brainbank 0.7.0 → 0.8.0

package/dist/chunk-OLDHLOMT.js

@@ -0,0 +1,69 @@
+import {
+  __name
+} from "./chunk-7QVYU63E.js";
+
+// src/services/daemon.ts
+import * as fs from "fs";
+import * as path from "path";
+import * as os from "os";
+var DEFAULT_PORT = 8181;
+function cacheDir() {
+  return path.join(os.homedir(), ".cache", "brainbank");
+}
+__name(cacheDir, "cacheDir");
+function pidPath() {
+  return path.join(cacheDir(), "server.pid");
+}
+__name(pidPath, "pidPath");
+function writePid(pid, port) {
+  const dir = cacheDir();
+  fs.mkdirSync(dir, { recursive: true });
+  fs.writeFileSync(pidPath(), JSON.stringify({ pid, port }));
+}
+__name(writePid, "writePid");
+function readPid() {
+  try {
+    const raw = fs.readFileSync(pidPath(), "utf8");
+    const info = JSON.parse(raw);
+    if (typeof info.pid !== "number" || typeof info.port !== "number") return null;
+    return info;
+  } catch {
+    return null;
+  }
+}
+__name(readPid, "readPid");
+function removePid() {
+  try {
+    fs.unlinkSync(pidPath());
+  } catch {
+  }
+}
+__name(removePid, "removePid");
+function isServerRunning() {
+  const info = readPid();
+  if (!info) return null;
+  try {
+    process.kill(info.pid, 0);
+    return info;
+  } catch {
+    removePid();
+    return null;
+  }
+}
+__name(isServerRunning, "isServerRunning");
+function getServerUrl() {
+  const info = isServerRunning();
+  if (!info) return null;
+  return `http://localhost:${info.port}`;
+}
+__name(getServerUrl, "getServerUrl");
+
+export {
+  DEFAULT_PORT,
+  writePid,
+  readPid,
+  removePid,
+  isServerRunning,
+  getServerUrl
+};
+//# sourceMappingURL=chunk-OLDHLOMT.js.map

package/dist/chunk-OLDHLOMT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/services/daemon.ts"],"sourcesContent":["/**\n * Daemon — PID file management for the BrainBank HTTP server.\n *\n * PID file: ~/.cache/brainbank/server.pid\n * Format: JSON { pid: number, port: number }\n *\n * Used by:\n *   - `brainbank daemon` to write PID on start\n *   - `brainbank daemon stop` to find and kill the daemon\n *   - `brainbank status` to report server state\n *   - CLI commands to detect running server for delegation\n */\n\nimport * as fs from 'node:fs';\nimport * as path from 'node:path';\nimport * as os from 'node:os';\n\nexport const DEFAULT_PORT = 8181;\n\ninterface PidInfo {\n    pid: number;\n    port: number;\n}\n\n/** Directory for PID file and other cache data. */\nfunction cacheDir(): string {\n    return path.join(os.homedir(), '.cache', 'brainbank');\n}\n\n/** Full path to the PID file. */\nfunction pidPath(): string {\n    return path.join(cacheDir(), 'server.pid');\n}\n\n/** Write the PID file after server starts. */\nexport function writePid(pid: number, port: number): void {\n    const dir = cacheDir();\n    fs.mkdirSync(dir, { recursive: true });\n    fs.writeFileSync(pidPath(), JSON.stringify({ pid, port } as PidInfo));\n}\n\n/** Read PID info from file. Returns null if missing or malformed. */\nexport function readPid(): PidInfo | null {\n    try {\n        const raw = fs.readFileSync(pidPath(), 'utf8');\n        const info = JSON.parse(raw) as PidInfo;\n        if (typeof info.pid !== 'number' || typeof info.port !== 'number') return null;\n        return info;\n    } catch {\n        return null;\n    }\n}\n\n/** Remove the PID file. */\nexport function removePid(): void {\n    try {\n        fs.unlinkSync(pidPath());\n    } catch {\n        // File doesn't exist — safe to ignore\n    }\n}\n\n/**\n * Check if the server process is still alive.\n * Uses `kill(pid, 0)` which doesn't actually send a signal —\n * it just checks whether the process exists.\n */\nexport function isServerRunning(): PidInfo | null {\n    const info = readPid();\n    if (!info) return null;\n\n    try {\n        process.kill(info.pid, 0);\n        return info;\n    } catch {\n        // Process not running — stale PID file\n        removePid();\n        return null;\n    }\n}\n\n/** Get the server URL if running, or null. */\nexport function getServerUrl(): string | null {\n    const info = isServerRunning();\n    if (!info) return null;\n    return `http://localhost:${info.port}`;\n}\n"],"mappings":";;;;;AAaA,YAAY,QAAQ;AACpB,YAAY,UAAU;AACtB,YAAY,QAAQ;AAEb,IAAM,eAAe;AAQ5B,SAAS,WAAmB;AACxB,SAAY,UAAQ,WAAQ,GAAG,UAAU,WAAW;AACxD;AAFS;AAKT,SAAS,UAAkB;AACvB,SAAY,UAAK,SAAS,GAAG,YAAY;AAC7C;AAFS;AAKF,SAAS,SAAS,KAAa,MAAoB;AACtD,QAAM,MAAM,SAAS;AACrB,EAAG,aAAU,KAAK,EAAE,WAAW,KAAK,CAAC;AACrC,EAAG,iBAAc,QAAQ,GAAG,KAAK,UAAU,EAAE,KAAK,KAAK,CAAY,CAAC;AACxE;AAJgB;AAOT,SAAS,UAA0B;AACtC,MAAI;AACA,UAAM,MAAS,gBAAa,QAAQ,GAAG,MAAM;AAC7C,UAAM,OAAO,KAAK,MAAM,GAAG;AAC3B,QAAI,OAAO,KAAK,QAAQ,YAAY,OAAO,KAAK,SAAS,SAAU,QAAO;AAC1E,WAAO;AAAA,EACX,QAAQ;AACJ,WAAO;AAAA,EACX;AACJ;AATgB;AAYT,SAAS,YAAkB;AAC9B,MAAI;AACA,IAAG,cAAW,QAAQ,CAAC;AAAA,EAC3B,QAAQ;AAAA,EAER;AACJ;AANgB;AAaT,SAAS,kBAAkC;AAC9C,QAAM,OAAO,QAAQ;AACrB,MAAI,CAAC,KAAM,QAAO;AAElB,MAAI;AACA,YAAQ,KAAK,KAAK,KAAK,CAAC;AACxB,WAAO;AAAA,EACX,QAAQ;AAEJ,cAAU;AACV,WAAO;AAAA,EACX;AACJ;AAZgB;AAeT,SAAS,eAA8B;AAC1C,QAAM,OAAO,gBAAgB;AAC7B,MAAI,CAAC,KAAM,QAAO;AAClB,SAAO,oBAAoB,KAAK,IAAI;AACxC;AAJgB;","names":[]}

package/dist/{chunk-424UFCY7.js → chunk-QTZNB6AK.js}

@@ -25,7 +25,11 @@ var LocalEmbedding = class {
     if (this._pipeline) return this._pipeline;
     if (this._pipelinePromise) return this._pipelinePromise;
     this._pipelinePromise = (async () => {
-      const { pipeline, env } = await import("@xenova/transformers");
+      const mod = await import(
+        /* webpackIgnore: true */
+        "@xenova/transformers"
+      );
+      const { pipeline, env } = mod;
       env.cacheDir = this._cacheDir;
       env.allowLocalModels = true;
       this._pipeline = await pipeline("feature-extraction", this._modelName, {
@@ -75,4 +79,4 @@ var LocalEmbedding = class {
 export {
   LocalEmbedding
 };
-//# sourceMappingURL=chunk-424UFCY7.js.map
+//# sourceMappingURL=chunk-QTZNB6AK.js.map

package/dist/chunk-QTZNB6AK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/providers/embeddings/local-embedding.ts"],"sourcesContent":["/**\n * BrainBank — Local Embedding Provider\n * \n * Uses @xenova/transformers with all-MiniLM-L6-v2 (384 dims, WASM).\n * Downloads ~23MB on first use, cached locally.\n * No external API calls — runs entirely in-process.\n */\n\nimport type { EmbeddingProvider } from '@/types.ts';\n\n/** Minimal interface for @xenova/transformers pipeline results. */\ninterface XenovaPipelineOutput {\n    data: Float32Array;\n}\n\n/** Callable pipeline returned by @xenova/transformers. */\ninterface XenovaPipeline {\n    (texts: string | string[], options: { pooling: string; normalize: boolean }): Promise<XenovaPipelineOutput>;\n}\n\n/** Configuration environment of @xenova/transformers. */\ninterface XenovaEnv {\n    cacheDir: string;\n    allowLocalModels: boolean;\n}\n\n/** Shape of the @xenova/transformers module used here. */\ninterface XenovaModule {\n    pipeline(task: string, model: string, options?: { quantized?: boolean }): Promise<XenovaPipeline>;\n    env: XenovaEnv;\n}\n\nexport class LocalEmbedding implements EmbeddingProvider {\n    readonly dims: number = 384;\n\n    private _pipeline: XenovaPipeline | null = null;\n    private _modelName: string;\n    private _cacheDir: string;\n\n    constructor(options: { model?: string; cacheDir?: string } = {}) {\n        this._modelName = options.model ?? 'Xenova/all-MiniLM-L6-v2';\n        this._cacheDir = options.cacheDir ?? '.model-cache';\n    }\n\n    private _pipelinePromise: Promise<XenovaPipeline> | null = null;\n\n    /**\n     * Lazy-load the transformer pipeline.\n     * Singleton — created once and reused.\n     * Promise-deduped to prevent concurrent downloads.\n     */\n    private async _getPipeline(): Promise<XenovaPipeline> {\n        if (this._pipeline) return this._pipeline;\n        if (this._pipelinePromise) return this._pipelinePromise;\n\n        this._pipelinePromise = (async () => {\n            const mod = await import(/* webpackIgnore: true */ '@xenova/transformers' as string) as XenovaModule;\n            const { pipeline, env } = mod;\n            env.cacheDir = this._cacheDir;\n            env.allowLocalModels = true;\n\n            this._pipeline = await pipeline('feature-extraction', this._modelName, {\n                quantized: true,\n            });\n\n            return this._pipeline!;\n        })();\n\n        try {\n            return await this._pipelinePromise;\n        } finally {\n            this._pipelinePromise = null;\n        }\n    }\n\n    /**\n     * Embed a single text string.\n     * Returns a normalized Float32Array of length 384.\n     */\n    async embed(text: string): Promise<Float32Array> {\n        const pipe = await this._getPipeline();\n        const output = await pipe(text, { pooling: 'mean', normalize: true });\n        return output.data as Float32Array;\n    }\n\n    /**\n     * Embed multiple texts using real batch processing.\n     * Chunks into groups of BATCH_SIZE to balance throughput vs memory.\n     */\n    async embedBatch(texts: string[]): Promise<Float32Array[]> {\n        if (texts.length === 0) return [];\n\n        const BATCH_SIZE = 32;\n        const pipe = await this._getPipeline();\n        const results: Float32Array[] = [];\n\n        for (let i = 0; i < texts.length; i += BATCH_SIZE) {\n            const batch = texts.slice(i, i + BATCH_SIZE);\n            const output = await pipe(batch, { pooling: 'mean', normalize: true });\n\n            // output.data is a flat Float32Array — must copy, not view,\n            // because the pipeline may reuse the underlying buffer\n            for (let j = 0; j < batch.length; j++) {\n                const start = j * this.dims;\n                results.push(output.data.slice(start, start + this.dims) as Float32Array);\n            }\n        }\n\n        return results;\n    }\n\n    async close(): Promise<void> {\n        this._pipeline = null;\n    }\n}\n"],"mappings":";;;;;AAgCO,IAAM,iBAAN,MAAkD;AAAA,EAhCzD,OAgCyD;AAAA;AAAA;AAAA,EAC5C,OAAe;AAAA,EAEhB,YAAmC;AAAA,EACnC;AAAA,EACA;AAAA,EAER,YAAY,UAAiD,CAAC,GAAG;AAC7D,SAAK,aAAa,QAAQ,SAAS;AACnC,SAAK,YAAY,QAAQ,YAAY;AAAA,EACzC;AAAA,EAEQ,mBAAmD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAO3D,MAAc,eAAwC;AAClD,QAAI,KAAK,UAAW,QAAO,KAAK;AAChC,QAAI,KAAK,iBAAkB,QAAO,KAAK;AAEvC,SAAK,oBAAoB,YAAY;AACjC,YAAM,MAAM,MAAM;AAAA;AAAA,QAAiC;AAAA,MAAgC;AACnF,YAAM,EAAE,UAAU,IAAI,IAAI;AAC1B,UAAI,WAAW,KAAK;AACpB,UAAI,mBAAmB;AAEvB,WAAK,YAAY,MAAM,SAAS,sBAAsB,KAAK,YAAY;AAAA,QACnE,WAAW;AAAA,MACf,CAAC;AAED,aAAO,KAAK;AAAA,IAChB,GAAG;AAEH,QAAI;AACA,aAAO,MAAM,KAAK;AAAA,IACtB,UAAE;AACE,WAAK,mBAAmB;AAAA,IAC5B;AAAA,EACJ;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,MAAM,MAAqC;AAC7C,UAAM,OAAO,MAAM,KAAK,aAAa;AACrC,UAAM,SAAS,MAAM,KAAK,MAAM,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AACpE,WAAO,OAAO;AAAA,EAClB;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAM,WAAW,OAA0C;AACvD,QAAI,MAAM,WAAW,EAAG,QAAO,CAAC;AAEhC,UAAM,aAAa;AACnB,UAAM,OAAO,MAAM,KAAK,aAAa;AACrC,UAAM,UAA0B,CAAC;AAEjC,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK,YAAY;AAC/C,YAAM,QAAQ,MAAM,MAAM,GAAG,IAAI,UAAU;AAC3C,YAAM,SAAS,MAAM,KAAK,OAAO,EAAE,SAAS,QAAQ,WAAW,KAAK,CAAC;AAIrE,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACnC,cAAM,QAAQ,IAAI,KAAK;AACvB,gBAAQ,KAAK,OAAO,KAAK,MAAM,OAAO,QAAQ,KAAK,IAAI,CAAiB;AAAA,MAC5E;AAAA,IACJ;AAEA,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,QAAuB;AACzB,SAAK,YAAY;AAAA,EACrB;AACJ;","names":[]}

package/dist/chunk-RFF7HMP6.js

@@ -0,0 +1,109 @@
+import {
+  __name
+} from "./chunk-7QVYU63E.js";
+
+// src/providers/pruners/haiku-expander.ts
+var DEFAULT_MODEL = "claude-haiku-4-5-20251001";
+var HaikuExpander = class {
+  static {
+    __name(this, "HaikuExpander");
+  }
+  _apiKey;
+  _model;
+  constructor(options = {}) {
+    this._apiKey = options.apiKey ?? process.env.ANTHROPIC_API_KEY ?? "";
+    this._model = options.model ?? DEFAULT_MODEL;
+    if (!this._apiKey) {
+      throw new Error(
+        "HaikuExpander: No API key provided. Set ANTHROPIC_API_KEY env var or pass apiKey option."
+      );
+    }
+  }
+  async expand(query, currentIds, manifest) {
+    if (manifest.length === 0) return { ids: [] };
+    const currentSet = new Set(currentIds);
+    const available = manifest.filter((m) => !currentSet.has(m.id));
+    if (available.length === 0) return { ids: [] };
+    const manifestLines = available.map(
+      (m) => `#${m.id} ${m.filePath} | ${m.chunkType} ${m.name} ${m.lines}`
+    ).join("\n");
+    const currentSummary = manifest.filter((m) => currentSet.has(m.id)).map((m) => `#${m.id} ${m.filePath} | ${m.chunkType} ${m.name}`).join("\n");
+    const prompt = `Task: "${query}"
+
+Already included chunks:
+${currentSummary}
+
+Available chunks NOT yet included:
+${manifestLines}
+
+You are a code context expander. The search already found the "included" chunks above.
+Review the "available" chunks and select any that would help an AI agent complete the task.
+
+Rules:
+- Select chunks that are structurally related (called by, imported by, or extend included code)
+- Select type definitions, interfaces, or configs needed to understand included code
+- Select initialization or setup code if the task involves debugging or modifying a feature
+- Do NOT select test files, documentation, or unrelated utilities
+- Be selective: only include chunks that fill clear gaps. Quality over quantity.
+- If nothing useful is available, return an empty ids array
+
+Respond with ONLY a JSON object:
+{ "ids": [42, 17, 89], "note": "Brief 1-2 sentence observation about the codebase relevant to the task" }
+
+The "note" is optional \u2014 use it to mention things like missing files, architectural patterns, deprecated modules, or important relationships you noticed. If nothing notable, omit it.
+If nothing to add: { "ids": [] }`;
+    try {
+      const response = await fetch("https://api.anthropic.com/v1/messages", {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          "x-api-key": this._apiKey,
+          "anthropic-version": "2023-06-01"
+        },
+        body: JSON.stringify({
+          model: this._model,
+          max_tokens: 512,
+          messages: [{
+            role: "user",
+            content: prompt
+          }]
+        })
+      });
+      if (!response.ok) {
+        return { ids: [] };
+      }
+      const data = await response.json();
+      const text = data.content?.[0]?.text ?? "";
+      return this._parseResponse(text, available);
+    } catch {
+      return { ids: [] };
+    }
+  }
+  /** Parse Haiku response — handles both `{ ids, note }` and bare `[...]` formats. */
+  _parseResponse(text, available) {
+    const validIds = new Set(available.map((m) => m.id));
+    const objMatch = text.match(/\{[\s\S]*"ids"\s*:\s*\[[\d\s,]*\][\s\S]*\}/);
+    if (objMatch) {
+      try {
+        const parsed = JSON.parse(objMatch[0]);
+        const ids = parsed.ids.filter((id) => validIds.has(id));
+        const note = parsed.note?.trim() || void 0;
+        return { ids, note };
+      } catch {
+      }
+    }
+    const arrMatch = text.match(/\[[\d\s,]*\]/);
+    if (arrMatch) {
+      const ids = JSON.parse(arrMatch[0]).filter((id) => validIds.has(id));
+      return { ids };
+    }
+    return { ids: [] };
+  }
+  async close() {
+  }
+};
+
+export {
+  HaikuExpander
+};
+//# sourceMappingURL=chunk-RFF7HMP6.js.map

package/dist/chunk-RFF7HMP6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/providers/pruners/haiku-expander.ts"],"sourcesContent":["/**\n * BrainBank — Haiku Expander\n *\n * LLM-powered context expansion using Anthropic's Haiku 4.5 model.\n * After search + pruning, reviews a manifest of available chunks\n * and requests additional IDs to include.\n *\n * Flow:\n *   1. Receives lightweight manifest (~20 chars per chunk)\n *   2. Haiku selects additional chunk IDs (just numbers, fast)\n *   3. Caller fetches those chunks from DB and splices into results\n *\n * Designed for minimal token usage:\n *   - Input: ~2,000-3,000 tokens (manifest)\n *   - Output: ~50-100 tokens (ID array)\n *   - Cost: ~$0.001 per call\n *   - Latency: ~300-600ms\n *\n * Fail-open: any error returns empty array (no expansion).\n */\n\nimport type { Expander, ExpanderManifestItem, ExpanderResult } from '@/types.ts';\n\nconst DEFAULT_MODEL = 'claude-haiku-4-5-20251001';\n\nexport interface HaikuExpanderOptions {\n    /** Anthropic API key. Falls back to ANTHROPIC_API_KEY env var. */\n    apiKey?: string;\n    /** Model to use. Default: claude-haiku-4-5-20251001 */\n    model?: string;\n}\n\nexport class HaikuExpander implements Expander {\n    private readonly _apiKey: string;\n    private readonly _model: string;\n\n    constructor(options: HaikuExpanderOptions = {}) {\n        this._apiKey = options.apiKey ?? process.env.ANTHROPIC_API_KEY ?? '';\n        this._model = options.model ?? DEFAULT_MODEL;\n\n        if (!this._apiKey) {\n            throw new Error(\n                'HaikuExpander: No API key provided. Set ANTHROPIC_API_KEY env var or pass apiKey option.',\n            );\n        }\n    }\n\n    async expand(\n        query: string,\n        currentIds: number[],\n        manifest: ExpanderManifestItem[],\n    ): Promise<ExpanderResult> {\n        if (manifest.length === 0) return { ids: [] };\n\n        // Filter out chunks already in results\n        const currentSet = new Set(currentIds);\n        const available = manifest.filter(m => !currentSet.has(m.id));\n        if (available.length === 0) return { ids: [] };\n\n        // Build compact manifest string\n        const manifestLines = available.map(m =>\n            `#${m.id} ${m.filePath} | ${m.chunkType} ${m.name} ${m.lines}`\n        ).join('\\n');\n\n        const currentSummary = manifest\n            .filter(m => currentSet.has(m.id))\n            .map(m => `#${m.id} ${m.filePath} | ${m.chunkType} ${m.name}`)\n            .join('\\n');\n\n        const prompt =\n            `Task: \"${query}\"\\n\\n` +\n            `Already included chunks:\\n${currentSummary}\\n\\n` +\n            `Available chunks NOT yet included:\\n${manifestLines}\\n\\n` +\n            `You are a code context expander. The search already found the \"included\" chunks above.\\n` +\n            `Review the \"available\" chunks and select any that would help an AI agent complete the task.\\n\\n` +\n            `Rules:\\n` +\n            `- Select chunks that are structurally related (called by, imported by, or extend included code)\\n` +\n            `- Select type definitions, interfaces, or configs needed to understand included code\\n` +\n            `- Select initialization or setup code if the task involves debugging or modifying a feature\\n` +\n            `- Do NOT select test files, documentation, or unrelated utilities\\n` +\n            `- Be selective: only include chunks that fill clear gaps. Quality over quantity.\\n` +\n            `- If nothing useful is available, return an empty ids array\\n\\n` +\n            `Respond with ONLY a JSON object:\\n` +\n            `{ \"ids\": [42, 17, 89], \"note\": \"Brief 1-2 sentence observation about the codebase relevant to the task\" }\\n\\n` +\n            `The \"note\" is optional — use it to mention things like missing files, architectural patterns, ` +\n            `deprecated modules, or important relationships you noticed. If nothing notable, omit it.\\n` +\n            `If nothing to add: { \"ids\": [] }`;\n\n        try {\n            const response = await fetch('https://api.anthropic.com/v1/messages', {\n                method: 'POST',\n                headers: {\n                    'Content-Type': 'application/json',\n                    'x-api-key': this._apiKey,\n                    'anthropic-version': '2023-06-01',\n                },\n                body: JSON.stringify({\n                    model: this._model,\n                    max_tokens: 512,\n                    messages: [{\n                        role: 'user',\n                        content: prompt,\n                    }],\n                }),\n            });\n\n            if (!response.ok) {\n                return { ids: [] };\n            }\n\n            const data = await response.json() as {\n                content: { type: string; text: string }[];\n            };\n\n            const text = data.content?.[0]?.text ?? '';\n            return this._parseResponse(text, available);\n        } catch {\n            return { ids: [] };\n        }\n    }\n\n    /** Parse Haiku response — handles both `{ ids, note }` and bare `[...]` formats. */\n    private _parseResponse(text: string, available: ExpanderManifestItem[]): ExpanderResult {\n        const validIds = new Set(available.map(m => m.id));\n\n        // Try JSON object first: { \"ids\": [...], \"note\": \"...\" }\n        const objMatch = text.match(/\\{[\\s\\S]*\"ids\"\\s*:\\s*\\[[\\d\\s,]*\\][\\s\\S]*\\}/);\n        if (objMatch) {\n            try {\n                const parsed = JSON.parse(objMatch[0]) as { ids: number[]; note?: string };\n                const ids = parsed.ids.filter(id => validIds.has(id));\n                const note = parsed.note?.trim() || undefined;\n                return { ids, note };\n            } catch {\n                // Fall through to array parsing\n            }\n        }\n\n        // Fallback: bare array [42, 17, 89]\n        const arrMatch = text.match(/\\[[\\d\\s,]*\\]/);\n        if (arrMatch) {\n            const ids = (JSON.parse(arrMatch[0]) as number[]).filter(id => validIds.has(id));\n            return { ids };\n        }\n\n        return { ids: [] };\n    }\n\n    async close(): Promise<void> {\n        // No resources to release (stateless HTTP)\n    }\n}\n"],"mappings":";;;;;AAuBA,IAAM,gBAAgB;AASf,IAAM,gBAAN,MAAwC;AAAA,EAhC/C,OAgC+C;AAAA;AAAA;AAAA,EAC1B;AAAA,EACA;AAAA,EAEjB,YAAY,UAAgC,CAAC,GAAG;AAC5C,SAAK,UAAU,QAAQ,UAAU,QAAQ,IAAI,qBAAqB;AAClE,SAAK,SAAS,QAAQ,SAAS;AAE/B,QAAI,CAAC,KAAK,SAAS;AACf,YAAM,IAAI;AAAA,QACN;AAAA,MACJ;AAAA,IACJ;AAAA,EACJ;AAAA,EAEA,MAAM,OACF,OACA,YACA,UACuB;AACvB,QAAI,SAAS,WAAW,EAAG,QAAO,EAAE,KAAK,CAAC,EAAE;AAG5C,UAAM,aAAa,IAAI,IAAI,UAAU;AACrC,UAAM,YAAY,SAAS,OAAO,OAAK,CAAC,WAAW,IAAI,EAAE,EAAE,CAAC;AAC5D,QAAI,UAAU,WAAW,EAAG,QAAO,EAAE,KAAK,CAAC,EAAE;AAG7C,UAAM,gBAAgB,UAAU;AAAA,MAAI,OAChC,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,MAAM,EAAE,SAAS,IAAI,EAAE,IAAI,IAAI,EAAE,KAAK;AAAA,IAChE,EAAE,KAAK,IAAI;AAEX,UAAM,iBAAiB,SAClB,OAAO,OAAK,WAAW,IAAI,EAAE,EAAE,CAAC,EAChC,IAAI,OAAK,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,MAAM,EAAE,SAAS,IAAI,EAAE,IAAI,EAAE,EAC5D,KAAK,IAAI;AAEd,UAAM,SACF,UAAU,KAAK;AAAA;AAAA;AAAA,EACc,cAAc;AAAA;AAAA;AAAA,EACJ,aAAa;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAgBxD,QAAI;AACA,YAAM,WAAW,MAAM,MAAM,yCAAyC;AAAA,QAClE,QAAQ;AAAA,QACR,SAAS;AAAA,UACL,gBAAgB;AAAA,UAChB,aAAa,KAAK;AAAA,UAClB,qBAAqB;AAAA,QACzB;AAAA,QACA,MAAM,KAAK,UAAU;AAAA,UACjB,OAAO,KAAK;AAAA,UACZ,YAAY;AAAA,UACZ,UAAU,CAAC;AAAA,YACP,MAAM;AAAA,YACN,SAAS;AAAA,UACb,CAAC;AAAA,QACL,CAAC;AAAA,MACL,CAAC;AAED,UAAI,CAAC,SAAS,IAAI;AACd,eAAO,EAAE,KAAK,CAAC,EAAE;AAAA,MACrB;AAEA,YAAM,OAAO,MAAM,SAAS,KAAK;AAIjC,YAAM,OAAO,KAAK,UAAU,CAAC,GAAG,QAAQ;AACxC,aAAO,KAAK,eAAe,MAAM,SAAS;AAAA,IAC9C,QAAQ;AACJ,aAAO,EAAE,KAAK,CAAC,EAAE;AAAA,IACrB;AAAA,EACJ;AAAA;AAAA,EAGQ,eAAe,MAAc,WAAmD;AACpF,UAAM,WAAW,IAAI,IAAI,UAAU,IAAI,OAAK,EAAE,EAAE,CAAC;AAGjD,UAAM,WAAW,KAAK,MAAM,4CAA4C;AACxE,QAAI,UAAU;AACV,UAAI;AACA,cAAM,SAAS,KAAK,MAAM,SAAS,CAAC,CAAC;AACrC,cAAM,MAAM,OAAO,IAAI,OAAO,QAAM,SAAS,IAAI,EAAE,CAAC;AACpD,cAAM,OAAO,OAAO,MAAM,KAAK,KAAK;AACpC,eAAO,EAAE,KAAK,KAAK;AAAA,MACvB,QAAQ;AAAA,MAER;AAAA,IACJ;AAGA,UAAM,WAAW,KAAK,MAAM,cAAc;AAC1C,QAAI,UAAU;AACV,YAAM,MAAO,KAAK,MAAM,SAAS,CAAC,CAAC,EAAe,OAAO,QAAM,SAAS,IAAI,EAAE,CAAC;AAC/E,aAAO,EAAE,IAAI;AAAA,IACjB;AAEA,WAAO,EAAE,KAAK,CAAC,EAAE;AAAA,EACrB;AAAA,EAEA,MAAM,QAAuB;AAAA,EAE7B;AACJ;","names":[]}

package/dist/{chunk-ZNLN2VWV.js → chunk-TD3TEFI3.js}

@@ -107,4 +107,4 @@ var Qwen3Reranker = class {
 export {
   Qwen3Reranker
 };
-//# sourceMappingURL=chunk-ZNLN2VWV.js.map
+//# sourceMappingURL=chunk-TD3TEFI3.js.map

package/dist/chunk-TD3TEFI3.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/providers/rerankers/qwen3-reranker.ts"],"sourcesContent":["/**\n * BrainBank — Qwen3 Local Reranker\n * \n * Cross-encoder reranker using Qwen3-Reranker-0.6B via node-llama-cpp.\n * Auto-downloads the GGUF model from HuggingFace (~640MB, cached).\n * \n * Based on QMD's reranker architecture:\n * - Lazy model loading (loads on first rank() call)\n * - Flash attention for 20× less VRAM\n * - Document deduplication (identical texts scored once)\n * - Tokenizer-based truncation for oversized documents\n */\n\nimport type { Reranker } from '@/types.ts';\nimport { homedir } from 'node:os';\nimport { join } from 'node:path';\nimport { existsSync, mkdirSync } from 'node:fs';\n\n/** Minimal interface for node-llama-cpp model methods used by the reranker. */\ninterface LlamaModel {\n    tokenize(text: string): number[];\n    detokenize(tokens: number[]): string;\n    createRankingContext(opts: { contextSize: number; flashAttention?: boolean }): Promise<LlamaRankingContext>;\n    dispose?(): Promise<void>;\n}\n\n/** Minimal interface for node-llama-cpp ranking context. */\ninterface LlamaRankingContext {\n    rankAll(query: string, documents: string[]): Promise<number[]>;\n    dispose(): Promise<void>;\n}\n\n/** Engine returned by getLlama(). */\ninterface LlamaEngine {\n    loadModel(opts: { modelPath: string }): Promise<LlamaModel>;\n}\n\n/** Shape of the node-llama-cpp module used here. */\ninterface NodeLlamaCppModule {\n    getLlama(): Promise<LlamaEngine>;\n    resolveModelFile(uri: string, cacheDir: string): Promise<string>;\n}\n\n// Default model — Qwen3-Reranker-0.6B quantized to Q8_0 (~640MB)\nconst DEFAULT_MODEL_URI = 'hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf';\n\n// Context size for reranking (Qwen3 template ~200 tokens overhead + query + doc)\nconst CONTEXT_SIZE = 2048;\n\n// Cache directory for downloaded models\nconst MODEL_CACHE_DIR = join(homedir(), '.cache', 'brainbank', 'models');\n\nexport interface Qwen3RerankerOptions {\n    /** HuggingFace model URI. Default: Qwen3-Reranker-0.6B-Q8_0 */\n    modelUri?: string;\n    /** Model cache directory. Default: ~/.cache/brainbank/models/ */\n    cacheDir?: string;\n    /** Context size for ranking. Default: 2048 */\n    contextSize?: number;\n}\n\nexport class Qwen3Reranker implements Reranker {\n    private readonly _modelUri: string;\n    private readonly _cacheDir: string;\n    private readonly _contextSize: number;\n\n    private _model: LlamaModel | null = null;\n    private _context: LlamaRankingContext | null = null;\n    private _loadPromise: Promise<void> | null = null;\n\n    constructor(options: Qwen3RerankerOptions = {}) {\n        this._modelUri = options.modelUri ?? DEFAULT_MODEL_URI;\n        this._cacheDir = options.cacheDir ?? MODEL_CACHE_DIR;\n        this._contextSize = options.contextSize ?? CONTEXT_SIZE;\n    }\n\n    /**\n     * Lazy-load the reranker model and create a ranking context.\n     * Model is auto-downloaded from HuggingFace on first use.\n     */\n    private async _ensureLoaded(): Promise<void> {\n        if (this._context) return;\n        if (this._loadPromise) {\n            await this._loadPromise;\n            return;\n        }\n\n        this._loadPromise = (async () => {\n            try {\n                // Dynamic import — node-llama-cpp is an optional peer dependency.\n                // String indirection prevents DTS from resolving the specifier at build time.\n                const llamaModule = 'node-llama-cpp';\n                const { getLlama, resolveModelFile } = await import(/* webpackIgnore: true */ llamaModule) as NodeLlamaCppModule;\n\n                // Ensure cache directory exists\n                if (!existsSync(this._cacheDir)) {\n                    mkdirSync(this._cacheDir, { recursive: true });\n                }\n\n                // Download model if needed (resolveModelFile handles caching)\n                const modelPath = await resolveModelFile(this._modelUri, this._cacheDir);\n\n                // Initialize llama engine\n                const llama = await getLlama();\n\n                // Load model\n                this._model = await llama.loadModel({ modelPath });\n\n                try {\n                    this._context = await this._model!.createRankingContext({\n                        contextSize: this._contextSize,\n                        flashAttention: true,\n                    });\n                } catch {\n                    // Flash attention might not be supported — retry without it\n                    this._context = await this._model!.createRankingContext({\n                        contextSize: this._contextSize,\n                    });\n                }\n            } finally {\n                this._loadPromise = null;\n            }\n        })();\n\n        await this._loadPromise;\n    }\n\n    /**\n     * Score each document's relevance to the query.\n     * Returns scores (0.0 - 1.0) in same order as input documents.\n     * \n     * Deduplicates identical documents to avoid redundant computation.\n     */\n    async rank(query: string, documents: string[]): Promise<number[]> {\n        if (documents.length === 0) return [];\n\n        await this._ensureLoaded();\n\n        // Deduplicate — identical texts get scored once\n        const uniqueTexts = [...new Set(documents)];\n        const textToScore = new Map<string, number>();\n\n        // Truncate documents that exceed context size\n        const truncated = uniqueTexts.map(text => {\n            if (this._model) {\n                const tokens = this._model.tokenize(text);\n                // Budget: contextSize - ~200 overhead - query tokens\n                const queryTokens = this._model.tokenize(query).length;\n                const maxDocTokens = this._contextSize - 200 - queryTokens;\n                if (tokens.length > maxDocTokens && maxDocTokens > 0) {\n                    return this._model.detokenize(tokens.slice(0, maxDocTokens));\n                }\n            }\n            return text;\n        });\n\n        // Rank all unique documents at once\n        const scores: number[] = await this._context!.rankAll(query, truncated);\n\n        // Map scores back\n        for (let i = 0; i < uniqueTexts.length; i++) {\n            textToScore.set(uniqueTexts[i], scores[i] ?? 0);\n        }\n\n        // Return scores in original document order\n        return documents.map(doc => textToScore.get(doc) ?? 0);\n    }\n\n    /** Release model resources. */\n    async close(): Promise<void> {\n        if (this._context) {\n            await this._context.dispose();\n            this._context = null;\n        }\n        if (this._model) {\n            await this._model.dispose?.();\n            this._model = null;\n        }\n    }\n}\n"],"mappings":";;;;;AAcA,SAAS,eAAe;AACxB,SAAS,YAAY;AACrB,SAAS,YAAY,iBAAiB;AA4BtC,IAAM,oBAAoB;AAG1B,IAAM,eAAe;AAGrB,IAAM,kBAAkB,KAAK,QAAQ,GAAG,UAAU,aAAa,QAAQ;AAWhE,IAAM,gBAAN,MAAwC;AAAA,EA7D/C,OA6D+C;AAAA;AAAA;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EAET,SAA4B;AAAA,EAC5B,WAAuC;AAAA,EACvC,eAAqC;AAAA,EAE7C,YAAY,UAAgC,CAAC,GAAG;AAC5C,SAAK,YAAY,QAAQ,YAAY;AACrC,SAAK,YAAY,QAAQ,YAAY;AACrC,SAAK,eAAe,QAAQ,eAAe;AAAA,EAC/C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,gBAA+B;AACzC,QAAI,KAAK,SAAU;AACnB,QAAI,KAAK,cAAc;AACnB,YAAM,KAAK;AACX;AAAA,IACJ;AAEA,SAAK,gBAAgB,YAAY;AAC7B,UAAI;AAGA,cAAM,cAAc;AACpB,cAAM,EAAE,UAAU,iBAAiB,IAAI,MAAM;AAAA;AAAA,UAAiC;AAAA;AAG9E,YAAI,CAAC,WAAW,KAAK,SAAS,GAAG;AAC7B,oBAAU,KAAK,WAAW,EAAE,WAAW,KAAK,CAAC;AAAA,QACjD;AAGA,cAAM,YAAY,MAAM,iBAAiB,KAAK,WAAW,KAAK,SAAS;AAGvE,cAAM,QAAQ,MAAM,SAAS;AAG7B,aAAK,SAAS,MAAM,MAAM,UAAU,EAAE,UAAU,CAAC;AAEjD,YAAI;AACA,eAAK,WAAW,MAAM,KAAK,OAAQ,qBAAqB;AAAA,YACpD,aAAa,KAAK;AAAA,YAClB,gBAAgB;AAAA,UACpB,CAAC;AAAA,QACL,QAAQ;AAEJ,eAAK,WAAW,MAAM,KAAK,OAAQ,qBAAqB;AAAA,YACpD,aAAa,KAAK;AAAA,UACtB,CAAC;AAAA,QACL;AAAA,MACJ,UAAE;AACE,aAAK,eAAe;AAAA,MACxB;AAAA,IACJ,GAAG;AAEH,UAAM,KAAK;AAAA,EACf;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,KAAK,OAAe,WAAwC;AAC9D,QAAI,UAAU,WAAW,EAAG,QAAO,CAAC;AAEpC,UAAM,KAAK,cAAc;AAGzB,UAAM,cAAc,CAAC,GAAG,IAAI,IAAI,SAAS,CAAC;AAC1C,UAAM,cAAc,oBAAI,IAAoB;AAG5C,UAAM,YAAY,YAAY,IAAI,UAAQ;AACtC,UAAI,KAAK,QAAQ;AACb,cAAM,SAAS,KAAK,OAAO,SAAS,IAAI;AAExC,cAAM,cAAc,KAAK,OAAO,SAAS,KAAK,EAAE;AAChD,cAAM,eAAe,KAAK,eAAe,MAAM;AAC/C,YAAI,OAAO,SAAS,gBAAgB,eAAe,GAAG;AAClD,iBAAO,KAAK,OAAO,WAAW,OAAO,MAAM,GAAG,YAAY,CAAC;AAAA,QAC/D;AAAA,MACJ;AACA,aAAO;AAAA,IACX,CAAC;AAGD,UAAM,SAAmB,MAAM,KAAK,SAAU,QAAQ,OAAO,SAAS;AAGtE,aAAS,IAAI,GAAG,IAAI,YAAY,QAAQ,KAAK;AACzC,kBAAY,IAAI,YAAY,CAAC,GAAG,OAAO,CAAC,KAAK,CAAC;AAAA,IAClD;AAGA,WAAO,UAAU,IAAI,SAAO,YAAY,IAAI,GAAG,KAAK,CAAC;AAAA,EACzD;AAAA;AAAA,EAGA,MAAM,QAAuB;AACzB,QAAI,KAAK,UAAU;AACf,YAAM,KAAK,SAAS,QAAQ;AAC5B,WAAK,WAAW;AAAA,IACpB;AACA,QAAI,KAAK,QAAQ;AACb,YAAM,KAAK,OAAO,UAAU;AAC5B,WAAK,SAAS;AAAA,IAClB;AAAA,EACJ;AACJ;","names":[]}