@remnic/core 9.3.726 → 9.3.727

package/dist/chunk-XGTBNVIU.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/local-model-endpoint.ts"],"sourcesContent":["/**\n * Local fine-tuned model endpoint resolver + openai-compatible caller\n * (issue #1585 model-lab integration).\n *\n * Remnic's two classification workloads — the extraction faithfulness gate\n * (#1576) and passive-correction intent detection (#1581) — can run on small\n * fine-tuned models served locally (Ollama / vLLM) instead of prompted\n * frontier LLMs. The model-lab (`model-lab/`) produces those models; this\n * module is the config-side half: it resolves the optional local endpoint\n * pointer from config and, when set, makes a minimal openai-compatible\n * chat-completions call to it.\n *\n * Design:\n *   - **Pointers are optional.** Empty defaults preserve the existing routing\n *     chain exactly (rule 39 / byte-identical pre-feature). A resolver returns\n *     `null` when the pointer is unset, so callers short-circuit before any\n *     network.\n *   - **Graceful degradation.** A local-endpoint failure (network, non-2xx,\n *     malformed body, timeout) returns `null`; callers fall back to the\n *     configured chain (checklist §4 — never block writes on a classifier).\n *   - **No inline casts.** The fetch JSON is narrowed with `in` / `typeof`\n *     guards (project rule: never assert a shape to read one property).\n *   - **Injectable fetch.** `callOpenAiCompatibleChat` takes an optional\n *     `fetchImpl` so the caller is unit-testable without a live server.\n */\n\nimport type { PluginConfig } from \"./types.js\";\n\n/** A resolved local endpoint: where to call + which model to name. */\nexport interface LocalModelEndpoint {\n  /** Base URL, e.g. `http://localhost:11434/v1` (no trailing slash required). */\n  baseUrl: string;\n  /** Model name the server expects, e.g. `remnic-faithfulness-gate-v1`. */\n  model: string;\n}\n\n/** A chat message in the openai-compatible shape this caller sends. */\nexport interface EndpointChatMessage {\n  role: \"system\" | \"user\" | \"assistant\";\n  content: string;\n}\n\n/** Options for a single local-endpoint chat call. */\nexport interface EndpointChatOptions {\n  /** Per-call timeout in ms. The caller's batch AbortSignal is NOT forwarded\n   *  (matches the local-LLM precedent: each attempt is bounded by its own\n   *  timer so one slow server can't starve the batch). */\n  timeoutMs: number;\n  /** Sampling temperature — classification wants low entropy (0.0–0.2). */\n  temperature?: number;\n  /** Max output tokens. */\n  maxTokens?: number;\n  /** Request JSON-object output (the gate emits a JSON verdict array). */\n  responseFormatJson?: boolean;\n}\n\n/** Result of a local-endpoint chat call. */\nexport interface EndpointChatResult {\n  content: string;\n  /** The model the server reports it used (falls back to the requested name). */\n  modelUsed: string;\n}\n\n/**\n * Resolve the faithfulness-gate local endpoint from config (issue #1585).\n *\n * Returns the endpoint only when BOTH a base URL and a model name are\n * configured; otherwise `null` (gate uses its existing routing chain). The\n * gate's verifier needs a model identity, so a base URL alone is intentionally\n * insufficient — silently calling whatever the server's default model is would\n * be an unreviewed behavioral change.\n */\nexport function resolveFaithfulnessGateEndpoint(\n  config: Pick<PluginConfig, \"extractionFaithfulnessModel\" | \"extractionFaithfulnessBaseUrl\">,\n): LocalModelEndpoint | null {\n  const baseUrl = config.extractionFaithfulnessBaseUrl.trim();\n  const model = config.extractionFaithfulnessModel.trim();\n  if (!baseUrl || !model) return null;\n  return { baseUrl, model };\n}\n\n/**\n * Resolve the correction-intent local endpoint from config (issue #1581/#1585).\n *\n * Same contract as the gate resolver. Consumption by the model-backed\n * detector is the consuming child's job — the rule-based\n * `detectPassiveCorrections` remains the default path; this pointer lets a\n * future model-backed detector route to a fine-tuned local classifier when an\n * operator configures one.\n */\nexport function resolveCorrectionIntentEndpoint(\n  config: Pick<PluginConfig, \"correctionIntentModel\" | \"correctionIntentBaseUrl\">,\n): LocalModelEndpoint | null {\n  const baseUrl = config.correctionIntentBaseUrl.trim();\n  const model = config.correctionIntentModel.trim();\n  if (!baseUrl || !model) return null;\n  return { baseUrl, model };\n}\n\ntype FetchLike = typeof fetch;\n\n/**\n * Make an openai-compatible chat-completions call to a local endpoint.\n *\n * POSTs to `${baseUrl}/chat/completions` with `{ model, messages, ... }` and\n * returns the first choice's content. Any failure (non-2xx, network error,\n * malformed body, timeout) returns `null` so callers fall back gracefully.\n *\n * The response is narrowed with runtime guards — never an unchecked cast — so\n * a surprising server response fails closed (`null`) rather than reading a\n * fabricated field.\n */\nexport async function callOpenAiCompatibleChat(\n  endpoint: LocalModelEndpoint,\n  messages: EndpointChatMessage[],\n  options: EndpointChatOptions,\n  fetchImpl: FetchLike = fetch,\n): Promise<EndpointChatResult | null> {\n  const url = joinBaseUrl(endpoint.baseUrl, \"/chat/completions\");\n  const controller = new AbortController();\n  const timer = setTimeout(() => controller.abort(), options.timeoutMs);\n  try {\n    const body: Record<string, unknown> = {\n      model: endpoint.model,\n      messages,\n      temperature: options.temperature ?? 0.1,\n    };\n    if (typeof options.maxTokens === \"number\") {\n      body.max_tokens = options.maxTokens;\n    }\n    if (options.responseFormatJson) {\n      body.response_format = { type: \"json_object\" };\n    }\n    let response: Response;\n    try {\n      response = await fetchImpl(url, {\n        method: \"POST\",\n        headers: { \"content-type\": \"application/json\" },\n        body: JSON.stringify(body),\n        signal: controller.signal,\n      });\n    } catch {\n      // Network error or abort — fail closed, let the caller fall back.\n      return null;\n    }\n    if (!response.ok) return null;\n    const parsed: unknown = await response.json().catch(() => null);\n    const content = extractFirstChoiceContent(parsed);\n    if (content === null) return null;\n    return { content, modelUsed: extractModel(parsed) ?? endpoint.model };\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/** Join a base URL and a path, tolerating a trailing slash on the base. */\nfunction joinBaseUrl(base: string, path: string): string {\n  return base.endsWith(\"/\") ? `${base.slice(0, -1)}${path}` : `${base}${path}`;\n}\n\n/**\n * Narrow the openai-compatible response and pull the first choice's content.\n *\n * `in` / `typeof` guards only — no `as` cast (project rule). Returns `null`\n * for any shape that isn't `{ choices: [{ message: { content: string } }, …] }`.\n */\nfunction extractFirstChoiceContent(data: unknown): string | null {\n  if (!data || typeof data !== \"object\") return null;\n  // After `in`, TS narrows the property to `unknown` — read it directly, no cast.\n  if (!(\"choices\" in data) || !Array.isArray(data.choices)) return null;\n  const choices: unknown[] = data.choices;\n  for (const choice of choices) {\n    if (!choice || typeof choice !== \"object\") continue;\n    if (!(\"message\" in choice)) continue;\n    const message = choice.message;\n    if (!message || typeof message !== \"object\") continue;\n    if (!(\"content\" in message)) continue;\n    const content = message.content;\n    if (typeof content === \"string\" && content.length > 0) return content;\n  }\n  return null;\n}\n\n/** Narrow the `model` field from the response (best-effort, may be absent). */\nfunction extractModel(data: unknown): string | null {\n  if (!data || typeof data !== \"object\" || !(\"model\" in data)) return null;\n  const model = data.model;\n  return typeof model === \"string\" && model.length > 0 ? model : null;\n}\n"],"mappings":";AAwEO,SAAS,gCACd,QAC2B;AAC3B,QAAM,UAAU,OAAO,8BAA8B,KAAK;AAC1D,QAAM,QAAQ,OAAO,4BAA4B,KAAK;AACtD,MAAI,CAAC,WAAW,CAAC,MAAO,QAAO;AAC/B,SAAO,EAAE,SAAS,MAAM;AAC1B;AAWO,SAAS,gCACd,QAC2B;AAC3B,QAAM,UAAU,OAAO,wBAAwB,KAAK;AACpD,QAAM,QAAQ,OAAO,sBAAsB,KAAK;AAChD,MAAI,CAAC,WAAW,CAAC,MAAO,QAAO;AAC/B,SAAO,EAAE,SAAS,MAAM;AAC1B;AAeA,eAAsB,yBACpB,UACA,UACA,SACA,YAAuB,OACa;AACpC,QAAM,MAAM,YAAY,SAAS,SAAS,mBAAmB;AAC7D,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,QAAQ,SAAS;AACpE,MAAI;AACF,UAAM,OAAgC;AAAA,MACpC,OAAO,SAAS;AAAA,MAChB;AAAA,MACA,aAAa,QAAQ,eAAe;AAAA,IACtC;AACA,QAAI,OAAO,QAAQ,cAAc,UAAU;AACzC,WAAK,aAAa,QAAQ;AAAA,IAC5B;AACA,QAAI,QAAQ,oBAAoB;AAC9B,WAAK,kBAAkB,EAAE,MAAM,cAAc;AAAA,IAC/C;AACA,QAAI;AACJ,QAAI;AACF,iBAAW,MAAM,UAAU,KAAK;AAAA,QAC9B,QAAQ;AAAA,QACR,SAAS,EAAE,gBAAgB,mBAAmB;AAAA,QAC9C,MAAM,KAAK,UAAU,IAAI;AAAA,QACzB,QAAQ,WAAW;AAAA,MACrB,CAAC;AAAA,IACH,QAAQ;AAEN,aAAO;AAAA,IACT;AACA,QAAI,CAAC,SAAS,GAAI,QAAO;AACzB,UAAM,SAAkB,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,IAAI;AAC9D,UAAM,UAAU,0BAA0B,MAAM;AAChD,QAAI,YAAY,KAAM,QAAO;AAC7B,WAAO,EAAE,SAAS,WAAW,aAAa,MAAM,KAAK,SAAS,MAAM;AAAA,EACtE,UAAE;AACA,iBAAa,KAAK;AAAA,EACpB;AACF;AAGA,SAAS,YAAY,MAAc,MAAsB;AACvD,SAAO,KAAK,SAAS,GAAG,IAAI,GAAG,KAAK,MAAM,GAAG,EAAE,CAAC,GAAG,IAAI,KAAK,GAAG,IAAI,GAAG,IAAI;AAC5E;AAQA,SAAS,0BAA0B,MAA8B;AAC/D,MAAI,CAAC,QAAQ,OAAO,SAAS,SAAU,QAAO;AAE9C,MAAI,EAAE,aAAa,SAAS,CAAC,MAAM,QAAQ,KAAK,OAAO,EAAG,QAAO;AACjE,QAAM,UAAqB,KAAK;AAChC,aAAW,UAAU,SAAS;AAC5B,QAAI,CAAC,UAAU,OAAO,WAAW,SAAU;AAC3C,QAAI,EAAE,aAAa,QAAS;AAC5B,UAAM,UAAU,OAAO;AACvB,QAAI,CAAC,WAAW,OAAO,YAAY,SAAU;AAC7C,QAAI,EAAE,aAAa,SAAU;AAC7B,UAAM,UAAU,QAAQ;AACxB,QAAI,OAAO,YAAY,YAAY,QAAQ,SAAS,EAAG,QAAO;AAAA,EAChE;AACA,SAAO;AACT;AAGA,SAAS,aAAa,MAA8B;AAClD,MAAI,CAAC,QAAQ,OAAO,SAAS,YAAY,EAAE,WAAW,MAAO,QAAO;AACpE,QAAM,QAAQ,KAAK;AACnB,SAAO,OAAO,UAAU,YAAY,MAAM,SAAS,IAAI,QAAQ;AACjE;","names":[]}