@theokit/sdk 5.13.0 → 5.15.0

package/dist/{chunk-H7G65PNI.js.map → chunk-VCQ37TC5.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/internal/memory/adapters/openai-compatible.ts"],"names":[],"mappings":";;;;;;;;AAyBA,IAAM,SAAA,GAAY,GAAA;AAClB,IAAM,WAAA,GAAc,CAAA;AASpB,IAAM,qBAAA,GAAwB,GAAA;AAC9B,IAAM,oBAAA,GAAuB,GAAA;AAkG7B,eAAsB,6BAAA,CACpB,KACA,OAAA,EAC2B;AAC3B,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,KAAA,IAAS,GAAA,CAAI,YAAA;AACnC,EAAA,MAAM,SAAS,OAAA,CAAQ,MAAA,IAAU,OAAA,CAAQ,GAAA,CAAI,IAAI,SAAS,CAAA;AAC1D,EAAA,IAAI,MAAA,KAAW,MAAA,IAAa,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG;AAC/C,IAAA,MAAM,IAAI,mBAAA,CAAoB,CAAA,EAAG,GAAA,CAAI,SAAS,CAAA,QAAA,CAAA,EAAY;AAAA,MACxD,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,MAAM,UAAA,GAAa,IAAI,UAAA,KAAe,MAAA,GAAY,QAAQ,GAAA,CAAI,GAAA,CAAI,UAAU,CAAA,GAAI,MAAA;AAChF,EAAA,MAAM,OAAA,GAAU,OAAA,CAAQ,OAAA,IAAW,UAAA,IAAc,GAAA,CAAI,cAAA;AACrD,EAAA,MAAM,SAAA,GAAY,UAAA,CAAW,OAAA,CAAQ,KAAK,CAAA;AAE1C,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,oBAAA;AAE/B,EAAA,MAAM,SAAA,GAAY,GAAA,CAAI,gBAAA,CAAiB,KAAK,CAAA;AAC5C,EAAA,IAAI,cAAc,MAAA,EAAW;AAC3B,IAAA,MAAM,IAAI,kBAAA;AAAA,MACR,CAAA,EAAG,GAAA,CAAI,EAAE,CAAA,iBAAA,EAAoB,KAAK,CAAA,6CAAA,EAAgD,MAAA,CAAO,IAAA,CAAK,GAAA,CAAI,gBAAgB,CAAA,CAAE,IAAA,CAAK,IAAI,CAAC,CAAA,2DAAA,CAAA;AAAA,MAC9H,EAAE,MAAM,yBAAA;AAA0B,KACpC;AAAA,EACF;AAKA,EAAA,MAAM,cAAA,GAAA,CAAkB,GAAA,CAAI,cAAA,IAAkB,gBAAA,EAAkB,OAAA;AAAA,IAC9D,SAAA;AAAA,IACA,mBAAmB,KAAK;AAAA,GAC1B;AAEA,EAAA,MAAM,KAAA,GAA+B;AAAA,IACnC,SAAA,EAAW,CAAA;AAAA,IACX,WAAA,EAAa,CAAA;AAAA,IACb,SAAA,EAAW,CAAA;AAAA,IACX,OAAA,EAAS;AAAA,GACX;AAEA,EAAA,OAAO;AAAA,IACL,IAAI,GAAA,CAAI,EAAA;AAAA,IACR,KAAA;AAAA,IACA,SAAA;AAAA,IACA,KAAA,EAAO,OAAO,EAAE,GAAG,KAAA,EAAM,CAAA;AAAA,IACzB,KAAA,EAAO,CAAC,KAAA,KACN,UAAA,CAAW;AAAA,MACT,KAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,SAAA;AAAA,MACA,MAAA;AAAA,MACA,OAAA;AAAA,MACA,cAAA;AAAA,MACA,SAAA;AAAA,MACA,YAAY,GAAA,CAAI,EAAA;AAAA,MAChB,OAAA,EAAS,GAAA,CAAI,OAAA,IAAW;AAAC,KAC1B;AAAA,GACL;AACF;AAgBA,eAAe,WAAW,KAAA,EAA6C;AACrE,EAAA,MAAM,EAAE,KAAA,EAAO,KAAA,EAAO,KAAA,EAAO,KAAA,EAAO,WAAU,GAAI,KAAA;AAClD,EAAA,MAAM,OAAA,GAAU,IAAI,KAAA,CAA4B,KAAA,CAAM,MAAM,CAAA;AAC5D,EAAA,MAAM,UAA+D,EAAC;AACtE,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,QAAQ,CAAA,EAAA,EAAK;AACrC,IAAA,aAAA,CAAc;AAAA,MACZ,KAAA,EAAO,CAAA;AAAA,MACP,IAAA,EAAM,KAAA,CAAM,CAAC,CAAA,IAAK,EAAA;AAAA,MAClB,KAAA;AAAA,MACA,SAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,OAAA;AAAA,MACA;AAAA,KACD,CAAA;AAAA,EACH;AACA,EAAA,MAAM,qBAAA,CAAsB,KAAA,EAAO,OAAA,EAAS,OAAO,CAAA;AACnD,EAAA,OAAO,OAAA,CAAQ,GAAA,CAAI,CAAC,CAAA,KAAM,CAAA,IAAK,IAAI,KAAA,CAAM,SAAS,CAAA,CAAE,IAAA,CAAK,CAAC,CAAC,CAAA;AAC7D;AAaA,SAAS,cAAc,IAAA,EAA+B;AACpD,EAAA,IAAI,IAAA,CAAK,IAAA,CAAK,IAAA,EAAK,CAAE,WAAW,CAAA,EAAG;AACjC,IAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,IAAI,MAAM,IAAA,CAAK,SAAS,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA;AAC3D,IAAA;AAAA,EACF;AACA,EAAA,MAAM,GAAA,GAAM,OAAA,CAAQ,IAAA,CAAK,KAAA,EAAO,KAAK,IAAI,CAAA;AACzC,EAAA,MAAM,MAAA,GAAS,IAAA,CAAK,KAAA,CAAM,GAAA,CAAI,GAAG,CAAA;AACjC,EAAA,IAAI,WAAW,MAAA,EAAW;AACxB,IAAA,IAAA,CAAK,MAAM,SAAA,IAAa,CAAA;AACxB,IAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,MAAA;AAC3B,IAAA;AAAA,EACF;AACA,EAAA,IAAA,CAAK,MAAM,WAAA,IAAe,CAAA;AAC1B,EAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,EAAE,KAAA,EAAO,IAAA,CAAK,OAAO,IAAA,EAAM,IAAA,CAAK,IAAA,EAAM,GAAA,EAAK,CAAA;AAC/D;AAKA,IAAM,sBAAA,GAAyB,CAAA;AAY/B,eAAe,qBAAA,CACb,KAAA,EACA,OAAA,EACA,OAAA,EACe;AACf,EAAA,MAAM,UAA8E,EAAC;AACrF,EAAA,KAAA,IAAS,SAAS,CAAA,EAAG,MAAA,GAAS,OAAA,CAAQ,MAAA,EAAQ,UAAU,SAAA,EAAW;AACjE,IAAA,OAAA,CAAQ,KAAK,OAAA,CAAQ,KAAA,CAAM,MAAA,EAAQ,MAAA,GAAS,SAAS,CAAC,CAAA;AAAA,EACxD;AAGA,EAAA,MAAM,kBAAA;AAAA,IAAmB,OAAA;AAAA,IAAS,sBAAA;AAAA,IAAwB,CAAC,KAAA,KACzD,YAAA,CAAa,KAAA,EAAO,OAAO,OAAO;AAAA,GACpC;AACF;AAEA,eAAe,YAAA,CACb,KAAA,EACA,KAAA,EACA,OAAA,EACe;AACf,EAAA,MAAM,OAAA,GAAU,MAAM,UAAA,CAAW;AAAA,IAC/B,QAAQ,KAAA,CAAM,MAAA;AAAA,IACd,SAAS,KAAA,CAAM,OAAA;AAAA,IACf,gBAAgB,KAAA,CAAM,cAAA;AAAA,IACtB,OAAO,KAAA,CAAM,KAAA;AAAA,IACb,QAAQ,KAAA,CAAM,GAAA,CAAI,CAAC,CAAA,KAAM,EAAE,IAAI,CAAA;AAAA,IAC/B,WAAW,KAAA,CAAM,SAAA;AAAA,IACjB,OAAO,KAAA,CAAM,KAAA;AAAA,IACb,YAAY,KAAA,CAAM,UAAA;AAAA,IAClB,SAAS,KAAA,CAAM;AAAA,GAChB,CAAA;AACD,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,QAAQ,CAAA,EAAA,EAAK;AACrC,IAAA,MAAM,IAAA,GAAO,MAAM,CAAC,CAAA;AACpB,IAAA,MAAM,MAAA,GAAS,QAAQ,CAAC,CAAA;AACxB,IAAA,IAAI,IAAA,KAAS,MAAA,IAAa,MAAA,KAAW,MAAA,EAAW;AAChD,IAAA,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,MAAA;AACtB,IAAA,KAAA,CAAM,KAAA,CAAM,GAAA,CAAI,IAAA,CAAK,GAAA,EAAK,MAAM,CAAA;AAAA,EAClC;AACF;AAcA,eAAe,WAAW,IAAA,EAAyC;AAEjE,EAAA,MAAM,GAAA,GAAM,CAAA,EAAG,IAAA,CAAK,OAAA,CAAQ,OAAA,CAAQ,OAAO,EAAE,CAAC,CAAA,EAAG,IAAA,CAAK,cAAc,CAAA,CAAA;AACpE,EAAA,IAAI,OAAA,GAAU,CAAA;AACd,EAAA,OAAO,IAAA,EAAM;AACX,IAAA,IAAA,CAAK,MAAM,SAAA,IAAa,CAAA;AACxB,IAAA,MAAM,QAAA,GAAW,MAAM,gBAAA,CAAiB,IAAA,EAAM,GAAG,CAAA;AACjD,IAAA,IAAI,QAAA,CAAS,EAAA;AACX,MAAA,OAAO,MAAM,mBAAmB,QAAA,EAAU,IAAA,CAAK,YAAY,IAAA,CAAK,cAAA,EAAgB,KAAK,OAAO,CAAA;AAC9F,IAAA,IAAI,WAAA,CAAY,QAAA,CAAS,MAAM,CAAA,IAAK,UAAU,WAAA,EAAa;AACzD,MAAA,OAAA,IAAW,CAAA;AACX,MAAA,IAAA,CAAK,MAAM,OAAA,IAAW,CAAA;AACtB,MAAA,MAAM,aAAA,GAAgB,eAAA,CAAgB,QAAA,CAAS,OAAO,CAAA;AACtD,MAAA,MAAM,KAAA;AAAA,QACJ,gBAAA,CAAiB;AAAA,UACf,OAAA;AAAA,UACA,MAAA,EAAQ,qBAAA;AAAA,UACR,KAAA,EAAO,oBAAA;AAAA,UACP,GAAI,kBAAkB,MAAA,GAAY,EAAE,cAAc,aAAA,GAAgB,GAAA,KAAS;AAAC,SAC7E;AAAA,OACH;AACA,MAAA;AAAA,IACF;AACA,IAAA,MAAM,MAAM,gBAAA,CAAiB,IAAA,EAAM,QAAQ,CAAA;AAAA,EAC7C;AACF;AAQA,eAAe,gBAAA,CAAiB,MAAoB,QAAA,EAAsC;AACxF,EAAA,MAAM,OAAO,MAAM,QAAA,CAAS,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AACjD,EAAA,IAAI,IAAA,GAAgB,IAAA;AACpB,EAAA,IAAI;AACF,IAAA,IAAA,GAAO,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EACxB,CAAA,CAAA,MAAQ;AAAA,EAER;AACA,EAAA,OAAO,wBAAA,CAAyB;AAAA,IAC9B,YAAY,IAAA,CAAK,UAAA;AAAA,IACjB,QAAQ,QAAA,CAAS,MAAA;AAAA,IACjB,IAAA;AAAA,IACA,SAAS,QAAA,CAAS,OAAA;AAAA,IAClB,UAAU,IAAA,CAAK;AAAA,GAChB,CAAA;AACH;AAEA,eAAe,gBAAA,CAAiB,MAAoB,GAAA,EAAgC;AAClF,EAAA,OAAO,IAAA,CAAK,UAAU,GAAA,EAAK;AAAA,IACzB,MAAA,EAAQ,MAAA;AAAA,IACR,OAAA,EAAS;AAAA,MACP,cAAA,EAAgB,kBAAA;AAAA,MAChB,GAAI,IAAA,CAAK,OAAA,CAAQ,WAAA,GAAc,IAAA,CAAK,MAAM,CAAA,IAAK,EAAE,aAAA,EAAe,CAAA,OAAA,EAAU,IAAA,CAAK,MAAM,CAAA,CAAA;AAAG,KAC1F;AAAA,IACA,MAAM,IAAA,CAAK,SAAA;AAAA,MACT,IAAA,CAAK,OAAA,CAAQ,IAAA,GAAO,IAAA,CAAK,OAAO,IAAA,CAAK,MAAM,CAAA,IAAK,EAAE,KAAA,EAAO,IAAA,CAAK,KAAA,EAAO,KAAA,EAAO,KAAK,MAAA;AAAO;AAC1F,GACD,CAAA;AACH;AAEA,eAAe,kBAAA,CACb,QAAA,EACA,UAAA,EACA,QAAA,EACA,OAAA,EACqB;AACrB,EAAA,MAAM,IAAA,GAAQ,MAAM,QAAA,CAAS,IAAA,EAAK;AAElC,EAAA,MAAM,UAAA,GAAa,OAAA,EAAS,OAAA,GAAU,IAAI,CAAA;AAC1C,EAAA,IAAI,UAAA,KAAe,QAAW,OAAO,UAAA;AACrC,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,IAAA,CAAK,IAAI,CAAA,EAAG;AAC7B,IAAA,MAAM,IAAI,YAAA,CAAa,CAAA,EAAG,UAAU,CAAA,CAAA,EAAI,QAAQ,CAAA,iBAAA,CAAA,EAAqB;AAAA,MACnE,IAAA,EAAM,4BAAA;AAAA,MACN,QAAA,EAAU;AAAA,QACR,QAAA,EAAU,UAAA;AAAA,QACV,QAAA;AAAA,QACA,IAAA,EAAM,iBAAA;AAAA,QACN,GAAA,EAAK;AAAA;AACP,KACD,CAAA;AAAA,EACH;AACA,EAAA,OAAO,KAAK,IAAA,CAAK,GAAA,CAAI,CAAC,CAAA,KAAM,EAAE,SAAS,CAAA;AACzC;AAEA,SAAS,YAAY,MAAA,EAAyB;AAC5C,EAAA,OAAO,MAAA,KAAW,GAAA,IAAQ,MAAA,IAAU,GAAA,IAAO,MAAA,GAAS,GAAA;AACtD;AAEA,SAAS,OAAA,CAAQ,OAAe,IAAA,EAAsB;AACpD,EAAA,OAAO,UAAA,CAAW,QAAQ,CAAA,CAAE,MAAA,CAAO,CAAA,EAAG,KAAK,CAAA,CAAA,EAAI,IAAI,CAAA,CAAE,CAAA,CAAE,MAAA,CAAO,KAAK,CAAA;AACrE;AAWA,SAAS,MAAM,EAAA,EAA2B;AACxC,EAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,YAAY,UAAA,CAAW,OAAA,EAAS,EAAE,CAAC,CAAA;AACzD","file":"chunk-H7G65PNI.js","sourcesContent":["import { createHash } from \"node:crypto\";\n\nimport { AuthenticationError, ConfigurationError, NetworkError } from \"../../../errors.js\";\nimport { mapWithConcurrency } from \"../../concurrency/map-with-concurrency.js\";\nimport { mapOpenAICompatibleError } from \"../../error-mappers/openai-compatible.js\";\nimport { parseRetryAfter } from \"../../error-mappers/shared.js\";\nimport { computeBackoffMs } from \"../../llm/retry.js\";\nimport { boundFetch } from \"../../runtime-fetch.js\";\nimport type {\n  CreateAdapterOptions,\n  EmbeddingRuntime,\n  EmbeddingRuntimeStats,\n} from \"../embedding-adapter.js\";\nimport { globalEmbeddingCache } from \"../embedding-cache.js\";\n\n/**\n * Shared factory for OpenAI-compatible embedding providers (OpenAI, Mistral,\n * DeepInfra, and anything else exposing `POST /v1/embeddings` with the\n * `{ model, input }` request shape and `{ data: [{ embedding }] }` response).\n *\n * Mirrors the inner mechanics of peer-project's batch embedding runtime.\n *\n * @internal\n */\n\nconst MAX_BATCH = 100;\nconst MAX_RETRIES = 2;\n/**\n * Deliberately tighter than the transport's 500ms/32s.\n *\n * An embedding retry sits inside a memory write on the agent's critical path and cannot be\n * cancelled (see `sleep` below), so the worst case has to be small enough to wait out: 2 retries\n * capped at 4s is at most ~8s. The transport can afford 32s because a streaming turn is already\n * long-running and IS abortable.\n */\nconst EMBED_BACKOFF_BASE_MS = 250;\nconst EMBED_BACKOFF_CAP_MS = 4_000;\n\n/**\n * What one provider adapter tells {@link createOpenAiCompatibleRuntime} about\n * its wire: where to POST, which environment variables carry the key and the\n * base URL, which model to use by default, and how wide each model's vectors\n * are.\n *\n * `id` is the provider id that ends up on errors and on\n * `EmbeddingRuntime.id`, and is also the identity the SQLite vector index stores\n * — changing it invalidates the vectors already on disk.\n *\n * `apiKeyEnv` is required, `baseUrlEnv` is not: a provider that declares no\n * base-URL variable can only be redirected by passing `baseUrl` explicitly.\n * Leave `dialect` unset for anything that speaks the OpenAI embedding wire.\n */\nexport interface OpenAiCompatibleConfig {\n  id: string;\n  defaultBaseUrl: string;\n  apiKeyEnv: string;\n  baseUrlEnv?: string;\n  defaultModel: string;\n  /**\n   * Dimension hint by model id. The chosen model MUST be in this table —\n   * unknown models throw `ConfigurationError(code: \"embedding_unknown_model\")`\n   * (EC-4 fix). This prevents vec0 virtual-table dimension mismatches that\n   * surface as cryptic errors downstream.\n   */\n  dimensionByModel: Record<string, number>;\n  /**\n   * Path component appended to `baseUrl` for the embeddings endpoint.\n   * Default `\"/v1/embeddings\"` (OpenAI canonical). REPLACES the default —\n   * does NOT concatenate. Used by DeepInfra (`\"/v1/openai/embeddings\"`),\n   * etc. (EC-2 fix).\n   */\n  embeddingsPath?: string;\n  /**\n   * theokit#159 — per-provider deviations from the OpenAI wire.\n   *\n   * The runtime was rigidly OpenAI-shaped: one auth header, one body, one response shape. Three of\n   * the ten catalog providers do not speak it, so every call they made was rejected — Azure\n   * authenticates with an `api-key` header and puts the deployment in the path, Cohere's `/v2/embed`\n   * takes `texts` and answers `{embeddings:{float}}`. Advertising a provider that cannot work is\n   * worse than not advertising it.\n   *\n   * Every hook is optional and its default is exactly the previous behaviour, so the seven\n   * providers that DO speak OpenAI are untouched.\n   */\n  dialect?: EmbeddingDialect;\n}\n\n/**\n * theokit#159 — the three points where a provider may deviate from the OpenAI embedding wire.\n *\n * Deliberately three narrow hooks rather than a general \"transform the request\" seam: these are the\n * only axes on which the supported providers actually differ, and a wider seam would invite\n * per-provider logic that belongs in the adapter.\n */\nexport interface EmbeddingDialect {\n  /** Replaces `authorization: Bearer <key>`. */\n  authHeaders?: (apiKey: string) => Record<string, string>;\n  /** Replaces the `{ model, input }` request body. */\n  body?: (model: string, inputs: ReadonlyArray<string>) => Record<string, unknown>;\n  /** Replaces reading vectors from `json.data[].embedding`. Return `undefined` when unparseable. */\n  vectors?: (json: unknown) => number[][] | undefined;\n}\n\n/**\n * Build an {@link EmbeddingRuntime} for a provider that speaks the OpenAI\n * embedding wire — or, through `cfg.dialect`, one that deviates from it in a\n * known way. Every adapter in the memory catalog is a thin call to this.\n *\n * Resolution order for each setting is explicit option, then environment\n * variable, then the config default. Two things are checked before any request\n * is sent, so a misconfiguration fails at creation rather than at first search:\n * a missing or empty key rejects with an `AuthenticationError` carrying\n * `embedding_missing_api_key`, and a model absent from `dimensionByModel`\n * rejects with a `ConfigurationError` carrying `embedding_unknown_model`. The\n * second check exists because the vector index is created at a fixed width — an\n * unknown dimension would surface much later as a vec0 mismatch.\n *\n * `{model}` in `embeddingsPath` is replaced with the URL-encoded model, which is\n * how Azure addresses a deployment in the path. `embeddingsPath` replaces the\n * default `/v1/embeddings` rather than being appended to it.\n *\n * The returned `embed` batches at 100 texts per request and runs at most 3\n * requests concurrently, preserving input order. Whitespace-only inputs never\n * reach the network and come back as an all-zero vector. Results are cached by\n * model and text; the cache defaults to a process-wide LRU of 5000 entries, so\n * pass `options.cache` when runtimes must not share entries.\n *\n * Failed requests retry up to twice on 429 and 5xx with a 50ms-per-attempt\n * linear backoff. Anything else, and a retry budget that runs out, throws the\n * typed error the OpenAI-compatible mapper builds from the status and body. A\n * response that parses but carries no vectors throws a `NetworkError` with\n * `embedding_invalid_response`. There is no request timeout here — cancellation\n * is the caller's to impose.\n */\nexport async function createOpenAiCompatibleRuntime(\n  cfg: OpenAiCompatibleConfig,\n  options: CreateAdapterOptions,\n): Promise<EmbeddingRuntime> {\n  const model = options.model ?? cfg.defaultModel;\n  const apiKey = options.apiKey ?? process.env[cfg.apiKeyEnv];\n  if (apiKey === undefined || apiKey.length === 0) {\n    throw new AuthenticationError(`${cfg.apiKeyEnv} missing`, {\n      code: \"embedding_missing_api_key\",\n    });\n  }\n  const envBaseUrl = cfg.baseUrlEnv !== undefined ? process.env[cfg.baseUrlEnv] : undefined;\n  const baseUrl = options.baseUrl ?? envBaseUrl ?? cfg.defaultBaseUrl;\n  const fetchImpl = boundFetch(options.fetch);\n  // T4.4 — default to process-wide singleton (was per-adapter instance).\n  const cache = options.cache ?? globalEmbeddingCache;\n  // EC-4: refuse unknown models to prevent vec0 dimension mismatches downstream.\n  const dimension = cfg.dimensionByModel[model];\n  if (dimension === undefined) {\n    throw new ConfigurationError(\n      `${cfg.id} adapter: model \"${model}\" is not in the dimension table. Use one of: ${Object.keys(cfg.dimensionByModel).join(\", \")}. To add it, update dimensionByModel in the adapter source.`,\n      { code: \"embedding_unknown_model\" },\n    );\n  }\n  // theokit#128: `{model}` is substituted, because Azure OpenAI addresses the deployment IN THE\n  // PATH (`/openai/deployments/{model}/embeddings`) rather than in the body. Without this the URL\n  // carried the literal placeholder and every Azure request 404'd. A path with no placeholder —\n  // every other provider — is passed through untouched.\n  const embeddingsPath = (cfg.embeddingsPath ?? \"/v1/embeddings\").replace(\n    \"{model}\",\n    encodeURIComponent(model),\n  );\n\n  const stats: EmbeddingRuntimeStats = {\n    cacheHits: 0,\n    cacheMisses: 0,\n    httpCalls: 0,\n    retries: 0,\n  };\n\n  return {\n    id: cfg.id,\n    model,\n    dimension,\n    stats: () => ({ ...stats }),\n    embed: (texts) =>\n      embedTexts({\n        texts,\n        cache,\n        stats,\n        model,\n        dimension,\n        apiKey,\n        baseUrl,\n        embeddingsPath,\n        fetchImpl,\n        providerId: cfg.id,\n        dialect: cfg.dialect ?? {},\n      }),\n  };\n}\n\ninterface EmbedTextsInput {\n  texts: ReadonlyArray<string>;\n  cache: NonNullable<CreateAdapterOptions[\"cache\"]>;\n  stats: EmbeddingRuntimeStats;\n  model: string;\n  dimension: number;\n  apiKey: string;\n  baseUrl: string;\n  embeddingsPath: string;\n  fetchImpl: typeof fetch;\n  providerId: string;\n  dialect: EmbeddingDialect;\n}\n\nasync function embedTexts(input: EmbedTextsInput): Promise<number[][]> {\n  const { texts, cache, stats, model, dimension } = input;\n  const results = new Array<number[] | undefined>(texts.length);\n  const pending: Array<{ index: number; text: string; key: string }> = [];\n  for (let i = 0; i < texts.length; i++) {\n    classifyEntry({\n      index: i,\n      text: texts[i] ?? \"\",\n      model,\n      dimension,\n      cache,\n      stats,\n      results,\n      pending,\n    });\n  }\n  await embedInBoundedBatches(input, pending, results);\n  return results.map((v) => v ?? new Array(dimension).fill(0));\n}\n\ninterface ClassifyEntryArgs {\n  index: number;\n  text: string;\n  model: string;\n  dimension: number;\n  cache: NonNullable<CreateAdapterOptions[\"cache\"]>;\n  stats: EmbeddingRuntimeStats;\n  results: Array<number[] | undefined>;\n  pending: Array<{ index: number; text: string; key: string }>;\n}\n\nfunction classifyEntry(args: ClassifyEntryArgs): void {\n  if (args.text.trim().length === 0) {\n    args.results[args.index] = new Array(args.dimension).fill(0);\n    return;\n  }\n  const key = hashKey(args.model, args.text);\n  const cached = args.cache.get(key);\n  if (cached !== undefined) {\n    args.stats.cacheHits += 1;\n    args.results[args.index] = cached;\n    return;\n  }\n  args.stats.cacheMisses += 1;\n  args.pending.push({ index: args.index, text: args.text, key });\n}\n\n// T4.3 — max concurrent embedding API requests per embed() call.\n// 3 is enough to saturate most embedding APIs without triggering\n// rate limits (each batch is already up to MAX_BATCH=100 texts).\nconst MAX_CONCURRENT_BATCHES = 3;\n\n/**\n * T4.3 — parallel embed batches (DR4 finding #3).\n *\n * Pre-T4.3 this was a serial `for` loop — each batch of 100 texts\n * awaited before the next HTTP request started. With 500 texts the\n * serial path took 5 × RTT; parallel takes max(RTT) × ceil(5/3).\n *\n * Uses bounded-concurrency `Promise.all` capped at 3 concurrent\n * HTTP requests to avoid overwhelming the embedding API.\n */\nasync function embedInBoundedBatches(\n  input: EmbedTextsInput,\n  pending: ReadonlyArray<{ index: number; text: string; key: string }>,\n  results: Array<number[] | undefined>,\n): Promise<void> {\n  const batches: Array<ReadonlyArray<{ index: number; text: string; key: string }>> = [];\n  for (let offset = 0; offset < pending.length; offset += MAX_BATCH) {\n    batches.push(pending.slice(offset, offset + MAX_BATCH));\n  }\n  // M0-2: bounded parallel via the shared ordered pool (was an inline\n  // acquire/release clone — see plan m0-foundation-expose-primitives).\n  await mapWithConcurrency(batches, MAX_CONCURRENT_BATCHES, (batch) =>\n    processBatch(input, batch, results),\n  );\n}\n\nasync function processBatch(\n  input: EmbedTextsInput,\n  batch: ReadonlyArray<{ index: number; text: string; key: string }>,\n  results: Array<number[] | undefined>,\n): Promise<void> {\n  const vectors = await embedBatch({\n    apiKey: input.apiKey,\n    baseUrl: input.baseUrl,\n    embeddingsPath: input.embeddingsPath,\n    model: input.model,\n    inputs: batch.map((b) => b.text),\n    fetchImpl: input.fetchImpl,\n    stats: input.stats,\n    providerId: input.providerId,\n    dialect: input.dialect,\n  });\n  for (let j = 0; j < batch.length; j++) {\n    const slot = batch[j];\n    const vector = vectors[j];\n    if (slot === undefined || vector === undefined) continue;\n    results[slot.index] = vector;\n    input.cache.set(slot.key, vector);\n  }\n}\n\ninterface BatchOptions {\n  apiKey: string;\n  baseUrl: string;\n  embeddingsPath: string;\n  model: string;\n  inputs: ReadonlyArray<string>;\n  fetchImpl: typeof fetch;\n  stats: EmbeddingRuntimeStats;\n  providerId: string;\n  dialect: EmbeddingDialect;\n}\n\nasync function embedBatch(opts: BatchOptions): Promise<number[][]> {\n  // EC-2: embeddingsPath REPLACES the suffix; never concatenates.\n  const url = `${opts.baseUrl.replace(/\\/$/, \"\")}${opts.embeddingsPath}`;\n  let attempt = 0;\n  while (true) {\n    opts.stats.httpCalls += 1;\n    const response = await postEmbedRequest(opts, url);\n    if (response.ok)\n      return await parseEmbedResponse(response, opts.providerId, opts.embeddingsPath, opts.dialect);\n    if (isRetryable(response.status) && attempt < MAX_RETRIES) {\n      attempt += 1;\n      opts.stats.retries += 1;\n      const retryAfterSec = parseRetryAfter(response.headers);\n      await sleep(\n        computeBackoffMs({\n          attempt,\n          baseMs: EMBED_BACKOFF_BASE_MS,\n          capMs: EMBED_BACKOFF_CAP_MS,\n          ...(retryAfterSec !== undefined ? { retryAfterMs: retryAfterSec * 1000 } : {}),\n        }),\n      );\n      continue;\n    }\n    throw await mappedEmbedError(opts, response);\n  }\n}\n\n/**\n * The terminal failure: read the body best-effort so the mapper sees the provider's error code, then\n * map it. Its own function because folding it back into the retry loop puts `embedBatch` at\n * cognitive complexity 11 against the repo max of 10 — and the sibling budget for those suppressions\n * is ratcheting DOWN, so adding one here to keep this loop whole would undo that.\n */\nasync function mappedEmbedError(opts: BatchOptions, response: Response): Promise<unknown> {\n  const text = await response.text().catch(() => \"\");\n  let body: unknown = text;\n  try {\n    body = JSON.parse(text);\n  } catch {\n    // not JSON — keep as string\n  }\n  return mapOpenAICompatibleError({\n    providerId: opts.providerId,\n    status: response.status,\n    body,\n    headers: response.headers,\n    endpoint: opts.embeddingsPath,\n  });\n}\n\nasync function postEmbedRequest(opts: BatchOptions, url: string): Promise<Response> {\n  return opts.fetchImpl(url, {\n    method: \"POST\",\n    headers: {\n      \"content-type\": \"application/json\",\n      ...(opts.dialect.authHeaders?.(opts.apiKey) ?? { authorization: `Bearer ${opts.apiKey}` }),\n    },\n    body: JSON.stringify(\n      opts.dialect.body?.(opts.model, opts.inputs) ?? { model: opts.model, input: opts.inputs },\n    ),\n  });\n}\n\nasync function parseEmbedResponse(\n  response: Response,\n  providerId: string,\n  endpoint: string,\n  dialect?: EmbeddingDialect,\n): Promise<number[][]> {\n  const json = (await response.json()) as { data?: Array<{ embedding: number[] }> };\n  // theokit#159 — a provider whose response is not OpenAI-shaped supplies its own reader.\n  const viaDialect = dialect?.vectors?.(json);\n  if (viaDialect !== undefined) return viaDialect;\n  if (!Array.isArray(json.data)) {\n    throw new NetworkError(`${providerId} ${endpoint} returned no data`, {\n      code: \"embedding_invalid_response\",\n      metadata: {\n        provider: providerId,\n        endpoint,\n        code: \"invalid_request\",\n        raw: json,\n      },\n    });\n  }\n  return json.data.map((d) => d.embedding);\n}\n\nfunction isRetryable(status: number): boolean {\n  return status === 429 || (status >= 500 && status < 600);\n}\n\nfunction hashKey(model: string, text: string): string {\n  return createHash(\"sha256\").update(`${model} ${text}`).digest(\"hex\");\n}\n\n/**\n * Not `sleepWithAbort`, and the reason is the parsimony ladder rather than an oversight.\n *\n * `EmbeddingRuntime.embed(texts)` takes no signal, and NONE of its seven callers — lance-index,\n * index-manager, vec-index, the two dreaming phases, scorers — holds one to pass. Adding the\n * parameter would add a seam nobody supplies, so cancellation is bounded instead of plumbed: at most\n * MAX_RETRIES sleeps of at most EMBED_BACKOFF_CAP_MS each. Give `embed()` a signal on the day a\n * caller has one, and this becomes `sleepWithAbort` in the same edit.\n */\nfunction sleep(ms: number): Promise<void> {\n  return new Promise((resolve) => setTimeout(resolve, ms));\n}\n"]}
+{"version":3,"sources":["../src/internal/memory/adapters/openai-compatible.ts"],"names":[],"mappings":";;;;;;;;AAyBA,IAAM,SAAA,GAAY,GAAA;AAClB,IAAM,WAAA,GAAc,CAAA;AASpB,IAAM,qBAAA,GAAwB,GAAA;AAC9B,IAAM,oBAAA,GAAuB,GAAA;AAkG7B,eAAsB,6BAAA,CACpB,KACA,OAAA,EAC2B;AAC3B,EAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,KAAA,IAAS,GAAA,CAAI,YAAA;AACnC,EAAA,MAAM,SAAS,OAAA,CAAQ,MAAA,IAAU,OAAA,CAAQ,GAAA,CAAI,IAAI,SAAS,CAAA;AAC1D,EAAA,IAAI,MAAA,KAAW,MAAA,IAAa,MAAA,CAAO,MAAA,KAAW,CAAA,EAAG;AAC/C,IAAA,MAAM,IAAI,mBAAA,CAAoB,CAAA,EAAG,GAAA,CAAI,SAAS,CAAA,QAAA,CAAA,EAAY;AAAA,MACxD,IAAA,EAAM;AAAA,KACP,CAAA;AAAA,EACH;AACA,EAAA,MAAM,UAAA,GAAa,IAAI,UAAA,KAAe,MAAA,GAAY,QAAQ,GAAA,CAAI,GAAA,CAAI,UAAU,CAAA,GAAI,MAAA;AAChF,EAAA,MAAM,OAAA,GAAU,OAAA,CAAQ,OAAA,IAAW,UAAA,IAAc,GAAA,CAAI,cAAA;AACrD,EAAA,MAAM,SAAA,GAAY,UAAA,CAAW,OAAA,CAAQ,KAAK,CAAA;AAE1C,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,oBAAA;AAE/B,EAAA,MAAM,SAAA,GAAY,GAAA,CAAI,gBAAA,CAAiB,KAAK,CAAA;AAC5C,EAAA,IAAI,cAAc,MAAA,EAAW;AAC3B,IAAA,MAAM,IAAI,kBAAA;AAAA,MACR,CAAA,EAAG,GAAA,CAAI,EAAE,CAAA,iBAAA,EAAoB,KAAK,CAAA,6CAAA,EAAgD,MAAA,CAAO,IAAA,CAAK,GAAA,CAAI,gBAAgB,CAAA,CAAE,IAAA,CAAK,IAAI,CAAC,CAAA,2DAAA,CAAA;AAAA,MAC9H,EAAE,MAAM,yBAAA;AAA0B,KACpC;AAAA,EACF;AAKA,EAAA,MAAM,cAAA,GAAA,CAAkB,GAAA,CAAI,cAAA,IAAkB,gBAAA,EAAkB,OAAA;AAAA,IAC9D,SAAA;AAAA,IACA,mBAAmB,KAAK;AAAA,GAC1B;AAEA,EAAA,MAAM,KAAA,GAA+B;AAAA,IACnC,SAAA,EAAW,CAAA;AAAA,IACX,WAAA,EAAa,CAAA;AAAA,IACb,SAAA,EAAW,CAAA;AAAA,IACX,OAAA,EAAS;AAAA,GACX;AAEA,EAAA,OAAO;AAAA,IACL,IAAI,GAAA,CAAI,EAAA;AAAA,IACR,KAAA;AAAA,IACA,SAAA;AAAA,IACA,KAAA,EAAO,OAAO,EAAE,GAAG,KAAA,EAAM,CAAA;AAAA,IACzB,KAAA,EAAO,CAAC,KAAA,KACN,UAAA,CAAW;AAAA,MACT,KAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,SAAA;AAAA,MACA,MAAA;AAAA,MACA,OAAA;AAAA,MACA,cAAA;AAAA,MACA,SAAA;AAAA,MACA,YAAY,GAAA,CAAI,EAAA;AAAA,MAChB,OAAA,EAAS,GAAA,CAAI,OAAA,IAAW;AAAC,KAC1B;AAAA,GACL;AACF;AAgBA,eAAe,WAAW,KAAA,EAA6C;AACrE,EAAA,MAAM,EAAE,KAAA,EAAO,KAAA,EAAO,KAAA,EAAO,KAAA,EAAO,WAAU,GAAI,KAAA;AAClD,EAAA,MAAM,OAAA,GAAU,IAAI,KAAA,CAA4B,KAAA,CAAM,MAAM,CAAA;AAC5D,EAAA,MAAM,UAA+D,EAAC;AACtE,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,QAAQ,CAAA,EAAA,EAAK;AACrC,IAAA,aAAA,CAAc;AAAA,MACZ,KAAA,EAAO,CAAA;AAAA,MACP,IAAA,EAAM,KAAA,CAAM,CAAC,CAAA,IAAK,EAAA;AAAA,MAClB,KAAA;AAAA,MACA,SAAA;AAAA,MACA,KAAA;AAAA,MACA,KAAA;AAAA,MACA,OAAA;AAAA,MACA;AAAA,KACD,CAAA;AAAA,EACH;AACA,EAAA,MAAM,qBAAA,CAAsB,KAAA,EAAO,OAAA,EAAS,OAAO,CAAA;AACnD,EAAA,OAAO,OAAA,CAAQ,GAAA,CAAI,CAAC,CAAA,KAAM,CAAA,IAAK,IAAI,KAAA,CAAM,SAAS,CAAA,CAAE,IAAA,CAAK,CAAC,CAAC,CAAA;AAC7D;AAaA,SAAS,cAAc,IAAA,EAA+B;AACpD,EAAA,IAAI,IAAA,CAAK,IAAA,CAAK,IAAA,EAAK,CAAE,WAAW,CAAA,EAAG;AACjC,IAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,IAAI,MAAM,IAAA,CAAK,SAAS,CAAA,CAAE,IAAA,CAAK,CAAC,CAAA;AAC3D,IAAA;AAAA,EACF;AACA,EAAA,MAAM,GAAA,GAAM,OAAA,CAAQ,IAAA,CAAK,KAAA,EAAO,KAAK,IAAI,CAAA;AACzC,EAAA,MAAM,MAAA,GAAS,IAAA,CAAK,KAAA,CAAM,GAAA,CAAI,GAAG,CAAA;AACjC,EAAA,IAAI,WAAW,MAAA,EAAW;AACxB,IAAA,IAAA,CAAK,MAAM,SAAA,IAAa,CAAA;AACxB,IAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,MAAA;AAC3B,IAAA;AAAA,EACF;AACA,EAAA,IAAA,CAAK,MAAM,WAAA,IAAe,CAAA;AAC1B,EAAA,IAAA,CAAK,OAAA,CAAQ,IAAA,CAAK,EAAE,KAAA,EAAO,IAAA,CAAK,OAAO,IAAA,EAAM,IAAA,CAAK,IAAA,EAAM,GAAA,EAAK,CAAA;AAC/D;AAKA,IAAM,sBAAA,GAAyB,CAAA;AAY/B,eAAe,qBAAA,CACb,KAAA,EACA,OAAA,EACA,OAAA,EACe;AACf,EAAA,MAAM,UAA8E,EAAC;AACrF,EAAA,KAAA,IAAS,SAAS,CAAA,EAAG,MAAA,GAAS,OAAA,CAAQ,MAAA,EAAQ,UAAU,SAAA,EAAW;AACjE,IAAA,OAAA,CAAQ,KAAK,OAAA,CAAQ,KAAA,CAAM,MAAA,EAAQ,MAAA,GAAS,SAAS,CAAC,CAAA;AAAA,EACxD;AAGA,EAAA,MAAM,kBAAA;AAAA,IAAmB,OAAA;AAAA,IAAS,sBAAA;AAAA,IAAwB,CAAC,KAAA,KACzD,YAAA,CAAa,KAAA,EAAO,OAAO,OAAO;AAAA,GACpC;AACF;AAEA,eAAe,YAAA,CACb,KAAA,EACA,KAAA,EACA,OAAA,EACe;AACf,EAAA,MAAM,OAAA,GAAU,MAAM,UAAA,CAAW;AAAA,IAC/B,QAAQ,KAAA,CAAM,MAAA;AAAA,IACd,SAAS,KAAA,CAAM,OAAA;AAAA,IACf,gBAAgB,KAAA,CAAM,cAAA;AAAA,IACtB,OAAO,KAAA,CAAM,KAAA;AAAA,IACb,QAAQ,KAAA,CAAM,GAAA,CAAI,CAAC,CAAA,KAAM,EAAE,IAAI,CAAA;AAAA,IAC/B,WAAW,KAAA,CAAM,SAAA;AAAA,IACjB,OAAO,KAAA,CAAM,KAAA;AAAA,IACb,YAAY,KAAA,CAAM,UAAA;AAAA,IAClB,SAAS,KAAA,CAAM;AAAA,GAChB,CAAA;AACD,EAAA,KAAA,IAAS,CAAA,GAAI,CAAA,EAAG,CAAA,GAAI,KAAA,CAAM,QAAQ,CAAA,EAAA,EAAK;AACrC,IAAA,MAAM,IAAA,GAAO,MAAM,CAAC,CAAA;AACpB,IAAA,MAAM,MAAA,GAAS,QAAQ,CAAC,CAAA;AACxB,IAAA,IAAI,IAAA,KAAS,MAAA,IAAa,MAAA,KAAW,MAAA,EAAW;AAChD,IAAA,OAAA,CAAQ,IAAA,CAAK,KAAK,CAAA,GAAI,MAAA;AACtB,IAAA,KAAA,CAAM,KAAA,CAAM,GAAA,CAAI,IAAA,CAAK,GAAA,EAAK,MAAM,CAAA;AAAA,EAClC;AACF;AAcA,eAAe,WAAW,IAAA,EAAyC;AAEjE,EAAA,MAAM,GAAA,GAAM,CAAA,EAAG,IAAA,CAAK,OAAA,CAAQ,OAAA,CAAQ,OAAO,EAAE,CAAC,CAAA,EAAG,IAAA,CAAK,cAAc,CAAA,CAAA;AACpE,EAAA,IAAI,OAAA,GAAU,CAAA;AACd,EAAA,OAAO,IAAA,EAAM;AACX,IAAA,IAAA,CAAK,MAAM,SAAA,IAAa,CAAA;AACxB,IAAA,MAAM,QAAA,GAAW,MAAM,gBAAA,CAAiB,IAAA,EAAM,GAAG,CAAA;AACjD,IAAA,IAAI,QAAA,CAAS,EAAA;AACX,MAAA,OAAO,MAAM,mBAAmB,QAAA,EAAU,IAAA,CAAK,YAAY,IAAA,CAAK,cAAA,EAAgB,KAAK,OAAO,CAAA;AAC9F,IAAA,IAAI,WAAA,CAAY,QAAA,CAAS,MAAM,CAAA,IAAK,UAAU,WAAA,EAAa;AACzD,MAAA,OAAA,IAAW,CAAA;AACX,MAAA,IAAA,CAAK,MAAM,OAAA,IAAW,CAAA;AACtB,MAAA,MAAM,aAAA,GAAgB,eAAA,CAAgB,QAAA,CAAS,OAAO,CAAA;AACtD,MAAA,MAAM,KAAA;AAAA,QACJ,gBAAA,CAAiB;AAAA,UACf,OAAA;AAAA,UACA,MAAA,EAAQ,qBAAA;AAAA,UACR,KAAA,EAAO,oBAAA;AAAA,UACP,GAAI,kBAAkB,MAAA,GAAY,EAAE,cAAc,aAAA,GAAgB,GAAA,KAAS;AAAC,SAC7E;AAAA,OACH;AACA,MAAA;AAAA,IACF;AACA,IAAA,MAAM,MAAM,gBAAA,CAAiB,IAAA,EAAM,QAAQ,CAAA;AAAA,EAC7C;AACF;AAQA,eAAe,gBAAA,CAAiB,MAAoB,QAAA,EAAsC;AACxF,EAAA,MAAM,OAAO,MAAM,QAAA,CAAS,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AACjD,EAAA,IAAI,IAAA,GAAgB,IAAA;AACpB,EAAA,IAAI;AACF,IAAA,IAAA,GAAO,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EACxB,CAAA,CAAA,MAAQ;AAAA,EAER;AACA,EAAA,OAAO,wBAAA,CAAyB;AAAA,IAC9B,YAAY,IAAA,CAAK,UAAA;AAAA,IACjB,QAAQ,QAAA,CAAS,MAAA;AAAA,IACjB,IAAA;AAAA,IACA,SAAS,QAAA,CAAS,OAAA;AAAA,IAClB,UAAU,IAAA,CAAK;AAAA,GAChB,CAAA;AACH;AAEA,eAAe,gBAAA,CAAiB,MAAoB,GAAA,EAAgC;AAClF,EAAA,OAAO,IAAA,CAAK,UAAU,GAAA,EAAK;AAAA,IACzB,MAAA,EAAQ,MAAA;AAAA,IACR,OAAA,EAAS;AAAA,MACP,cAAA,EAAgB,kBAAA;AAAA,MAChB,GAAI,IAAA,CAAK,OAAA,CAAQ,WAAA,GAAc,IAAA,CAAK,MAAM,CAAA,IAAK,EAAE,aAAA,EAAe,CAAA,OAAA,EAAU,IAAA,CAAK,MAAM,CAAA,CAAA;AAAG,KAC1F;AAAA,IACA,MAAM,IAAA,CAAK,SAAA;AAAA,MACT,IAAA,CAAK,OAAA,CAAQ,IAAA,GAAO,IAAA,CAAK,OAAO,IAAA,CAAK,MAAM,CAAA,IAAK,EAAE,KAAA,EAAO,IAAA,CAAK,KAAA,EAAO,KAAA,EAAO,KAAK,MAAA;AAAO;AAC1F,GACD,CAAA;AACH;AAEA,eAAe,kBAAA,CACb,QAAA,EACA,UAAA,EACA,QAAA,EACA,OAAA,EACqB;AACrB,EAAA,MAAM,IAAA,GAAQ,MAAM,QAAA,CAAS,IAAA,EAAK;AAElC,EAAA,MAAM,UAAA,GAAa,OAAA,EAAS,OAAA,GAAU,IAAI,CAAA;AAC1C,EAAA,IAAI,UAAA,KAAe,QAAW,OAAO,UAAA;AACrC,EAAA,IAAI,CAAC,KAAA,CAAM,OAAA,CAAQ,IAAA,CAAK,IAAI,CAAA,EAAG;AAC7B,IAAA,MAAM,IAAI,YAAA,CAAa,CAAA,EAAG,UAAU,CAAA,CAAA,EAAI,QAAQ,CAAA,iBAAA,CAAA,EAAqB;AAAA,MACnE,IAAA,EAAM,4BAAA;AAAA,MACN,QAAA,EAAU;AAAA,QACR,QAAA,EAAU,UAAA;AAAA,QACV,QAAA;AAAA,QACA,IAAA,EAAM,iBAAA;AAAA,QACN,GAAA,EAAK;AAAA;AACP,KACD,CAAA;AAAA,EACH;AACA,EAAA,OAAO,KAAK,IAAA,CAAK,GAAA,CAAI,CAAC,CAAA,KAAM,EAAE,SAAS,CAAA;AACzC;AAEA,SAAS,YAAY,MAAA,EAAyB;AAC5C,EAAA,OAAO,MAAA,KAAW,GAAA,IAAQ,MAAA,IAAU,GAAA,IAAO,MAAA,GAAS,GAAA;AACtD;AAEA,SAAS,OAAA,CAAQ,OAAe,IAAA,EAAsB;AACpD,EAAA,OAAO,UAAA,CAAW,QAAQ,CAAA,CAAE,MAAA,CAAO,CAAA,EAAG,KAAK,CAAA,CAAA,EAAI,IAAI,CAAA,CAAE,CAAA,CAAE,MAAA,CAAO,KAAK,CAAA;AACrE;AAWA,SAAS,MAAM,EAAA,EAA2B;AACxC,EAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,YAAY,UAAA,CAAW,OAAA,EAAS,EAAE,CAAC,CAAA;AACzD","file":"chunk-VCQ37TC5.js","sourcesContent":["import { createHash } from \"node:crypto\";\n\nimport { AuthenticationError, ConfigurationError, NetworkError } from \"../../../errors.js\";\nimport { mapWithConcurrency } from \"../../concurrency/map-with-concurrency.js\";\nimport { mapOpenAICompatibleError } from \"../../error-mappers/openai-compatible.js\";\nimport { parseRetryAfter } from \"../../error-mappers/shared.js\";\nimport { computeBackoffMs } from \"../../llm/retry.js\";\nimport { boundFetch } from \"../../runtime-fetch.js\";\nimport type {\n  CreateAdapterOptions,\n  EmbeddingRuntime,\n  EmbeddingRuntimeStats,\n} from \"../embedding-adapter.js\";\nimport { globalEmbeddingCache } from \"../embedding-cache.js\";\n\n/**\n * Shared factory for OpenAI-compatible embedding providers (OpenAI, Mistral,\n * DeepInfra, and anything else exposing `POST /v1/embeddings` with the\n * `{ model, input }` request shape and `{ data: [{ embedding }] }` response).\n *\n * Mirrors the inner mechanics of peer-project's batch embedding runtime.\n *\n * @internal\n */\n\nconst MAX_BATCH = 100;\nconst MAX_RETRIES = 2;\n/**\n * Deliberately tighter than the transport's 500ms/32s.\n *\n * An embedding retry sits inside a memory write on the agent's critical path and cannot be\n * cancelled (see `sleep` below), so the worst case has to be small enough to wait out: 2 retries\n * capped at 4s is at most ~8s. The transport can afford 32s because a streaming turn is already\n * long-running and IS abortable.\n */\nconst EMBED_BACKOFF_BASE_MS = 250;\nconst EMBED_BACKOFF_CAP_MS = 4_000;\n\n/**\n * What one provider adapter tells {@link createOpenAiCompatibleRuntime} about\n * its wire: where to POST, which environment variables carry the key and the\n * base URL, which model to use by default, and how wide each model's vectors\n * are.\n *\n * `id` is the provider id that ends up on errors and on\n * `EmbeddingRuntime.id`, and is also the identity the SQLite vector index stores\n * — changing it invalidates the vectors already on disk.\n *\n * `apiKeyEnv` is required, `baseUrlEnv` is not: a provider that declares no\n * base-URL variable can only be redirected by passing `baseUrl` explicitly.\n * Leave `dialect` unset for anything that speaks the OpenAI embedding wire.\n */\nexport interface OpenAiCompatibleConfig {\n  id: string;\n  defaultBaseUrl: string;\n  apiKeyEnv: string;\n  baseUrlEnv?: string;\n  defaultModel: string;\n  /**\n   * Dimension hint by model id. The chosen model MUST be in this table —\n   * unknown models throw `ConfigurationError(code: \"embedding_unknown_model\")`\n   * (EC-4 fix). This prevents vec0 virtual-table dimension mismatches that\n   * surface as cryptic errors downstream.\n   */\n  dimensionByModel: Record<string, number>;\n  /**\n   * Path component appended to `baseUrl` for the embeddings endpoint.\n   * Default `\"/v1/embeddings\"` (OpenAI canonical). REPLACES the default —\n   * does NOT concatenate. Used by DeepInfra (`\"/v1/openai/embeddings\"`),\n   * etc. (EC-2 fix).\n   */\n  embeddingsPath?: string;\n  /**\n   * theokit#159 — per-provider deviations from the OpenAI wire.\n   *\n   * The runtime was rigidly OpenAI-shaped: one auth header, one body, one response shape. Three of\n   * the ten catalog providers do not speak it, so every call they made was rejected — Azure\n   * authenticates with an `api-key` header and puts the deployment in the path, Cohere's `/v2/embed`\n   * takes `texts` and answers `{embeddings:{float}}`. Advertising a provider that cannot work is\n   * worse than not advertising it.\n   *\n   * Every hook is optional and its default is exactly the previous behaviour, so the seven\n   * providers that DO speak OpenAI are untouched.\n   */\n  dialect?: EmbeddingDialect;\n}\n\n/**\n * theokit#159 — the three points where a provider may deviate from the OpenAI embedding wire.\n *\n * Deliberately three narrow hooks rather than a general \"transform the request\" seam: these are the\n * only axes on which the supported providers actually differ, and a wider seam would invite\n * per-provider logic that belongs in the adapter.\n */\nexport interface EmbeddingDialect {\n  /** Replaces `authorization: Bearer <key>`. */\n  authHeaders?: (apiKey: string) => Record<string, string>;\n  /** Replaces the `{ model, input }` request body. */\n  body?: (model: string, inputs: ReadonlyArray<string>) => Record<string, unknown>;\n  /** Replaces reading vectors from `json.data[].embedding`. Return `undefined` when unparseable. */\n  vectors?: (json: unknown) => number[][] | undefined;\n}\n\n/**\n * Build an {@link EmbeddingRuntime} for a provider that speaks the OpenAI\n * embedding wire — or, through `cfg.dialect`, one that deviates from it in a\n * known way. Every adapter in the memory catalog is a thin call to this.\n *\n * Resolution order for each setting is explicit option, then environment\n * variable, then the config default. Two things are checked before any request\n * is sent, so a misconfiguration fails at creation rather than at first search:\n * a missing or empty key rejects with an `AuthenticationError` carrying\n * `embedding_missing_api_key`, and a model absent from `dimensionByModel`\n * rejects with a `ConfigurationError` carrying `embedding_unknown_model`. The\n * second check exists because the vector index is created at a fixed width — an\n * unknown dimension would surface much later as a vec0 mismatch.\n *\n * `{model}` in `embeddingsPath` is replaced with the URL-encoded model, which is\n * how Azure addresses a deployment in the path. `embeddingsPath` replaces the\n * default `/v1/embeddings` rather than being appended to it.\n *\n * The returned `embed` batches at 100 texts per request and runs at most 3\n * requests concurrently, preserving input order. Whitespace-only inputs never\n * reach the network and come back as an all-zero vector. Results are cached by\n * model and text; the cache defaults to a process-wide LRU of 5000 entries, so\n * pass `options.cache` when runtimes must not share entries.\n *\n * Failed requests retry up to twice on 429 and 5xx with a 50ms-per-attempt\n * linear backoff. Anything else, and a retry budget that runs out, throws the\n * typed error the OpenAI-compatible mapper builds from the status and body. A\n * response that parses but carries no vectors throws a `NetworkError` with\n * `embedding_invalid_response`. There is no request timeout here — cancellation\n * is the caller's to impose.\n */\nexport async function createOpenAiCompatibleRuntime(\n  cfg: OpenAiCompatibleConfig,\n  options: CreateAdapterOptions,\n): Promise<EmbeddingRuntime> {\n  const model = options.model ?? cfg.defaultModel;\n  const apiKey = options.apiKey ?? process.env[cfg.apiKeyEnv];\n  if (apiKey === undefined || apiKey.length === 0) {\n    throw new AuthenticationError(`${cfg.apiKeyEnv} missing`, {\n      code: \"embedding_missing_api_key\",\n    });\n  }\n  const envBaseUrl = cfg.baseUrlEnv !== undefined ? process.env[cfg.baseUrlEnv] : undefined;\n  const baseUrl = options.baseUrl ?? envBaseUrl ?? cfg.defaultBaseUrl;\n  const fetchImpl = boundFetch(options.fetch);\n  // T4.4 — default to process-wide singleton (was per-adapter instance).\n  const cache = options.cache ?? globalEmbeddingCache;\n  // EC-4: refuse unknown models to prevent vec0 dimension mismatches downstream.\n  const dimension = cfg.dimensionByModel[model];\n  if (dimension === undefined) {\n    throw new ConfigurationError(\n      `${cfg.id} adapter: model \"${model}\" is not in the dimension table. Use one of: ${Object.keys(cfg.dimensionByModel).join(\", \")}. To add it, update dimensionByModel in the adapter source.`,\n      { code: \"embedding_unknown_model\" },\n    );\n  }\n  // theokit#128: `{model}` is substituted, because Azure OpenAI addresses the deployment IN THE\n  // PATH (`/openai/deployments/{model}/embeddings`) rather than in the body. Without this the URL\n  // carried the literal placeholder and every Azure request 404'd. A path with no placeholder —\n  // every other provider — is passed through untouched.\n  const embeddingsPath = (cfg.embeddingsPath ?? \"/v1/embeddings\").replace(\n    \"{model}\",\n    encodeURIComponent(model),\n  );\n\n  const stats: EmbeddingRuntimeStats = {\n    cacheHits: 0,\n    cacheMisses: 0,\n    httpCalls: 0,\n    retries: 0,\n  };\n\n  return {\n    id: cfg.id,\n    model,\n    dimension,\n    stats: () => ({ ...stats }),\n    embed: (texts) =>\n      embedTexts({\n        texts,\n        cache,\n        stats,\n        model,\n        dimension,\n        apiKey,\n        baseUrl,\n        embeddingsPath,\n        fetchImpl,\n        providerId: cfg.id,\n        dialect: cfg.dialect ?? {},\n      }),\n  };\n}\n\ninterface EmbedTextsInput {\n  texts: ReadonlyArray<string>;\n  cache: NonNullable<CreateAdapterOptions[\"cache\"]>;\n  stats: EmbeddingRuntimeStats;\n  model: string;\n  dimension: number;\n  apiKey: string;\n  baseUrl: string;\n  embeddingsPath: string;\n  fetchImpl: typeof fetch;\n  providerId: string;\n  dialect: EmbeddingDialect;\n}\n\nasync function embedTexts(input: EmbedTextsInput): Promise<number[][]> {\n  const { texts, cache, stats, model, dimension } = input;\n  const results = new Array<number[] | undefined>(texts.length);\n  const pending: Array<{ index: number; text: string; key: string }> = [];\n  for (let i = 0; i < texts.length; i++) {\n    classifyEntry({\n      index: i,\n      text: texts[i] ?? \"\",\n      model,\n      dimension,\n      cache,\n      stats,\n      results,\n      pending,\n    });\n  }\n  await embedInBoundedBatches(input, pending, results);\n  return results.map((v) => v ?? new Array(dimension).fill(0));\n}\n\ninterface ClassifyEntryArgs {\n  index: number;\n  text: string;\n  model: string;\n  dimension: number;\n  cache: NonNullable<CreateAdapterOptions[\"cache\"]>;\n  stats: EmbeddingRuntimeStats;\n  results: Array<number[] | undefined>;\n  pending: Array<{ index: number; text: string; key: string }>;\n}\n\nfunction classifyEntry(args: ClassifyEntryArgs): void {\n  if (args.text.trim().length === 0) {\n    args.results[args.index] = new Array(args.dimension).fill(0);\n    return;\n  }\n  const key = hashKey(args.model, args.text);\n  const cached = args.cache.get(key);\n  if (cached !== undefined) {\n    args.stats.cacheHits += 1;\n    args.results[args.index] = cached;\n    return;\n  }\n  args.stats.cacheMisses += 1;\n  args.pending.push({ index: args.index, text: args.text, key });\n}\n\n// T4.3 — max concurrent embedding API requests per embed() call.\n// 3 is enough to saturate most embedding APIs without triggering\n// rate limits (each batch is already up to MAX_BATCH=100 texts).\nconst MAX_CONCURRENT_BATCHES = 3;\n\n/**\n * T4.3 — parallel embed batches (DR4 finding #3).\n *\n * Pre-T4.3 this was a serial `for` loop — each batch of 100 texts\n * awaited before the next HTTP request started. With 500 texts the\n * serial path took 5 × RTT; parallel takes max(RTT) × ceil(5/3).\n *\n * Uses bounded-concurrency `Promise.all` capped at 3 concurrent\n * HTTP requests to avoid overwhelming the embedding API.\n */\nasync function embedInBoundedBatches(\n  input: EmbedTextsInput,\n  pending: ReadonlyArray<{ index: number; text: string; key: string }>,\n  results: Array<number[] | undefined>,\n): Promise<void> {\n  const batches: Array<ReadonlyArray<{ index: number; text: string; key: string }>> = [];\n  for (let offset = 0; offset < pending.length; offset += MAX_BATCH) {\n    batches.push(pending.slice(offset, offset + MAX_BATCH));\n  }\n  // M0-2: bounded parallel via the shared ordered pool (was an inline\n  // acquire/release clone — see plan m0-foundation-expose-primitives).\n  await mapWithConcurrency(batches, MAX_CONCURRENT_BATCHES, (batch) =>\n    processBatch(input, batch, results),\n  );\n}\n\nasync function processBatch(\n  input: EmbedTextsInput,\n  batch: ReadonlyArray<{ index: number; text: string; key: string }>,\n  results: Array<number[] | undefined>,\n): Promise<void> {\n  const vectors = await embedBatch({\n    apiKey: input.apiKey,\n    baseUrl: input.baseUrl,\n    embeddingsPath: input.embeddingsPath,\n    model: input.model,\n    inputs: batch.map((b) => b.text),\n    fetchImpl: input.fetchImpl,\n    stats: input.stats,\n    providerId: input.providerId,\n    dialect: input.dialect,\n  });\n  for (let j = 0; j < batch.length; j++) {\n    const slot = batch[j];\n    const vector = vectors[j];\n    if (slot === undefined || vector === undefined) continue;\n    results[slot.index] = vector;\n    input.cache.set(slot.key, vector);\n  }\n}\n\ninterface BatchOptions {\n  apiKey: string;\n  baseUrl: string;\n  embeddingsPath: string;\n  model: string;\n  inputs: ReadonlyArray<string>;\n  fetchImpl: typeof fetch;\n  stats: EmbeddingRuntimeStats;\n  providerId: string;\n  dialect: EmbeddingDialect;\n}\n\nasync function embedBatch(opts: BatchOptions): Promise<number[][]> {\n  // EC-2: embeddingsPath REPLACES the suffix; never concatenates.\n  const url = `${opts.baseUrl.replace(/\\/$/, \"\")}${opts.embeddingsPath}`;\n  let attempt = 0;\n  while (true) {\n    opts.stats.httpCalls += 1;\n    const response = await postEmbedRequest(opts, url);\n    if (response.ok)\n      return await parseEmbedResponse(response, opts.providerId, opts.embeddingsPath, opts.dialect);\n    if (isRetryable(response.status) && attempt < MAX_RETRIES) {\n      attempt += 1;\n      opts.stats.retries += 1;\n      const retryAfterSec = parseRetryAfter(response.headers);\n      await sleep(\n        computeBackoffMs({\n          attempt,\n          baseMs: EMBED_BACKOFF_BASE_MS,\n          capMs: EMBED_BACKOFF_CAP_MS,\n          ...(retryAfterSec !== undefined ? { retryAfterMs: retryAfterSec * 1000 } : {}),\n        }),\n      );\n      continue;\n    }\n    throw await mappedEmbedError(opts, response);\n  }\n}\n\n/**\n * The terminal failure: read the body best-effort so the mapper sees the provider's error code, then\n * map it. Its own function because folding it back into the retry loop puts `embedBatch` at\n * cognitive complexity 11 against the repo max of 10 — and the sibling budget for those suppressions\n * is ratcheting DOWN, so adding one here to keep this loop whole would undo that.\n */\nasync function mappedEmbedError(opts: BatchOptions, response: Response): Promise<unknown> {\n  const text = await response.text().catch(() => \"\");\n  let body: unknown = text;\n  try {\n    body = JSON.parse(text);\n  } catch {\n    // not JSON — keep as string\n  }\n  return mapOpenAICompatibleError({\n    providerId: opts.providerId,\n    status: response.status,\n    body,\n    headers: response.headers,\n    endpoint: opts.embeddingsPath,\n  });\n}\n\nasync function postEmbedRequest(opts: BatchOptions, url: string): Promise<Response> {\n  return opts.fetchImpl(url, {\n    method: \"POST\",\n    headers: {\n      \"content-type\": \"application/json\",\n      ...(opts.dialect.authHeaders?.(opts.apiKey) ?? { authorization: `Bearer ${opts.apiKey}` }),\n    },\n    body: JSON.stringify(\n      opts.dialect.body?.(opts.model, opts.inputs) ?? { model: opts.model, input: opts.inputs },\n    ),\n  });\n}\n\nasync function parseEmbedResponse(\n  response: Response,\n  providerId: string,\n  endpoint: string,\n  dialect?: EmbeddingDialect,\n): Promise<number[][]> {\n  const json = (await response.json()) as { data?: Array<{ embedding: number[] }> };\n  // theokit#159 — a provider whose response is not OpenAI-shaped supplies its own reader.\n  const viaDialect = dialect?.vectors?.(json);\n  if (viaDialect !== undefined) return viaDialect;\n  if (!Array.isArray(json.data)) {\n    throw new NetworkError(`${providerId} ${endpoint} returned no data`, {\n      code: \"embedding_invalid_response\",\n      metadata: {\n        provider: providerId,\n        endpoint,\n        code: \"invalid_request\",\n        raw: json,\n      },\n    });\n  }\n  return json.data.map((d) => d.embedding);\n}\n\nfunction isRetryable(status: number): boolean {\n  return status === 429 || (status >= 500 && status < 600);\n}\n\nfunction hashKey(model: string, text: string): string {\n  return createHash(\"sha256\").update(`${model} ${text}`).digest(\"hex\");\n}\n\n/**\n * Not `sleepWithAbort`, and the reason is the parsimony ladder rather than an oversight.\n *\n * `EmbeddingRuntime.embed(texts)` takes no signal, and NONE of its seven callers — lance-index,\n * index-manager, vec-index, the two dreaming phases, scorers — holds one to pass. Adding the\n * parameter would add a seam nobody supplies, so cancellation is bounded instead of plumbed: at most\n * MAX_RETRIES sleeps of at most EMBED_BACKOFF_CAP_MS each. Give `embed()` a signal on the day a\n * caller has one, and this becomes `sleepWithAbort` in the same edit.\n */\nfunction sleep(ms: number): Promise<void> {\n  return new Promise((resolve) => setTimeout(resolve, ms));\n}\n"]}

package/dist/{chunk-SDZ4XDWS.cjs → chunk-VWYFGE27.cjs}

@@ -5,8 +5,8 @@ var chunkSWALPLFQ_cjs = require('./chunk-SWALPLFQ.cjs');
 var chunk3KIUY67P_cjs = require('./chunk-3KIUY67P.cjs');
 var chunk3WKWBYPO_cjs = require('./chunk-3WKWBYPO.cjs');
 var chunkAT2PESXB_cjs = require('./chunk-AT2PESXB.cjs');
-var chunkG7XBVM3T_cjs = require('./chunk-G7XBVM3T.cjs');
-var chunkZF2LDKQQ_cjs = require('./chunk-ZF2LDKQQ.cjs');
+var chunkMHUOVS5I_cjs = require('./chunk-MHUOVS5I.cjs');
+var chunkU3DW3WDE_cjs = require('./chunk-U3DW3WDE.cjs');
 var chunkZOFRNNZD_cjs = require('./chunk-ZOFRNNZD.cjs');
 var chunkCEGU5AVD_cjs = require('./chunk-CEGU5AVD.cjs');
 var chunkRNXXYF22_cjs = require('./chunk-RNXXYF22.cjs');
@@ -462,7 +462,7 @@ function assertWritable(fact, text) {
   }
 }
 function appendFactToMarkdown(cwd, fact, targetDir = projectMemoryDir(cwd), options = {}) {
-  return chunkZF2LDKQQ_cjs.withCwdMutex(targetDir, async () => {
+  return chunkU3DW3WDE_cjs.withCwdMutex(targetDir, async () => {
     const text = chunkRNXXYF22_cjs.redactSecrets(fact.text);
     assertWritable(fact, text);
     const maxFacts = resolveMaxFacts(options.maxFacts);
@@ -474,7 +474,7 @@ var CEILING_LOCK = { retries: 10 };
 async function withCeilingLock(targetDir, fn) {
   const lockTarget = path.resolve(targetDir);
   await promises.mkdir(path.dirname(lockTarget), { recursive: true });
-  return chunkG7XBVM3T_cjs.withFileLock(lockTarget, fn, CEILING_LOCK);
+  return chunkMHUOVS5I_cjs.withFileLock(lockTarget, fn, CEILING_LOCK);
 }
 async function writeFactFile(targetDir, fact, text, maxFacts) {
   const title = fact.title?.trim() ?? titleForFact(text);
@@ -583,8 +583,8 @@ async function writeSessionSummary(input) {
   if (input.status !== "finished") return;
   const path = sessionSummaryPath(input.memoryRoot, input.runId);
   await promises.mkdir(sessionsDir(input.memoryRoot), { recursive: true });
-  const safeUser = chunkRNXXYF22_cjs.redactSecrets(truncate(input.userText));
-  const safeAssistant = chunkRNXXYF22_cjs.redactSecrets(truncate(input.assistantText));
+  const safeUser = truncate(chunkRNXXYF22_cjs.redactSecrets(input.userText));
+  const safeAssistant = truncate(chunkRNXXYF22_cjs.redactSecrets(input.assistantText));
   const iso = new Date(input.at).toISOString();
   const body = [
     "---",
@@ -751,7 +751,7 @@ function defaultIndexPath(root) {
 }
 function requireLance() {
   try {
-    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-SDZ4XDWS.cjs', document.baseURI).href)));
+    const r = module$1.createRequire((typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('chunk-VWYFGE27.cjs', document.baseURI).href)));
     return r("@lancedb/lancedb");
   } catch (cause) {
     throw new chunkCEGU5AVD_cjs.ConfigurationError(
@@ -901,5 +901,5 @@ exports.sha256 = sha256;
 exports.truncateSnippet = truncateSnippet;
 exports.wikiDir = wikiDir;
 exports.writeSessionSummary = writeSessionSummary;
-//# sourceMappingURL=chunk-SDZ4XDWS.cjs.map
-//# sourceMappingURL=chunk-SDZ4XDWS.cjs.map
+//# sourceMappingURL=chunk-VWYFGE27.cjs.map
+//# sourceMappingURL=chunk-VWYFGE27.cjs.map