relevance score.\n    const stride = logits.dims.length >= 2 ? logits.dims[1] : 1;\n    const data = logits.data;\n\n    const scored: RerankResult[] = candidates.map((c, i) => ({\n      id: c.id,\n      score: Number(data[i * stride]),\n    }));\n\n    return scored.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n\n  // Release the underlying ONNX session before process exit. Without this,\n  // the runtime's worker threads race during C++ destructor teardown and\n  // surface as `mutex lock failed: Invalid argument` on macOS.\n  async dispose(): Promise<void> {\n    const model = this.model;\n    this.model = null;\n    this.tokenizer = null;\n    this.loadPromise = null;\n    if (!model) return;\n    try {\n      const m = model as unknown as { dispose?: () => Promise<unknown> };\n      if (typeof m.dispose === 'function') await m.dispose();\n    } catch (err) {\n      log.debug('reranker dispose failed', { error: err instanceof Error ? err.message : String(err) });\n    }\n  }\n}\n"],"mappings":"AAAA,SAAS,YAAY;AACrB;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAMP,SAAS,oBAAoB;AAC7B,SAAS,iBAAiB;AAE1B,MAAM,MAAM,aAAa,UAAU;AAsBnC,SAAS,cAAc,KAAqB;AAC1C,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,QAAM,wBACJ,kHAAkH;AAAA,IAChH;AAAA,EACF;AACF,MAAI,uBAAuB;AACzB,WAAO,IAAI;AAAA,MACT,sEAAiE,OAAO;AAAA,IAC1E;AAAA,EACF;AACA,SAAO,IAAI,MAAM,kCAAkC,OAAO,EAAE;AAC9D;AAEO,MAAM,2BAAqD;AAAA,EACxD,YAA8B;AAAA,EAC9B,QAAsB;AAAA,EACtB,cAAsE;AAAA,EACrE;AAAA,EAET,cAAc;AACZ,SAAK,UAAU;AAAA,EACjB;AAAA,EAEA,MAAM,SAAwB;AAC5B,UAAM,KAAK,KAAK;AAAA,EAClB;AAAA,EAEQ,OAAwD;AAC9D,QAAI,KAAK,aAAa,KAAK,OAAO;AAChC,aAAO,QAAQ,QAAQ,EAAE,WAAW,KAAK,WAAW,OAAO,KAAK,MAAM,CAAC;AAAA,IACzE;AACA,QAAI,KAAK,YAAa,QAAO,KAAK;AAElC,QAAI,KAAK,wBAAwB,EAAE,SAAS,KAAK,QAAQ,CAAC;AAC1D,UAAM,WAAW,KAAK,UAAU,EAAE,SAAS,cAAc;AAGzD,QAAI,WAAW;AAEf,SAAK,cAAc,QAAQ,IAAI;AAAA,MAC7B,cAAc,gBAAgB,KAAK,OAAO;AAAA,MAC1C,mCAAmC,gBAAgB,KAAK,OAAO;AAAA,IACjE,CAAC,EACE,KAAK,CAAC,CAAC,WAAW,KAAK,MAAM;AAC5B,WAAK,YAAY;AACjB,WAAK,QAAQ;AACb,aAAO,EAAE,WAAW,MAAM;AAAA,IAC5B,CAAC,EACA,MAAM,CAAC,QAAiB;AACvB,WAAK,cAAc;AACnB,YAAM,cAAc,GAAG;AAAA,IACzB,CAAC;AAEH,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,OACJ,OACA,YACA,OAAO,WAAW,QACO;AACzB,QAAI,WAAW,WAAW,EAAG,QAAO,CAAC;AAErC,UAAM,EAAE,WAAW,MAAM,IAAI,MAAM,KAAK,KAAK;AAG7C,UAAM,UAAU,WAAW,IAAI,MAAM,KAAK;AAC1C,UAAM,OAAO,WAAW,IAAI,CAAC,MAAM,EAAE,IAAI;AAEzC,UAAM,SAAS,UAAU,SAAS;AAAA,MAChC,WAAW;AAAA,MACX,SAAS;AAAA,MACT,YAAY;AAAA,IACd,CAAC;AAED,UAAM,UAAW,MAAM,MAAM,MAAM;AACnC,UAAM,SAAS,QAAQ;AAIvB,UAAM,SAAS,OAAO,KAAK,UAAU,IAAI,OAAO,KAAK,CAAC,IAAI;AAC1D,UAAM,OAAO,OAAO;AAEpB,UAAM,SAAyB,WAAW,IAAI,CAAC,GAAG,OAAO;AAAA,MACvD,IAAI,EAAE;AAAA,MACN,OAAO,OAAO,KAAK,IAAI,MAAM,CAAC;AAAA,IAChC,EAAE;AAEF,WAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,GAAG,IAAI;AAAA,EAC/D;AAAA;AAAA;AAAA;AAAA,EAKA,MAAM,UAAyB;AAC7B,UAAM,QAAQ,KAAK;AACnB,SAAK,QAAQ;AACb,SAAK,YAAY;AACjB,SAAK,cAAc;AACnB,QAAI,CAAC,MAAO;AACZ,QAAI;AACF,YAAM,IAAI;AACV,UAAI,OAAO,EAAE,YAAY,WAAY,OAAM,EAAE,QAAQ;AAAA,IACvD,SAAS,KAAK;AACZ,UAAI,MAAM,2BAA2B,EAAE,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAAE,CAAC;AAAA,IAClG;AAAA,EACF;AACF;","names":[]}

package/dist/tools/cache.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cache.d.ts","sourceRoot":"","sources":["../../src/tools/cache.ts"],"names":[],"mappings":"AAaA,OAAO,KAAK,EAAE,UAAU,EAAE,WAAW,EAAiC,MAAM,aAAa,CAAC;AAC1F,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAQtD,wBAAsB,WAAW,CAAC,KAAK,EAAE,UAAU,EAAE,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC,CAmH/F"}
+{"version":3,"file":"cache.d.ts","sourceRoot":"","sources":["../../src/tools/cache.ts"],"names":[],"mappings":"AAaA,OAAO,KAAK,EAAE,UAAU,EAAE,WAAW,EAAiC,MAAM,aAAa,CAAC;AAC1F,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAQtD,wBAAsB,WAAW,CAAC,KAAK,EAAE,UAAU,EAAE,MAAM,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC,CAqH/F"}

package/dist/tools/cache.js

@@ -102,12 +102,14 @@ async function handleCache(input, router) {
       query: input.query,
       urlPattern: input.url_pattern,
       since: input.since,
-      mode: input.mode
+      mode: input.mode,
+      limit: input.limit
     });
     const results = searchCacheFiltered({
       query: input.query,
       urlPattern: input.url_pattern,
-      since: input.since
+      since: input.since,
+      limit: input.limit
     });
     return {
       results: results.map((r) => ({