@effect-uai/jina 0.6.0 → 0.7.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"JinaEmbedding.d.mts","names":[],"sources":["../src/JinaEmbedding.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;
|
|
1
|
+
{"version":3,"file":"JinaEmbedding.d.mts","names":[],"sources":["../src/JinaEmbedding.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;KAoCY,QAAA;;;;AAAZ;;;;;AA8BA;;;;;AAEA;;;;;;KAFY,YAAA;AAAA,KAEA,gBAAA,GAAmB,IAAA,CAAK,kBAAA;EAUF,2EARvB,KAAA,EAAO,kBAAA;EAFa;;;;;EAAA,SAQpB,IAAA,GAAO,QAAA,EAEP;EAAA,SAAA,QAAA,GAAW,YAAA;AAAA;AAAA,KAGV,oBAAA,GAAuB,IAAA,CAAK,gBAAA;EAAA,SAC7B,MAAA,EAAQ,aAAA,CAAc,UAAA;AAAA;AAAA,KAGrB,oBAAA;EAAA,SACD,KAAA,aAAkB,YAAA,0BACzB,OAAA,EAAS,IAAA,CAAK,gBAAA;IAAA,SAA2C,QAAA,GAAW,CAAA;EAAA,MACjE,MAAA,CAAO,MAAA,CAAO,aAAA,CAAc,CAAA,GAAI,OAAA,CAAQ,OAAA;EAAA,SACpC,SAAA,aAAsB,YAAA,0BAC7B,OAAA,EAAS,IAAA,CAAK,oBAAA;IAAA,SAA+C,QAAA,GAAW,CAAA;EAAA,MACrE,MAAA,CAAO,MAAA,CAAO,iBAAA,CAAkB,CAAA,GAAI,OAAA,CAAQ,OAAA;AAAA;AAAA,cAClD,kBAAA;;;;;;cAOY,aAAA,SAAsB,kBAAA;AAAA,KAIvB,MAAA;EAAA,SACD,MAAA,EAAQ,QAAA,CAAS,QAAA;EAAA,SACjB,OAAA;AAAA;;;;;cA8YE,IAAA,GACX,GAAA,EAAK,MAAA,KACJ,MAAA,CAAO,MAAA,CAAO,oBAAA,SAA6B,UAAA,CAAW,UAAA;;;;;;;;cAuB5C,KAAA,GACX,GAAA,EAAK,MAAA,KACJ,KAAA,CAAM,KAAA,CAAM,aAAA,GAAgB,cAAA,SAAuB,UAAA,CAAW,UAAA"}
|
package/dist/JinaEmbedding.mjs
CHANGED
|
@@ -2,7 +2,7 @@ import { t as __exportAll } from "./chunk-CfYAbeIz.mjs";
|
|
|
2
2
|
import { Context, Effect, Encoding, Layer, Match, Result, Schema } from "effect";
|
|
3
3
|
import { HttpClient, HttpClientRequest } from "effect/unstable/http";
|
|
4
4
|
import * as AiError from "@effect-uai/core/AiError";
|
|
5
|
-
import { EmbeddingModel } from "@effect-uai/core/EmbeddingModel";
|
|
5
|
+
import { EmbeddingModel, assertEncoding } from "@effect-uai/core/EmbeddingModel";
|
|
6
6
|
//#region src/JinaEmbedding.ts
|
|
7
7
|
var JinaEmbedding_exports = /* @__PURE__ */ __exportAll({
|
|
8
8
|
JinaEmbedding: () => JinaEmbedding,
|
|
@@ -22,10 +22,10 @@ var JinaEmbedding = class extends Context.Service()("@betalyra/effect-uai/provid
|
|
|
22
22
|
*/
|
|
23
23
|
const imageSourceToItem = Match.type().pipe(Match.tag("url", (s) => ({ image: s.url })), Match.tag("base64", (s) => ({ image: s.base64 })), Match.tag("bytes", (s) => ({ image: Encoding.encodeBase64(s.bytes) })), Match.exhaustive);
|
|
24
24
|
const contentPartToItem = Match.type().pipe(Match.when({ text: Match.string }, ({ text }) => ({ text })), Match.when({ image: Match.any }, ({ image }) => imageSourceToItem(image)), Match.exhaustive);
|
|
25
|
-
const multiPartContentRejected = new AiError.
|
|
25
|
+
const multiPartContentRejected = new AiError.Unsupported({
|
|
26
26
|
provider: "jina",
|
|
27
|
-
|
|
28
|
-
|
|
27
|
+
capability: "multiPartInput",
|
|
28
|
+
reason: "Jina treats each input[] entry as one item; multi-part `content[]` can't be fused into one vector. Split into separate `inputs[]` entries."
|
|
29
29
|
});
|
|
30
30
|
/**
|
|
31
31
|
* Project an `EmbedInput` to one wire item. A single-part `content[]` is
|
|
@@ -246,19 +246,19 @@ const layer = (cfg) => {
|
|
|
246
246
|
const generic = Layer.effect(EmbeddingModel, Effect.map(make(cfg), (s) => ({
|
|
247
247
|
embed: (req) => {
|
|
248
248
|
const task = mapGenericTask(req.task);
|
|
249
|
-
return s.embed({
|
|
249
|
+
return assertEncoding(req.encoding, ["float32", "binary"], "jina").pipe(Effect.andThen(s.embed({
|
|
250
250
|
...req,
|
|
251
251
|
model: req.model,
|
|
252
252
|
...task !== void 0 && { task }
|
|
253
|
-
});
|
|
253
|
+
})));
|
|
254
254
|
},
|
|
255
255
|
embedMany: (req) => {
|
|
256
256
|
const task = mapGenericTask(req.task);
|
|
257
|
-
return s.embedMany({
|
|
257
|
+
return assertEncoding(req.encoding, ["float32", "binary"], "jina").pipe(Effect.andThen(s.embedMany({
|
|
258
258
|
...req,
|
|
259
259
|
model: req.model,
|
|
260
260
|
...task !== void 0 && { task }
|
|
261
|
-
});
|
|
261
|
+
})));
|
|
262
262
|
}
|
|
263
263
|
})));
|
|
264
264
|
return Layer.merge(typed, generic);
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"JinaEmbedding.mjs","names":[],"sources":["../src/JinaEmbedding.ts"],"sourcesContent":["import { Context, Effect, Encoding, Layer, Match, Redacted, Result, Schema } from \"effect\"\nimport { HttpClient, HttpClientRequest } from \"effect/unstable/http\"\nimport * as AiError from \"@effect-uai/core/AiError\"\nimport type { Embedding, EmbedContentPart, EmbedInput, Usage } from \"@effect-uai/core/Embedding\"\n\n// Internal open-union response shape — impls return this; the typed\n// `make()` wrapper casts to the narrowed `EmbedResponse<E>`.\ntype AnyEmbedResponse = { readonly embedding: Embedding; readonly usage: Usage }\ntype AnyEmbedManyResponse = { readonly embeddings: ReadonlyArray<Embedding>; readonly usage: Usage }\nimport {\n type CommonEmbedManyRequest,\n type CommonEmbedRequest,\n type EmbedEncoding,\n EmbeddingModel,\n type EmbeddingModelService,\n type EmbedManyResponse,\n type EmbedResponse,\n} from \"@effect-uai/core/EmbeddingModel\"\nimport type { ImageSource } from \"@effect-uai/core/Image\"\nimport type { JinaEmbeddingModel } from \"./models.js\"\n\n// ---------------------------------------------------------------------------\n// Public types\n// ---------------------------------------------------------------------------\n\n/**\n * Jina's task vocabulary. v4 and v3/v5 share the dotted-pair form:\n * - `retrieval.query` / `retrieval.passage` — asymmetric retrieval.\n * - `text-matching` — symmetric similarity between two texts.\n * - `code.query` / `code.passage` — code retrieval / matching (v4-only).\n * - `classification` / `separation` — v3 / v5 only.\n *\n * The `(string & {})` tail accepts any string so newly-released task\n * types work without an SDK update.\n */\nexport type JinaTask =\n | \"retrieval.query\"\n | \"retrieval.passage\"\n | \"text-matching\"\n | \"code.query\"\n | \"code.passage\"\n | \"classification\"\n | \"separation\"\n // eslint-disable-next-line @typescript-eslint/ban-types\n | (string & {})\n\n/**\n * Jina's supported `encoding` values. Provider-narrowed from the cross-\n * provider `EmbedEncoding` union.\n *\n * Compatibility with models is checked at the response level rather than\n * pre-flight: if you ask for `multivector` against a model that doesn't\n * produce one, Jina's API rejects the request with a clear error; if\n * you ask for `sparse` and the chosen model returns dense, our decoder\n * fails with a typed mismatch. No hardcoded model-encoding table to\n * maintain.\n *\n * Wire mapping:\n * - `multivector` → request sets `return_multivector: true`\n * (`jina-embeddings-v4` + ColBERT models).\n * - `sparse` → no flag; the model itself (`elser-v2`) determines the\n * output shape.\n * - `binary` → request sets `embedding_type: \"binary\"`.\n * - `float32` → no flag; default JSON `number[]`.\n */\nexport type JinaEncoding = \"float32\" | \"binary\" | \"sparse\" | \"multivector\"\n\nexport type JinaEmbedRequest = Omit<CommonEmbedRequest, \"model\" | \"task\" | \"encoding\"> & {\n /** Narrows `CommonEmbedRequest.model` to the typed Jina union. */\n readonly model: JinaEmbeddingModel\n /**\n * Widens the cross-provider `\"query\" | \"document\"` to Jina's task enum.\n * Required by v3+ for retrieval-quality results. v4 unifies query and\n * document under `retrieval`.\n */\n readonly task?: JinaTask\n /** Narrows `CommonEmbedRequest.encoding` to Jina's supported set. */\n readonly encoding?: JinaEncoding\n}\n\nexport type JinaEmbedManyRequest = Omit<JinaEmbedRequest, \"input\"> & {\n readonly inputs: ReadonlyArray<EmbedInput>\n}\n\nexport type JinaEmbeddingService = {\n readonly embed: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) => Effect.Effect<EmbedResponse<E>, AiError.AiError>\n readonly embedMany: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) => Effect.Effect<EmbedManyResponse<E>, AiError.AiError>\n}\n\n/**\n * Provider-typed service tag. Yield this for Jina-typed embedding\n * requests; yield the generic `EmbeddingModel` tag for provider-portable\n * code. Both are registered by `layer`.\n */\nexport class JinaEmbedding extends Context.Service<JinaEmbedding, JinaEmbeddingService>()(\n \"@betalyra/effect-uai/providers/jina/JinaEmbedding\",\n) {}\n\nexport type Config = {\n readonly apiKey: Redacted.Redacted\n readonly baseUrl?: string\n}\n\n// ---------------------------------------------------------------------------\n// Codec - request body\n// ---------------------------------------------------------------------------\n\ntype WireTextItem = {\n readonly text: string\n}\ntype WireImageItem = {\n readonly image: string\n}\ntype WireItem = WireTextItem | WireImageItem\n\n/**\n * Jina v4 accepts `{ image: <url> }` or `{ image: <base64-string> }`.\n * The server detects which form by looking at the value. We pass URLs\n * through directly and base64-encode raw bytes.\n */\nconst imageSourceToItem: (s: ImageSource) => WireItem = Match.type<ImageSource>().pipe(\n Match.tag(\"url\", (s): WireItem => ({ image: s.url })),\n Match.tag(\"base64\", (s): WireItem => ({ image: s.base64 })),\n Match.tag(\"bytes\", (s): WireItem => ({ image: Encoding.encodeBase64(s.bytes) })),\n Match.exhaustive,\n)\n\nconst contentPartToItem: (part: EmbedContentPart) => WireItem = Match.type<EmbedContentPart>().pipe(\n Match.when({ text: Match.string }, ({ text }): WireItem => ({ text })),\n Match.when({ image: Match.any }, ({ image }) => imageSourceToItem(image)),\n Match.exhaustive,\n)\n\nconst multiPartContentRejected: AiError.AiError = new AiError.InvalidRequest({\n provider: \"jina\",\n param: \"input.content\",\n raw: \"Jina treats each input[] entry as one item; multi-part `content[]` would lose the grouping. Split into separate `inputs[]` entries.\",\n})\n\n/**\n * Project an `EmbedInput` to one wire item. A single-part `content[]` is\n * unwrapped; multi-part `content[]` is rejected because Jina's flat\n * `input[]` would lose the grouping.\n */\nconst inputToItem: (input: EmbedInput) => Effect.Effect<WireItem, AiError.AiError> =\n Match.type<EmbedInput>().pipe(\n Match.when(\n Match.string,\n (s): Effect.Effect<WireItem, AiError.AiError> => Effect.succeed({ text: s }),\n ),\n Match.when(\n { text: Match.string },\n ({ text }): Effect.Effect<WireItem, AiError.AiError> => Effect.succeed({ text }),\n ),\n Match.when(\n { image: Match.any },\n ({ image }): Effect.Effect<WireItem, AiError.AiError> =>\n Effect.succeed(imageSourceToItem(image)),\n ),\n Match.when(\n { content: Match.any },\n ({ content }): Effect.Effect<WireItem, AiError.AiError> =>\n content.length === 1\n ? Effect.succeed(contentPartToItem(content[0]!))\n : Effect.fail(multiPartContentRejected),\n ),\n Match.exhaustive,\n )\n\ntype WireBody = {\n readonly model: string\n readonly input: ReadonlyArray<WireItem>\n readonly task?: string\n readonly dimensions?: number\n readonly embedding_type?: \"float\" | \"base64\" | \"binary\" | \"ubinary\"\n readonly return_multivector?: boolean\n}\n\ntype WireQuant = WireBody[\"embedding_type\"]\n\n/**\n * Map our `JinaEncoding` to Jina's `embedding_type` wire field. Only\n * `binary` actually changes the wire transport; the others either omit\n * the field (`float32`) or ignore it (`sparse` / `multivector` use\n * separate signalling — model choice and `return_multivector`).\n */\nconst encodingToQuant: (encoding: JinaEncoding | undefined) => WireQuant = Match.type<\n JinaEncoding | undefined\n>().pipe(\n Match.when(\"binary\", (): WireQuant => \"binary\"),\n Match.orElse((): WireQuant => undefined),\n)\n\nconst buildBody = (request: {\n readonly model: JinaEmbeddingModel\n readonly task?: JinaTask\n readonly dimensions?: number\n readonly encoding?: JinaEncoding\n readonly items: ReadonlyArray<WireItem>\n}): WireBody => {\n const quant = encodingToQuant(request.encoding)\n return {\n model: request.model,\n input: request.items,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(quant !== undefined && { embedding_type: quant }),\n ...(request.encoding === \"multivector\" && { return_multivector: true }),\n }\n}\n\n// ---------------------------------------------------------------------------\n// Codec - response\n// ---------------------------------------------------------------------------\n\n/**\n * Wire response payloads. Split into four discriminated schemas so the\n * decoder can narrow precisely via `Schema.is`:\n *\n * - **Dense float** (`object: \"embedding\"`, `embedding: number[]`) —\n * default JSON form.\n * - **Dense base64** (`object: \"embedding\"`, `embedding: string`) — set\n * when `embedding_type` was `base64` or `binary`.\n * - **Sparse** (`object: \"embedding\"`, `embedding: Record<string,\n * number>`) — token-keyed weights from an ELSER-style model.\n * - **Multivector** (`object: \"embeddings\"`, plural) — token-level\n * vectors when `return_multivector: true` or a ColBERT model is used.\n */\nconst WireDenseFloatPayload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.Array(Schema.Number),\n})\ntype WireDenseFloatPayload = typeof WireDenseFloatPayload.Type\n\nconst WireDenseBase64Payload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.String,\n})\ntype WireDenseBase64Payload = typeof WireDenseBase64Payload.Type\n\nconst WireSparsePayload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.Record(Schema.String, Schema.Number),\n})\ntype WireSparsePayload = typeof WireSparsePayload.Type\n\nconst WireMultiPayload = Schema.Struct({\n object: Schema.Literal(\"embeddings\"),\n index: Schema.Number,\n embeddings: Schema.Array(Schema.Array(Schema.Number)),\n tokenized_input: Schema.optional(Schema.Array(Schema.String)),\n})\ntype WireMultiPayload = typeof WireMultiPayload.Type\n\n// Order matters: more specific schemas first. Dense float (array) and\n// multivector (different `object`) are unambiguous; sparse vs base64\n// disambiguate by the `embedding` field type.\nconst WirePayload = Schema.Union([\n WireDenseFloatPayload,\n WireSparsePayload,\n WireDenseBase64Payload,\n WireMultiPayload,\n])\ntype WirePayload = typeof WirePayload.Type\n\nconst isMultiPayload = Schema.is(WireMultiPayload)\nconst isSparsePayload = Schema.is(WireSparsePayload)\nconst isDenseBase64Payload = Schema.is(WireDenseBase64Payload)\n\nconst WireUsage = Schema.Struct({\n total_tokens: Schema.Number,\n prompt_tokens: Schema.optional(Schema.Number),\n})\ntype WireUsage = typeof WireUsage.Type\n\nconst WireResponse = Schema.Struct({\n data: Schema.Array(WirePayload),\n model: Schema.String,\n usage: WireUsage,\n})\n\nconst transportFailure = (cause: unknown): AiError.AiError =>\n new AiError.Unavailable({ provider: \"jina\", raw: cause })\n\nconst encodingMismatch = (expected: JinaEncoding, got: string): AiError.AiError =>\n new AiError.InvalidRequest({\n provider: \"jina\",\n param: \"encoding\",\n raw: `requested encoding=\"${expected}\" but the response contains a ${got} embedding - confirm the chosen model supports the requested encoding`,\n })\n\n/** Decode a base64 wire string to the right dense `Embedding` arm. */\nconst decodeBase64Embedding = (\n b64: string,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<Embedding, AiError.AiError> =>\n Result.match(Encoding.decodeBase64(b64), {\n onFailure: (cause) => Effect.fail(transportFailure(cause)),\n onSuccess: (bytes) =>\n Match.value(encoding).pipe(\n Match.when(\n \"binary\",\n (): Effect.Effect<Embedding, AiError.AiError> =>\n Effect.succeed({ _tag: \"binary\", vector: bytes }),\n ),\n Match.orElse(\n (): Effect.Effect<Embedding, AiError.AiError> =>\n Effect.succeed({\n _tag: \"float32\",\n vector: new Float32Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 4),\n }),\n ),\n ),\n })\n\n/**\n * Project a wire payload to our `Embedding` discriminated union and\n * verify the requested encoding matches what came back.\n */\nconst payloadToEmbedding = (\n payload: WirePayload,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<Embedding, AiError.AiError> => {\n if (isMultiPayload(payload)) {\n return encoding === \"multivector\"\n ? Effect.succeed<Embedding>({\n _tag: \"multivector\",\n vectors: payload.embeddings.map((v) => Float32Array.from(v)),\n })\n : Effect.fail(encodingMismatch(encoding ?? \"float32\", \"multivector\"))\n }\n if (isSparsePayload(payload)) {\n return encoding === \"sparse\"\n ? Effect.succeed<Embedding>({ _tag: \"sparse\", weights: payload.embedding })\n : Effect.fail(encodingMismatch(encoding ?? \"float32\", \"sparse\"))\n }\n // Dense (float array or base64 string).\n if (encoding === \"multivector\" || encoding === \"sparse\") {\n return Effect.fail(encodingMismatch(encoding, \"dense\"))\n }\n return isDenseBase64Payload(payload)\n ? decodeBase64Embedding(payload.embedding, encoding)\n : Effect.succeed<Embedding>({\n _tag: \"float32\",\n vector: Float32Array.from(payload.embedding),\n })\n}\n\nconst indexOf = (p: WirePayload): number => p.index\n\nconst usageOf = (u: WireUsage): Usage => ({ inputTokens: u.total_tokens })\n\n// ---------------------------------------------------------------------------\n// HTTP errors\n// ---------------------------------------------------------------------------\n\nconst httpStatusError = (status: number, body: string): AiError.AiError => {\n const provider = \"jina\"\n const raw = body\n if (status === 429) return new AiError.RateLimited({ provider, raw })\n if (status === 408 || status === 504) return new AiError.Timeout({ provider, raw })\n if (status === 401) return new AiError.AuthFailed({ provider, subtype: \"auth\", raw })\n if (status === 403) return new AiError.AuthFailed({ provider, subtype: \"permission\", raw })\n if (status === 402) return new AiError.AuthFailed({ provider, subtype: \"billing\", raw })\n if (status === 413) return new AiError.ContextLengthExceeded({ provider, raw })\n if (status >= 500) return new AiError.Unavailable({ provider, status, raw })\n return new AiError.InvalidRequest({ provider, raw })\n}\n\n// ---------------------------------------------------------------------------\n// HTTP plumbing\n// ---------------------------------------------------------------------------\n\nconst baseUrl = (cfg: Config): string => cfg.baseUrl ?? \"https://api.jina.ai/v1\"\n\nconst postEmbed = (\n cfg: Config,\n body: WireBody,\n): Effect.Effect<typeof WireResponse.Type, AiError.AiError, HttpClient.HttpClient> =>\n Effect.gen(function* () {\n const client = yield* HttpClient.HttpClient\n const httpRequest = HttpClientRequest.post(`${baseUrl(cfg)}/embeddings`).pipe(\n HttpClientRequest.bearerToken(cfg.apiKey),\n HttpClientRequest.bodyJsonUnsafe(body),\n )\n const response = yield* client.execute(httpRequest).pipe(Effect.mapError(transportFailure))\n if (response.status >= 400) {\n const text = yield* response.text.pipe(Effect.orElseSucceed(() => \"\"))\n return yield* httpStatusError(response.status, text)\n }\n const json = yield* response.json.pipe(Effect.mapError(transportFailure))\n return yield* Schema.decodeUnknownEffect(WireResponse)(json).pipe(\n Effect.mapError(transportFailure),\n )\n })\n\n// ---------------------------------------------------------------------------\n// Service implementation\n// ---------------------------------------------------------------------------\n\nconst orderedEmbeddings = (\n data: ReadonlyArray<WirePayload>,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<ReadonlyArray<Embedding>, AiError.AiError> =>\n Effect.forEach(\n [...data].sort((a, b) => indexOf(a) - indexOf(b)),\n (item) => payloadToEmbedding(item, encoding),\n )\n\nconst embedImpl =\n (cfg: Config) =>\n (\n request: JinaEmbedRequest,\n ): Effect.Effect<AnyEmbedResponse, AiError.AiError, HttpClient.HttpClient> =>\n inputToItem(request.input).pipe(\n Effect.flatMap((item) =>\n postEmbed(\n cfg,\n buildBody({\n model: request.model,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(request.encoding !== undefined && { encoding: request.encoding }),\n items: [item],\n }),\n ),\n ),\n Effect.flatMap((decoded) => {\n const first = decoded.data[0]\n if (first === undefined) {\n return Effect.fail(transportFailure(\"Jina returned empty `data` array\"))\n }\n return payloadToEmbedding(first, request.encoding).pipe(\n Effect.map(\n (embedding): AnyEmbedResponse => ({ embedding, usage: usageOf(decoded.usage) }),\n ),\n )\n }),\n )\n\nconst embedManyImpl =\n (cfg: Config) =>\n (\n request: JinaEmbedManyRequest,\n ): Effect.Effect<AnyEmbedManyResponse, AiError.AiError, HttpClient.HttpClient> =>\n Effect.forEach(request.inputs, inputToItem).pipe(\n Effect.flatMap((items) =>\n postEmbed(\n cfg,\n buildBody({\n model: request.model,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(request.encoding !== undefined && { encoding: request.encoding }),\n items,\n }),\n ),\n ),\n Effect.flatMap((decoded) =>\n orderedEmbeddings(decoded.data, request.encoding).pipe(\n Effect.map(\n (embeddings): AnyEmbedManyResponse => ({\n embeddings,\n usage: usageOf(decoded.usage),\n }),\n ),\n ),\n ),\n )\n\n/**\n * Map cross-provider `query`/`document` to Jina's dotted-pair task\n * vocabulary. `undefined` stays `undefined`.\n */\nconst mapGenericTask: (t: \"query\" | \"document\" | undefined) => JinaTask | undefined = Match.type<\n \"query\" | \"document\" | undefined\n>().pipe(\n Match.when(\"query\", (): JinaTask => \"retrieval.query\"),\n Match.when(\"document\", (): JinaTask => \"retrieval.passage\"),\n Match.orElse((): JinaTask | undefined => undefined),\n)\n\n// ---------------------------------------------------------------------------\n// Constructors\n// ---------------------------------------------------------------------------\n\n/**\n * Build a `JinaEmbeddingService` value. For Layer-based setup, prefer\n * `layer`.\n */\nexport const make = (\n cfg: Config,\n): Effect.Effect<JinaEmbeddingService, never, HttpClient.HttpClient> =>\n Effect.map(HttpClient.HttpClient.asEffect(), (client) => ({\n embed: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) =>\n embedImpl(cfg)(request as JinaEmbedRequest).pipe(\n Effect.provideService(HttpClient.HttpClient, client),\n ) as Effect.Effect<EmbedResponse<E>, AiError.AiError>,\n embedMany: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) =>\n embedManyImpl(cfg)(request as JinaEmbedManyRequest).pipe(\n Effect.provideService(HttpClient.HttpClient, client),\n ) as Effect.Effect<EmbedManyResponse<E>, AiError.AiError>,\n }))\n\n/**\n * Layer that registers both the provider-specific `JinaEmbedding` tag\n * and the generic `EmbeddingModel` tag, sharing one underlying\n * implementation. The generic registration maps cross-provider\n * `query`/`document` to Jina's `retrieval.query` / `retrieval.passage`\n * task strings; other request fields pass through.\n */\nexport const layer = (\n cfg: Config,\n): Layer.Layer<JinaEmbedding | EmbeddingModel, never, HttpClient.HttpClient> => {\n const typed = Layer.effect(JinaEmbedding, make(cfg))\n const generic = Layer.effect(\n EmbeddingModel,\n Effect.map(\n make(cfg),\n (s): EmbeddingModelService => ({\n embed: <E extends EmbedEncoding | undefined = undefined>(\n req: Omit<CommonEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) => {\n const task = mapGenericTask(req.task)\n return s.embed({\n ...req,\n model: req.model as JinaEmbeddingModel,\n ...(task !== undefined && { task }),\n } as JinaEmbedRequest) as Effect.Effect<EmbedResponse<E>, AiError.AiError>\n },\n embedMany: <E extends EmbedEncoding | undefined = undefined>(\n req: Omit<CommonEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) => {\n const task = mapGenericTask(req.task)\n return s.embedMany({\n ...req,\n model: req.model as JinaEmbeddingModel,\n ...(task !== undefined && { task }),\n } as JinaEmbedManyRequest) as Effect.Effect<EmbedManyResponse<E>, AiError.AiError>\n },\n }),\n ),\n )\n return Layer.merge(typed, generic)\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAkGA,IAAa,gBAAb,cAAmC,QAAQ,SAA8C,CACvF,oDACD,CAAC;;;;;;AAwBF,MAAM,oBAAkD,MAAM,MAAmB,CAAC,KAChF,MAAM,IAAI,QAAQ,OAAiB,EAAE,OAAO,EAAE,KAAK,EAAE,EACrD,MAAM,IAAI,WAAW,OAAiB,EAAE,OAAO,EAAE,QAAQ,EAAE,EAC3D,MAAM,IAAI,UAAU,OAAiB,EAAE,OAAO,SAAS,aAAa,EAAE,MAAM,EAAE,EAAE,EAChF,MAAM,WACP;AAED,MAAM,oBAA0D,MAAM,MAAwB,CAAC,KAC7F,MAAM,KAAK,EAAE,MAAM,MAAM,QAAQ,GAAG,EAAE,YAAsB,EAAE,MAAM,EAAE,EACtE,MAAM,KAAK,EAAE,OAAO,MAAM,KAAK,GAAG,EAAE,YAAY,kBAAkB,MAAM,CAAC,EACzE,MAAM,WACP;AAED,MAAM,2BAA4C,IAAI,QAAQ,eAAe;CAC3E,UAAU;CACV,OAAO;CACP,KAAK;CACN,CAAC;;;;;;AAOF,MAAM,cACJ,MAAM,MAAkB,CAAC,KACvB,MAAM,KACJ,MAAM,SACL,MAAgD,OAAO,QAAQ,EAAE,MAAM,GAAG,CAAC,CAC7E,EACD,MAAM,KACJ,EAAE,MAAM,MAAM,QAAQ,GACrB,EAAE,WAAqD,OAAO,QAAQ,EAAE,MAAM,CAAC,CACjF,EACD,MAAM,KACJ,EAAE,OAAO,MAAM,KAAK,GACnB,EAAE,YACD,OAAO,QAAQ,kBAAkB,MAAM,CAAC,CAC3C,EACD,MAAM,KACJ,EAAE,SAAS,MAAM,KAAK,GACrB,EAAE,cACD,QAAQ,WAAW,IACf,OAAO,QAAQ,kBAAkB,QAAQ,GAAI,CAAC,GAC9C,OAAO,KAAK,yBAAyB,CAC5C,EACD,MAAM,WACP;;;;;;;AAmBH,MAAM,kBAAqE,MAAM,MAE9E,CAAC,KACF,MAAM,KAAK,gBAA2B,SAAS,EAC/C,MAAM,aAAwB,KAAA,EAAU,CACzC;AAED,MAAM,aAAa,YAMH;CACd,MAAM,QAAQ,gBAAgB,QAAQ,SAAS;AAC/C,QAAO;EACL,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;EACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;EAC1E,GAAI,UAAU,KAAA,KAAa,EAAE,gBAAgB,OAAO;EACpD,GAAI,QAAQ,aAAa,iBAAiB,EAAE,oBAAoB,MAAM;EACvE;;;;;;;;;;;;;;;AAoBH,MAAM,wBAAwB,OAAO,OAAO;CAC1C,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO,MAAM,OAAO,OAAO;CACvC,CAAC;AAGF,MAAM,yBAAyB,OAAO,OAAO;CAC3C,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO;CACnB,CAAC;AAGF,MAAM,oBAAoB,OAAO,OAAO;CACtC,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO,OAAO,OAAO,QAAQ,OAAO,OAAO;CACvD,CAAC;AAGF,MAAM,mBAAmB,OAAO,OAAO;CACrC,QAAQ,OAAO,QAAQ,aAAa;CACpC,OAAO,OAAO;CACd,YAAY,OAAO,MAAM,OAAO,MAAM,OAAO,OAAO,CAAC;CACrD,iBAAiB,OAAO,SAAS,OAAO,MAAM,OAAO,OAAO,CAAC;CAC9D,CAAC;AAMF,MAAM,cAAc,OAAO,MAAM;CAC/B;CACA;CACA;CACA;CACD,CAAC;AAGF,MAAM,iBAAiB,OAAO,GAAG,iBAAiB;AAClD,MAAM,kBAAkB,OAAO,GAAG,kBAAkB;AACpD,MAAM,uBAAuB,OAAO,GAAG,uBAAuB;AAE9D,MAAM,YAAY,OAAO,OAAO;CAC9B,cAAc,OAAO;CACrB,eAAe,OAAO,SAAS,OAAO,OAAO;CAC9C,CAAC;AAGF,MAAM,eAAe,OAAO,OAAO;CACjC,MAAM,OAAO,MAAM,YAAY;CAC/B,OAAO,OAAO;CACd,OAAO;CACR,CAAC;AAEF,MAAM,oBAAoB,UACxB,IAAI,QAAQ,YAAY;CAAE,UAAU;CAAQ,KAAK;CAAO,CAAC;AAE3D,MAAM,oBAAoB,UAAwB,QAChD,IAAI,QAAQ,eAAe;CACzB,UAAU;CACV,OAAO;CACP,KAAK,uBAAuB,SAAS,gCAAgC,IAAI;CAC1E,CAAC;;AAGJ,MAAM,yBACJ,KACA,aAEA,OAAO,MAAM,SAAS,aAAa,IAAI,EAAE;CACvC,YAAY,UAAU,OAAO,KAAK,iBAAiB,MAAM,CAAC;CAC1D,YAAY,UACV,MAAM,MAAM,SAAS,CAAC,KACpB,MAAM,KACJ,gBAEE,OAAO,QAAQ;EAAE,MAAM;EAAU,QAAQ;EAAO,CAAC,CACpD,EACD,MAAM,aAEF,OAAO,QAAQ;EACb,MAAM;EACN,QAAQ,IAAI,aAAa,MAAM,QAAQ,MAAM,YAAY,MAAM,aAAa,EAAE;EAC/E,CAAC,CACL,CACF;CACJ,CAAC;;;;;AAMJ,MAAM,sBACJ,SACA,aAC8C;AAC9C,KAAI,eAAe,QAAQ,CACzB,QAAO,aAAa,gBAChB,OAAO,QAAmB;EACxB,MAAM;EACN,SAAS,QAAQ,WAAW,KAAK,MAAM,aAAa,KAAK,EAAE,CAAC;EAC7D,CAAC,GACF,OAAO,KAAK,iBAAiB,YAAY,WAAW,cAAc,CAAC;AAEzE,KAAI,gBAAgB,QAAQ,CAC1B,QAAO,aAAa,WAChB,OAAO,QAAmB;EAAE,MAAM;EAAU,SAAS,QAAQ;EAAW,CAAC,GACzE,OAAO,KAAK,iBAAiB,YAAY,WAAW,SAAS,CAAC;AAGpE,KAAI,aAAa,iBAAiB,aAAa,SAC7C,QAAO,OAAO,KAAK,iBAAiB,UAAU,QAAQ,CAAC;AAEzD,QAAO,qBAAqB,QAAQ,GAChC,sBAAsB,QAAQ,WAAW,SAAS,GAClD,OAAO,QAAmB;EACxB,MAAM;EACN,QAAQ,aAAa,KAAK,QAAQ,UAAU;EAC7C,CAAC;;AAGR,MAAM,WAAW,MAA2B,EAAE;AAE9C,MAAM,WAAW,OAAyB,EAAE,aAAa,EAAE,cAAc;AAMzE,MAAM,mBAAmB,QAAgB,SAAkC;CACzE,MAAM,WAAW;CACjB,MAAM,MAAM;AACZ,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,YAAY;EAAE;EAAU;EAAK,CAAC;AACrE,KAAI,WAAW,OAAO,WAAW,IAAK,QAAO,IAAI,QAAQ,QAAQ;EAAE;EAAU;EAAK,CAAC;AACnF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAQ;EAAK,CAAC;AACrF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAc;EAAK,CAAC;AAC3F,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAW;EAAK,CAAC;AACxF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,sBAAsB;EAAE;EAAU;EAAK,CAAC;AAC/E,KAAI,UAAU,IAAK,QAAO,IAAI,QAAQ,YAAY;EAAE;EAAU;EAAQ;EAAK,CAAC;AAC5E,QAAO,IAAI,QAAQ,eAAe;EAAE;EAAU;EAAK,CAAC;;AAOtD,MAAM,WAAW,QAAwB,IAAI,WAAW;AAExD,MAAM,aACJ,KACA,SAEA,OAAO,IAAI,aAAa;CACtB,MAAM,SAAS,OAAO,WAAW;CACjC,MAAM,cAAc,kBAAkB,KAAK,GAAG,QAAQ,IAAI,CAAC,aAAa,CAAC,KACvE,kBAAkB,YAAY,IAAI,OAAO,EACzC,kBAAkB,eAAe,KAAK,CACvC;CACD,MAAM,WAAW,OAAO,OAAO,QAAQ,YAAY,CAAC,KAAK,OAAO,SAAS,iBAAiB,CAAC;AAC3F,KAAI,SAAS,UAAU,KAAK;EAC1B,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,oBAAoB,GAAG,CAAC;AACtE,SAAO,OAAO,gBAAgB,SAAS,QAAQ,KAAK;;CAEtD,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,SAAS,iBAAiB,CAAC;AACzE,QAAO,OAAO,OAAO,oBAAoB,aAAa,CAAC,KAAK,CAAC,KAC3D,OAAO,SAAS,iBAAiB,CAClC;EACD;AAMJ,MAAM,qBACJ,MACA,aAEA,OAAO,QACL,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,MAAM,QAAQ,EAAE,GAAG,QAAQ,EAAE,CAAC,GAChD,SAAS,mBAAmB,MAAM,SAAS,CAC7C;AAEH,MAAM,aACH,SAEC,YAEA,YAAY,QAAQ,MAAM,CAAC,KACzB,OAAO,SAAS,SACd,UACE,KACA,UAAU;CACR,OAAO,QAAQ;CACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;CACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;CAC1E,GAAI,QAAQ,aAAa,KAAA,KAAa,EAAE,UAAU,QAAQ,UAAU;CACpE,OAAO,CAAC,KAAK;CACd,CAAC,CACH,CACF,EACD,OAAO,SAAS,YAAY;CAC1B,MAAM,QAAQ,QAAQ,KAAK;AAC3B,KAAI,UAAU,KAAA,EACZ,QAAO,OAAO,KAAK,iBAAiB,mCAAmC,CAAC;AAE1E,QAAO,mBAAmB,OAAO,QAAQ,SAAS,CAAC,KACjD,OAAO,KACJ,eAAiC;EAAE;EAAW,OAAO,QAAQ,QAAQ,MAAM;EAAE,EAC/E,CACF;EACD,CACH;AAEL,MAAM,iBACH,SAEC,YAEA,OAAO,QAAQ,QAAQ,QAAQ,YAAY,CAAC,KAC1C,OAAO,SAAS,UACd,UACE,KACA,UAAU;CACR,OAAO,QAAQ;CACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;CACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;CAC1E,GAAI,QAAQ,aAAa,KAAA,KAAa,EAAE,UAAU,QAAQ,UAAU;CACpE;CACD,CAAC,CACH,CACF,EACD,OAAO,SAAS,YACd,kBAAkB,QAAQ,MAAM,QAAQ,SAAS,CAAC,KAChD,OAAO,KACJ,gBAAsC;CACrC;CACA,OAAO,QAAQ,QAAQ,MAAM;CAC9B,EACF,CACF,CACF,CACF;;;;;AAML,MAAM,iBAAgF,MAAM,MAEzF,CAAC,KACF,MAAM,KAAK,eAAyB,kBAAkB,EACtD,MAAM,KAAK,kBAA4B,oBAAoB,EAC3D,MAAM,aAAmC,KAAA,EAAU,CACpD;;;;;AAUD,MAAa,QACX,QAEA,OAAO,IAAI,WAAW,WAAW,UAAU,GAAG,YAAY;CACxD,QACE,YAEA,UAAU,IAAI,CAAC,QAA4B,CAAC,KAC1C,OAAO,eAAe,WAAW,YAAY,OAAO,CACrD;CACH,YACE,YAEA,cAAc,IAAI,CAAC,QAAgC,CAAC,KAClD,OAAO,eAAe,WAAW,YAAY,OAAO,CACrD;CACJ,EAAE;;;;;;;;AASL,MAAa,SACX,QAC8E;CAC9E,MAAM,QAAQ,MAAM,OAAO,eAAe,KAAK,IAAI,CAAC;CACpD,MAAM,UAAU,MAAM,OACpB,gBACA,OAAO,IACL,KAAK,IAAI,GACR,OAA8B;EAC7B,QACE,QACG;GACH,MAAM,OAAO,eAAe,IAAI,KAAK;AACrC,UAAO,EAAE,MAAM;IACb,GAAG;IACH,OAAO,IAAI;IACX,GAAI,SAAS,KAAA,KAAa,EAAE,MAAM;IACnC,CAAqB;;EAExB,YACE,QACG;GACH,MAAM,OAAO,eAAe,IAAI,KAAK;AACrC,UAAO,EAAE,UAAU;IACjB,GAAG;IACH,OAAO,IAAI;IACX,GAAI,SAAS,KAAA,KAAa,EAAE,MAAM;IACnC,CAAyB;;EAE7B,EACF,CACF;AACD,QAAO,MAAM,MAAM,OAAO,QAAQ"}
|
|
1
|
+
{"version":3,"file":"JinaEmbedding.mjs","names":[],"sources":["../src/JinaEmbedding.ts"],"sourcesContent":["import { Context, Effect, Encoding, Layer, Match, Redacted, Result, Schema } from \"effect\"\nimport { HttpClient, HttpClientRequest } from \"effect/unstable/http\"\nimport * as AiError from \"@effect-uai/core/AiError\"\nimport type { Embedding, EmbedContentPart, EmbedInput, Usage } from \"@effect-uai/core/Embedding\"\n\n// Internal open-union response shape — impls return this; the typed\n// `make()` wrapper casts to the narrowed `EmbedResponse<E>`.\ntype AnyEmbedResponse = { readonly embedding: Embedding; readonly usage: Usage }\ntype AnyEmbedManyResponse = { readonly embeddings: ReadonlyArray<Embedding>; readonly usage: Usage }\nimport {\n assertEncoding,\n type CommonEmbedManyRequest,\n type CommonEmbedRequest,\n type EmbedEncoding,\n EmbeddingModel,\n type EmbeddingModelService,\n type EmbedManyResponse,\n type EmbedResponse,\n} from \"@effect-uai/core/EmbeddingModel\"\nimport type { ImageSource } from \"@effect-uai/core/Image\"\nimport type { JinaEmbeddingModel } from \"./models.js\"\n\n// ---------------------------------------------------------------------------\n// Public types\n// ---------------------------------------------------------------------------\n\n/**\n * Jina's task vocabulary. v4 and v3/v5 share the dotted-pair form:\n * - `retrieval.query` / `retrieval.passage` — asymmetric retrieval.\n * - `text-matching` — symmetric similarity between two texts.\n * - `code.query` / `code.passage` — code retrieval / matching (v4-only).\n * - `classification` / `separation` — v3 / v5 only.\n *\n * The `(string & {})` tail accepts any string so newly-released task\n * types work without an SDK update.\n */\nexport type JinaTask =\n | \"retrieval.query\"\n | \"retrieval.passage\"\n | \"text-matching\"\n | \"code.query\"\n | \"code.passage\"\n | \"classification\"\n | \"separation\"\n // eslint-disable-next-line @typescript-eslint/ban-types\n | (string & {})\n\n/**\n * Jina's supported `encoding` values. Provider-narrowed from the cross-\n * provider `EmbedEncoding` union.\n *\n * Compatibility with models is checked at the response level rather than\n * pre-flight: if you ask for `multivector` against a model that doesn't\n * produce one, Jina's API rejects the request with a clear error; if\n * you ask for `sparse` and the chosen model returns dense, our decoder\n * fails with a typed mismatch. No hardcoded model-encoding table to\n * maintain.\n *\n * Wire mapping:\n * - `multivector` → request sets `return_multivector: true`\n * (`jina-embeddings-v4` + ColBERT models).\n * - `sparse` → no flag; the model itself (`elser-v2`) determines the\n * output shape.\n * - `binary` → request sets `embedding_type: \"binary\"`.\n * - `float32` → no flag; default JSON `number[]`.\n */\nexport type JinaEncoding = \"float32\" | \"binary\" | \"sparse\" | \"multivector\"\n\nexport type JinaEmbedRequest = Omit<CommonEmbedRequest, \"model\" | \"task\" | \"encoding\"> & {\n /** Narrows `CommonEmbedRequest.model` to the typed Jina union. */\n readonly model: JinaEmbeddingModel\n /**\n * Widens the cross-provider `\"query\" | \"document\"` to Jina's task enum.\n * Required by v3+ for retrieval-quality results. v4 unifies query and\n * document under `retrieval`.\n */\n readonly task?: JinaTask\n /** Narrows `CommonEmbedRequest.encoding` to Jina's supported set. */\n readonly encoding?: JinaEncoding\n}\n\nexport type JinaEmbedManyRequest = Omit<JinaEmbedRequest, \"input\"> & {\n readonly inputs: ReadonlyArray<EmbedInput>\n}\n\nexport type JinaEmbeddingService = {\n readonly embed: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) => Effect.Effect<EmbedResponse<E>, AiError.AiError>\n readonly embedMany: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) => Effect.Effect<EmbedManyResponse<E>, AiError.AiError>\n}\n\n/**\n * Provider-typed service tag. Yield this for Jina-typed embedding\n * requests; yield the generic `EmbeddingModel` tag for provider-portable\n * code. Both are registered by `layer`.\n */\nexport class JinaEmbedding extends Context.Service<JinaEmbedding, JinaEmbeddingService>()(\n \"@betalyra/effect-uai/providers/jina/JinaEmbedding\",\n) {}\n\nexport type Config = {\n readonly apiKey: Redacted.Redacted\n readonly baseUrl?: string\n}\n\n// ---------------------------------------------------------------------------\n// Codec - request body\n// ---------------------------------------------------------------------------\n\ntype WireTextItem = {\n readonly text: string\n}\ntype WireImageItem = {\n readonly image: string\n}\ntype WireItem = WireTextItem | WireImageItem\n\n/**\n * Jina v4 accepts `{ image: <url> }` or `{ image: <base64-string> }`.\n * The server detects which form by looking at the value. We pass URLs\n * through directly and base64-encode raw bytes.\n */\nconst imageSourceToItem: (s: ImageSource) => WireItem = Match.type<ImageSource>().pipe(\n Match.tag(\"url\", (s): WireItem => ({ image: s.url })),\n Match.tag(\"base64\", (s): WireItem => ({ image: s.base64 })),\n Match.tag(\"bytes\", (s): WireItem => ({ image: Encoding.encodeBase64(s.bytes) })),\n Match.exhaustive,\n)\n\nconst contentPartToItem: (part: EmbedContentPart) => WireItem = Match.type<EmbedContentPart>().pipe(\n Match.when({ text: Match.string }, ({ text }): WireItem => ({ text })),\n Match.when({ image: Match.any }, ({ image }) => imageSourceToItem(image)),\n Match.exhaustive,\n)\n\n// Jina's flat `input[]` can't fuse a multi-part `content[]` into one vector.\n// Bucket 1: silently splitting or dropping parts would change what the vector\n// represents (and the vector count). Reject as `Unsupported` (capability gap),\n// not `InvalidRequest` (wire-shape) - Jina does carry images, just not fused.\nconst multiPartContentRejected: AiError.AiError = new AiError.Unsupported({\n provider: \"jina\",\n capability: \"multiPartInput\",\n reason:\n \"Jina treats each input[] entry as one item; multi-part `content[]` can't be fused into one vector. Split into separate `inputs[]` entries.\",\n})\n\n/**\n * Project an `EmbedInput` to one wire item. A single-part `content[]` is\n * unwrapped; multi-part `content[]` is rejected because Jina's flat\n * `input[]` would lose the grouping.\n */\nconst inputToItem: (input: EmbedInput) => Effect.Effect<WireItem, AiError.AiError> =\n Match.type<EmbedInput>().pipe(\n Match.when(\n Match.string,\n (s): Effect.Effect<WireItem, AiError.AiError> => Effect.succeed({ text: s }),\n ),\n Match.when(\n { text: Match.string },\n ({ text }): Effect.Effect<WireItem, AiError.AiError> => Effect.succeed({ text }),\n ),\n Match.when(\n { image: Match.any },\n ({ image }): Effect.Effect<WireItem, AiError.AiError> =>\n Effect.succeed(imageSourceToItem(image)),\n ),\n Match.when(\n { content: Match.any },\n ({ content }): Effect.Effect<WireItem, AiError.AiError> =>\n content.length === 1\n ? Effect.succeed(contentPartToItem(content[0]!))\n : Effect.fail(multiPartContentRejected),\n ),\n Match.exhaustive,\n )\n\ntype WireBody = {\n readonly model: string\n readonly input: ReadonlyArray<WireItem>\n readonly task?: string\n readonly dimensions?: number\n readonly embedding_type?: \"float\" | \"base64\" | \"binary\" | \"ubinary\"\n readonly return_multivector?: boolean\n}\n\ntype WireQuant = WireBody[\"embedding_type\"]\n\n/**\n * Map our `JinaEncoding` to Jina's `embedding_type` wire field. Only\n * `binary` actually changes the wire transport; the others either omit\n * the field (`float32`) or ignore it (`sparse` / `multivector` use\n * separate signalling — model choice and `return_multivector`).\n */\nconst encodingToQuant: (encoding: JinaEncoding | undefined) => WireQuant = Match.type<\n JinaEncoding | undefined\n>().pipe(\n Match.when(\"binary\", (): WireQuant => \"binary\"),\n Match.orElse((): WireQuant => undefined),\n)\n\nconst buildBody = (request: {\n readonly model: JinaEmbeddingModel\n readonly task?: JinaTask\n readonly dimensions?: number\n readonly encoding?: JinaEncoding\n readonly items: ReadonlyArray<WireItem>\n}): WireBody => {\n const quant = encodingToQuant(request.encoding)\n return {\n model: request.model,\n input: request.items,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(quant !== undefined && { embedding_type: quant }),\n ...(request.encoding === \"multivector\" && { return_multivector: true }),\n }\n}\n\n// ---------------------------------------------------------------------------\n// Codec - response\n// ---------------------------------------------------------------------------\n\n/**\n * Wire response payloads. Split into four discriminated schemas so the\n * decoder can narrow precisely via `Schema.is`:\n *\n * - **Dense float** (`object: \"embedding\"`, `embedding: number[]`) —\n * default JSON form.\n * - **Dense base64** (`object: \"embedding\"`, `embedding: string`) — set\n * when `embedding_type` was `base64` or `binary`.\n * - **Sparse** (`object: \"embedding\"`, `embedding: Record<string,\n * number>`) — token-keyed weights from an ELSER-style model.\n * - **Multivector** (`object: \"embeddings\"`, plural) — token-level\n * vectors when `return_multivector: true` or a ColBERT model is used.\n */\nconst WireDenseFloatPayload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.Array(Schema.Number),\n})\ntype WireDenseFloatPayload = typeof WireDenseFloatPayload.Type\n\nconst WireDenseBase64Payload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.String,\n})\ntype WireDenseBase64Payload = typeof WireDenseBase64Payload.Type\n\nconst WireSparsePayload = Schema.Struct({\n object: Schema.Literal(\"embedding\"),\n index: Schema.Number,\n embedding: Schema.Record(Schema.String, Schema.Number),\n})\ntype WireSparsePayload = typeof WireSparsePayload.Type\n\nconst WireMultiPayload = Schema.Struct({\n object: Schema.Literal(\"embeddings\"),\n index: Schema.Number,\n embeddings: Schema.Array(Schema.Array(Schema.Number)),\n tokenized_input: Schema.optional(Schema.Array(Schema.String)),\n})\ntype WireMultiPayload = typeof WireMultiPayload.Type\n\n// Order matters: more specific schemas first. Dense float (array) and\n// multivector (different `object`) are unambiguous; sparse vs base64\n// disambiguate by the `embedding` field type.\nconst WirePayload = Schema.Union([\n WireDenseFloatPayload,\n WireSparsePayload,\n WireDenseBase64Payload,\n WireMultiPayload,\n])\ntype WirePayload = typeof WirePayload.Type\n\nconst isMultiPayload = Schema.is(WireMultiPayload)\nconst isSparsePayload = Schema.is(WireSparsePayload)\nconst isDenseBase64Payload = Schema.is(WireDenseBase64Payload)\n\nconst WireUsage = Schema.Struct({\n total_tokens: Schema.Number,\n prompt_tokens: Schema.optional(Schema.Number),\n})\ntype WireUsage = typeof WireUsage.Type\n\nconst WireResponse = Schema.Struct({\n data: Schema.Array(WirePayload),\n model: Schema.String,\n usage: WireUsage,\n})\n\nconst transportFailure = (cause: unknown): AiError.AiError =>\n new AiError.Unavailable({ provider: \"jina\", raw: cause })\n\nconst encodingMismatch = (expected: JinaEncoding, got: string): AiError.AiError =>\n new AiError.InvalidRequest({\n provider: \"jina\",\n param: \"encoding\",\n raw: `requested encoding=\"${expected}\" but the response contains a ${got} embedding - confirm the chosen model supports the requested encoding`,\n })\n\n/** Decode a base64 wire string to the right dense `Embedding` arm. */\nconst decodeBase64Embedding = (\n b64: string,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<Embedding, AiError.AiError> =>\n Result.match(Encoding.decodeBase64(b64), {\n onFailure: (cause) => Effect.fail(transportFailure(cause)),\n onSuccess: (bytes) =>\n Match.value(encoding).pipe(\n Match.when(\n \"binary\",\n (): Effect.Effect<Embedding, AiError.AiError> =>\n Effect.succeed({ _tag: \"binary\", vector: bytes }),\n ),\n Match.orElse(\n (): Effect.Effect<Embedding, AiError.AiError> =>\n Effect.succeed({\n _tag: \"float32\",\n vector: new Float32Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 4),\n }),\n ),\n ),\n })\n\n/**\n * Project a wire payload to our `Embedding` discriminated union and\n * verify the requested encoding matches what came back.\n */\nconst payloadToEmbedding = (\n payload: WirePayload,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<Embedding, AiError.AiError> => {\n if (isMultiPayload(payload)) {\n return encoding === \"multivector\"\n ? Effect.succeed<Embedding>({\n _tag: \"multivector\",\n vectors: payload.embeddings.map((v) => Float32Array.from(v)),\n })\n : Effect.fail(encodingMismatch(encoding ?? \"float32\", \"multivector\"))\n }\n if (isSparsePayload(payload)) {\n return encoding === \"sparse\"\n ? Effect.succeed<Embedding>({ _tag: \"sparse\", weights: payload.embedding })\n : Effect.fail(encodingMismatch(encoding ?? \"float32\", \"sparse\"))\n }\n // Dense (float array or base64 string).\n if (encoding === \"multivector\" || encoding === \"sparse\") {\n return Effect.fail(encodingMismatch(encoding, \"dense\"))\n }\n return isDenseBase64Payload(payload)\n ? decodeBase64Embedding(payload.embedding, encoding)\n : Effect.succeed<Embedding>({\n _tag: \"float32\",\n vector: Float32Array.from(payload.embedding),\n })\n}\n\nconst indexOf = (p: WirePayload): number => p.index\n\nconst usageOf = (u: WireUsage): Usage => ({ inputTokens: u.total_tokens })\n\n// ---------------------------------------------------------------------------\n// HTTP errors\n// ---------------------------------------------------------------------------\n\nconst httpStatusError = (status: number, body: string): AiError.AiError => {\n const provider = \"jina\"\n const raw = body\n if (status === 429) return new AiError.RateLimited({ provider, raw })\n if (status === 408 || status === 504) return new AiError.Timeout({ provider, raw })\n if (status === 401) return new AiError.AuthFailed({ provider, subtype: \"auth\", raw })\n if (status === 403) return new AiError.AuthFailed({ provider, subtype: \"permission\", raw })\n if (status === 402) return new AiError.AuthFailed({ provider, subtype: \"billing\", raw })\n if (status === 413) return new AiError.ContextLengthExceeded({ provider, raw })\n if (status >= 500) return new AiError.Unavailable({ provider, status, raw })\n return new AiError.InvalidRequest({ provider, raw })\n}\n\n// ---------------------------------------------------------------------------\n// HTTP plumbing\n// ---------------------------------------------------------------------------\n\nconst baseUrl = (cfg: Config): string => cfg.baseUrl ?? \"https://api.jina.ai/v1\"\n\nconst postEmbed = (\n cfg: Config,\n body: WireBody,\n): Effect.Effect<typeof WireResponse.Type, AiError.AiError, HttpClient.HttpClient> =>\n Effect.gen(function* () {\n const client = yield* HttpClient.HttpClient\n const httpRequest = HttpClientRequest.post(`${baseUrl(cfg)}/embeddings`).pipe(\n HttpClientRequest.bearerToken(cfg.apiKey),\n HttpClientRequest.bodyJsonUnsafe(body),\n )\n const response = yield* client.execute(httpRequest).pipe(Effect.mapError(transportFailure))\n if (response.status >= 400) {\n const text = yield* response.text.pipe(Effect.orElseSucceed(() => \"\"))\n return yield* httpStatusError(response.status, text)\n }\n const json = yield* response.json.pipe(Effect.mapError(transportFailure))\n return yield* Schema.decodeUnknownEffect(WireResponse)(json).pipe(\n Effect.mapError(transportFailure),\n )\n })\n\n// ---------------------------------------------------------------------------\n// Service implementation\n// ---------------------------------------------------------------------------\n\nconst orderedEmbeddings = (\n data: ReadonlyArray<WirePayload>,\n encoding: JinaEncoding | undefined,\n): Effect.Effect<ReadonlyArray<Embedding>, AiError.AiError> =>\n Effect.forEach(\n [...data].sort((a, b) => indexOf(a) - indexOf(b)),\n (item) => payloadToEmbedding(item, encoding),\n )\n\nconst embedImpl =\n (cfg: Config) =>\n (\n request: JinaEmbedRequest,\n ): Effect.Effect<AnyEmbedResponse, AiError.AiError, HttpClient.HttpClient> =>\n inputToItem(request.input).pipe(\n Effect.flatMap((item) =>\n postEmbed(\n cfg,\n buildBody({\n model: request.model,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(request.encoding !== undefined && { encoding: request.encoding }),\n items: [item],\n }),\n ),\n ),\n Effect.flatMap((decoded) => {\n const first = decoded.data[0]\n if (first === undefined) {\n return Effect.fail(transportFailure(\"Jina returned empty `data` array\"))\n }\n return payloadToEmbedding(first, request.encoding).pipe(\n Effect.map(\n (embedding): AnyEmbedResponse => ({ embedding, usage: usageOf(decoded.usage) }),\n ),\n )\n }),\n )\n\nconst embedManyImpl =\n (cfg: Config) =>\n (\n request: JinaEmbedManyRequest,\n ): Effect.Effect<AnyEmbedManyResponse, AiError.AiError, HttpClient.HttpClient> =>\n Effect.forEach(request.inputs, inputToItem).pipe(\n Effect.flatMap((items) =>\n postEmbed(\n cfg,\n buildBody({\n model: request.model,\n ...(request.task !== undefined && { task: request.task }),\n ...(request.dimensions !== undefined && { dimensions: request.dimensions }),\n ...(request.encoding !== undefined && { encoding: request.encoding }),\n items,\n }),\n ),\n ),\n Effect.flatMap((decoded) =>\n orderedEmbeddings(decoded.data, request.encoding).pipe(\n Effect.map(\n (embeddings): AnyEmbedManyResponse => ({\n embeddings,\n usage: usageOf(decoded.usage),\n }),\n ),\n ),\n ),\n )\n\n/**\n * Map cross-provider `query`/`document` to Jina's dotted-pair task\n * vocabulary. `undefined` stays `undefined`.\n */\nconst mapGenericTask: (t: \"query\" | \"document\" | undefined) => JinaTask | undefined = Match.type<\n \"query\" | \"document\" | undefined\n>().pipe(\n Match.when(\"query\", (): JinaTask => \"retrieval.query\"),\n Match.when(\"document\", (): JinaTask => \"retrieval.passage\"),\n Match.orElse((): JinaTask | undefined => undefined),\n)\n\n// ---------------------------------------------------------------------------\n// Constructors\n// ---------------------------------------------------------------------------\n\n/**\n * Build a `JinaEmbeddingService` value. For Layer-based setup, prefer\n * `layer`.\n */\nexport const make = (\n cfg: Config,\n): Effect.Effect<JinaEmbeddingService, never, HttpClient.HttpClient> =>\n Effect.map(HttpClient.HttpClient.asEffect(), (client) => ({\n embed: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) =>\n embedImpl(cfg)(request as JinaEmbedRequest).pipe(\n Effect.provideService(HttpClient.HttpClient, client),\n ) as Effect.Effect<EmbedResponse<E>, AiError.AiError>,\n embedMany: <E extends JinaEncoding | undefined = undefined>(\n request: Omit<JinaEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) =>\n embedManyImpl(cfg)(request as JinaEmbedManyRequest).pipe(\n Effect.provideService(HttpClient.HttpClient, client),\n ) as Effect.Effect<EmbedManyResponse<E>, AiError.AiError>,\n }))\n\n/**\n * Layer that registers both the provider-specific `JinaEmbedding` tag\n * and the generic `EmbeddingModel` tag, sharing one underlying\n * implementation. The generic registration maps cross-provider\n * `query`/`document` to Jina's `retrieval.query` / `retrieval.passage`\n * task strings; other request fields pass through.\n */\nexport const layer = (\n cfg: Config,\n): Layer.Layer<JinaEmbedding | EmbeddingModel, never, HttpClient.HttpClient> => {\n const typed = Layer.effect(JinaEmbedding, make(cfg))\n const generic = Layer.effect(\n EmbeddingModel,\n Effect.map(\n make(cfg),\n (s): EmbeddingModelService => ({\n // Jina emits dense float32 / binary (binary is bit-quantized, packed\n // into int8 bytes on the wire - NOT scalar int8 per dimension; sparse /\n // multivector live on the provider-typed JinaEncoding, off the\n // cross-provider set). Reject a scalar `int8` encoding (bucket 1) so it\n // isn't returned as a mislabeled vector, matching OpenAI / Gemini.\n embed: <E extends EmbedEncoding | undefined = undefined>(\n req: Omit<CommonEmbedRequest, \"encoding\"> & { readonly encoding?: E },\n ) => {\n const task = mapGenericTask(req.task)\n return assertEncoding(req.encoding, [\"float32\", \"binary\"], \"jina\").pipe(\n Effect.andThen(\n s.embed({\n ...req,\n model: req.model as JinaEmbeddingModel,\n ...(task !== undefined && { task }),\n } as JinaEmbedRequest),\n ),\n ) as Effect.Effect<EmbedResponse<E>, AiError.AiError>\n },\n embedMany: <E extends EmbedEncoding | undefined = undefined>(\n req: Omit<CommonEmbedManyRequest, \"encoding\"> & { readonly encoding?: E },\n ) => {\n const task = mapGenericTask(req.task)\n return assertEncoding(req.encoding, [\"float32\", \"binary\"], \"jina\").pipe(\n Effect.andThen(\n s.embedMany({\n ...req,\n model: req.model as JinaEmbeddingModel,\n ...(task !== undefined && { task }),\n } as JinaEmbedManyRequest),\n ),\n ) as Effect.Effect<EmbedManyResponse<E>, AiError.AiError>\n },\n }),\n ),\n )\n return Layer.merge(typed, generic)\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAmGA,IAAa,gBAAb,cAAmC,QAAQ,SAA8C,CACvF,oDACD,CAAC;;;;;;AAwBF,MAAM,oBAAkD,MAAM,MAAmB,CAAC,KAChF,MAAM,IAAI,QAAQ,OAAiB,EAAE,OAAO,EAAE,KAAK,EAAE,EACrD,MAAM,IAAI,WAAW,OAAiB,EAAE,OAAO,EAAE,QAAQ,EAAE,EAC3D,MAAM,IAAI,UAAU,OAAiB,EAAE,OAAO,SAAS,aAAa,EAAE,MAAM,EAAE,EAAE,EAChF,MAAM,WACP;AAED,MAAM,oBAA0D,MAAM,MAAwB,CAAC,KAC7F,MAAM,KAAK,EAAE,MAAM,MAAM,QAAQ,GAAG,EAAE,YAAsB,EAAE,MAAM,EAAE,EACtE,MAAM,KAAK,EAAE,OAAO,MAAM,KAAK,GAAG,EAAE,YAAY,kBAAkB,MAAM,CAAC,EACzE,MAAM,WACP;AAMD,MAAM,2BAA4C,IAAI,QAAQ,YAAY;CACxE,UAAU;CACV,YAAY;CACZ,QACE;CACH,CAAC;;;;;;AAOF,MAAM,cACJ,MAAM,MAAkB,CAAC,KACvB,MAAM,KACJ,MAAM,SACL,MAAgD,OAAO,QAAQ,EAAE,MAAM,GAAG,CAAC,CAC7E,EACD,MAAM,KACJ,EAAE,MAAM,MAAM,QAAQ,GACrB,EAAE,WAAqD,OAAO,QAAQ,EAAE,MAAM,CAAC,CACjF,EACD,MAAM,KACJ,EAAE,OAAO,MAAM,KAAK,GACnB,EAAE,YACD,OAAO,QAAQ,kBAAkB,MAAM,CAAC,CAC3C,EACD,MAAM,KACJ,EAAE,SAAS,MAAM,KAAK,GACrB,EAAE,cACD,QAAQ,WAAW,IACf,OAAO,QAAQ,kBAAkB,QAAQ,GAAI,CAAC,GAC9C,OAAO,KAAK,yBAAyB,CAC5C,EACD,MAAM,WACP;;;;;;;AAmBH,MAAM,kBAAqE,MAAM,MAE9E,CAAC,KACF,MAAM,KAAK,gBAA2B,SAAS,EAC/C,MAAM,aAAwB,KAAA,EAAU,CACzC;AAED,MAAM,aAAa,YAMH;CACd,MAAM,QAAQ,gBAAgB,QAAQ,SAAS;AAC/C,QAAO;EACL,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;EACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;EAC1E,GAAI,UAAU,KAAA,KAAa,EAAE,gBAAgB,OAAO;EACpD,GAAI,QAAQ,aAAa,iBAAiB,EAAE,oBAAoB,MAAM;EACvE;;;;;;;;;;;;;;;AAoBH,MAAM,wBAAwB,OAAO,OAAO;CAC1C,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO,MAAM,OAAO,OAAO;CACvC,CAAC;AAGF,MAAM,yBAAyB,OAAO,OAAO;CAC3C,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO;CACnB,CAAC;AAGF,MAAM,oBAAoB,OAAO,OAAO;CACtC,QAAQ,OAAO,QAAQ,YAAY;CACnC,OAAO,OAAO;CACd,WAAW,OAAO,OAAO,OAAO,QAAQ,OAAO,OAAO;CACvD,CAAC;AAGF,MAAM,mBAAmB,OAAO,OAAO;CACrC,QAAQ,OAAO,QAAQ,aAAa;CACpC,OAAO,OAAO;CACd,YAAY,OAAO,MAAM,OAAO,MAAM,OAAO,OAAO,CAAC;CACrD,iBAAiB,OAAO,SAAS,OAAO,MAAM,OAAO,OAAO,CAAC;CAC9D,CAAC;AAMF,MAAM,cAAc,OAAO,MAAM;CAC/B;CACA;CACA;CACA;CACD,CAAC;AAGF,MAAM,iBAAiB,OAAO,GAAG,iBAAiB;AAClD,MAAM,kBAAkB,OAAO,GAAG,kBAAkB;AACpD,MAAM,uBAAuB,OAAO,GAAG,uBAAuB;AAE9D,MAAM,YAAY,OAAO,OAAO;CAC9B,cAAc,OAAO;CACrB,eAAe,OAAO,SAAS,OAAO,OAAO;CAC9C,CAAC;AAGF,MAAM,eAAe,OAAO,OAAO;CACjC,MAAM,OAAO,MAAM,YAAY;CAC/B,OAAO,OAAO;CACd,OAAO;CACR,CAAC;AAEF,MAAM,oBAAoB,UACxB,IAAI,QAAQ,YAAY;CAAE,UAAU;CAAQ,KAAK;CAAO,CAAC;AAE3D,MAAM,oBAAoB,UAAwB,QAChD,IAAI,QAAQ,eAAe;CACzB,UAAU;CACV,OAAO;CACP,KAAK,uBAAuB,SAAS,gCAAgC,IAAI;CAC1E,CAAC;;AAGJ,MAAM,yBACJ,KACA,aAEA,OAAO,MAAM,SAAS,aAAa,IAAI,EAAE;CACvC,YAAY,UAAU,OAAO,KAAK,iBAAiB,MAAM,CAAC;CAC1D,YAAY,UACV,MAAM,MAAM,SAAS,CAAC,KACpB,MAAM,KACJ,gBAEE,OAAO,QAAQ;EAAE,MAAM;EAAU,QAAQ;EAAO,CAAC,CACpD,EACD,MAAM,aAEF,OAAO,QAAQ;EACb,MAAM;EACN,QAAQ,IAAI,aAAa,MAAM,QAAQ,MAAM,YAAY,MAAM,aAAa,EAAE;EAC/E,CAAC,CACL,CACF;CACJ,CAAC;;;;;AAMJ,MAAM,sBACJ,SACA,aAC8C;AAC9C,KAAI,eAAe,QAAQ,CACzB,QAAO,aAAa,gBAChB,OAAO,QAAmB;EACxB,MAAM;EACN,SAAS,QAAQ,WAAW,KAAK,MAAM,aAAa,KAAK,EAAE,CAAC;EAC7D,CAAC,GACF,OAAO,KAAK,iBAAiB,YAAY,WAAW,cAAc,CAAC;AAEzE,KAAI,gBAAgB,QAAQ,CAC1B,QAAO,aAAa,WAChB,OAAO,QAAmB;EAAE,MAAM;EAAU,SAAS,QAAQ;EAAW,CAAC,GACzE,OAAO,KAAK,iBAAiB,YAAY,WAAW,SAAS,CAAC;AAGpE,KAAI,aAAa,iBAAiB,aAAa,SAC7C,QAAO,OAAO,KAAK,iBAAiB,UAAU,QAAQ,CAAC;AAEzD,QAAO,qBAAqB,QAAQ,GAChC,sBAAsB,QAAQ,WAAW,SAAS,GAClD,OAAO,QAAmB;EACxB,MAAM;EACN,QAAQ,aAAa,KAAK,QAAQ,UAAU;EAC7C,CAAC;;AAGR,MAAM,WAAW,MAA2B,EAAE;AAE9C,MAAM,WAAW,OAAyB,EAAE,aAAa,EAAE,cAAc;AAMzE,MAAM,mBAAmB,QAAgB,SAAkC;CACzE,MAAM,WAAW;CACjB,MAAM,MAAM;AACZ,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,YAAY;EAAE;EAAU;EAAK,CAAC;AACrE,KAAI,WAAW,OAAO,WAAW,IAAK,QAAO,IAAI,QAAQ,QAAQ;EAAE;EAAU;EAAK,CAAC;AACnF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAQ;EAAK,CAAC;AACrF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAc;EAAK,CAAC;AAC3F,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,WAAW;EAAE;EAAU,SAAS;EAAW;EAAK,CAAC;AACxF,KAAI,WAAW,IAAK,QAAO,IAAI,QAAQ,sBAAsB;EAAE;EAAU;EAAK,CAAC;AAC/E,KAAI,UAAU,IAAK,QAAO,IAAI,QAAQ,YAAY;EAAE;EAAU;EAAQ;EAAK,CAAC;AAC5E,QAAO,IAAI,QAAQ,eAAe;EAAE;EAAU;EAAK,CAAC;;AAOtD,MAAM,WAAW,QAAwB,IAAI,WAAW;AAExD,MAAM,aACJ,KACA,SAEA,OAAO,IAAI,aAAa;CACtB,MAAM,SAAS,OAAO,WAAW;CACjC,MAAM,cAAc,kBAAkB,KAAK,GAAG,QAAQ,IAAI,CAAC,aAAa,CAAC,KACvE,kBAAkB,YAAY,IAAI,OAAO,EACzC,kBAAkB,eAAe,KAAK,CACvC;CACD,MAAM,WAAW,OAAO,OAAO,QAAQ,YAAY,CAAC,KAAK,OAAO,SAAS,iBAAiB,CAAC;AAC3F,KAAI,SAAS,UAAU,KAAK;EAC1B,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,oBAAoB,GAAG,CAAC;AACtE,SAAO,OAAO,gBAAgB,SAAS,QAAQ,KAAK;;CAEtD,MAAM,OAAO,OAAO,SAAS,KAAK,KAAK,OAAO,SAAS,iBAAiB,CAAC;AACzE,QAAO,OAAO,OAAO,oBAAoB,aAAa,CAAC,KAAK,CAAC,KAC3D,OAAO,SAAS,iBAAiB,CAClC;EACD;AAMJ,MAAM,qBACJ,MACA,aAEA,OAAO,QACL,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,MAAM,QAAQ,EAAE,GAAG,QAAQ,EAAE,CAAC,GAChD,SAAS,mBAAmB,MAAM,SAAS,CAC7C;AAEH,MAAM,aACH,SAEC,YAEA,YAAY,QAAQ,MAAM,CAAC,KACzB,OAAO,SAAS,SACd,UACE,KACA,UAAU;CACR,OAAO,QAAQ;CACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;CACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;CAC1E,GAAI,QAAQ,aAAa,KAAA,KAAa,EAAE,UAAU,QAAQ,UAAU;CACpE,OAAO,CAAC,KAAK;CACd,CAAC,CACH,CACF,EACD,OAAO,SAAS,YAAY;CAC1B,MAAM,QAAQ,QAAQ,KAAK;AAC3B,KAAI,UAAU,KAAA,EACZ,QAAO,OAAO,KAAK,iBAAiB,mCAAmC,CAAC;AAE1E,QAAO,mBAAmB,OAAO,QAAQ,SAAS,CAAC,KACjD,OAAO,KACJ,eAAiC;EAAE;EAAW,OAAO,QAAQ,QAAQ,MAAM;EAAE,EAC/E,CACF;EACD,CACH;AAEL,MAAM,iBACH,SAEC,YAEA,OAAO,QAAQ,QAAQ,QAAQ,YAAY,CAAC,KAC1C,OAAO,SAAS,UACd,UACE,KACA,UAAU;CACR,OAAO,QAAQ;CACf,GAAI,QAAQ,SAAS,KAAA,KAAa,EAAE,MAAM,QAAQ,MAAM;CACxD,GAAI,QAAQ,eAAe,KAAA,KAAa,EAAE,YAAY,QAAQ,YAAY;CAC1E,GAAI,QAAQ,aAAa,KAAA,KAAa,EAAE,UAAU,QAAQ,UAAU;CACpE;CACD,CAAC,CACH,CACF,EACD,OAAO,SAAS,YACd,kBAAkB,QAAQ,MAAM,QAAQ,SAAS,CAAC,KAChD,OAAO,KACJ,gBAAsC;CACrC;CACA,OAAO,QAAQ,QAAQ,MAAM;CAC9B,EACF,CACF,CACF,CACF;;;;;AAML,MAAM,iBAAgF,MAAM,MAEzF,CAAC,KACF,MAAM,KAAK,eAAyB,kBAAkB,EACtD,MAAM,KAAK,kBAA4B,oBAAoB,EAC3D,MAAM,aAAmC,KAAA,EAAU,CACpD;;;;;AAUD,MAAa,QACX,QAEA,OAAO,IAAI,WAAW,WAAW,UAAU,GAAG,YAAY;CACxD,QACE,YAEA,UAAU,IAAI,CAAC,QAA4B,CAAC,KAC1C,OAAO,eAAe,WAAW,YAAY,OAAO,CACrD;CACH,YACE,YAEA,cAAc,IAAI,CAAC,QAAgC,CAAC,KAClD,OAAO,eAAe,WAAW,YAAY,OAAO,CACrD;CACJ,EAAE;;;;;;;;AASL,MAAa,SACX,QAC8E;CAC9E,MAAM,QAAQ,MAAM,OAAO,eAAe,KAAK,IAAI,CAAC;CACpD,MAAM,UAAU,MAAM,OACpB,gBACA,OAAO,IACL,KAAK,IAAI,GACR,OAA8B;EAM7B,QACE,QACG;GACH,MAAM,OAAO,eAAe,IAAI,KAAK;AACrC,UAAO,eAAe,IAAI,UAAU,CAAC,WAAW,SAAS,EAAE,OAAO,CAAC,KACjE,OAAO,QACL,EAAE,MAAM;IACN,GAAG;IACH,OAAO,IAAI;IACX,GAAI,SAAS,KAAA,KAAa,EAAE,MAAM;IACnC,CAAqB,CACvB,CACF;;EAEH,YACE,QACG;GACH,MAAM,OAAO,eAAe,IAAI,KAAK;AACrC,UAAO,eAAe,IAAI,UAAU,CAAC,WAAW,SAAS,EAAE,OAAO,CAAC,KACjE,OAAO,QACL,EAAE,UAAU;IACV,GAAG;IACH,OAAO,IAAI;IACX,GAAI,SAAS,KAAA,KAAa,EAAE,MAAM;IACnC,CAAyB,CAC3B,CACF;;EAEJ,EACF,CACF;AACD,QAAO,MAAM,MAAM,OAAO,QAAQ"}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@effect-uai/jina",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.7.0",
|
|
4
4
|
"description": "Jina AI provider for @effect-uai/core (embeddings, rerank).",
|
|
5
5
|
"keywords": [
|
|
6
6
|
"ai",
|
|
@@ -45,7 +45,7 @@
|
|
|
45
45
|
"devDependencies": {
|
|
46
46
|
"effect": "4.0.0-beta.57",
|
|
47
47
|
"typescript": "^6.0.3",
|
|
48
|
-
"@effect-uai/core": "0.
|
|
48
|
+
"@effect-uai/core": "0.7.0"
|
|
49
49
|
},
|
|
50
50
|
"peerDependencies": {
|
|
51
51
|
"@effect-uai/core": ">=0.2.0 <1",
|
package/src/JinaEmbedding.ts
CHANGED
|
@@ -8,6 +8,7 @@ import type { Embedding, EmbedContentPart, EmbedInput, Usage } from "@effect-uai
|
|
|
8
8
|
type AnyEmbedResponse = { readonly embedding: Embedding; readonly usage: Usage }
|
|
9
9
|
type AnyEmbedManyResponse = { readonly embeddings: ReadonlyArray<Embedding>; readonly usage: Usage }
|
|
10
10
|
import {
|
|
11
|
+
assertEncoding,
|
|
11
12
|
type CommonEmbedManyRequest,
|
|
12
13
|
type CommonEmbedRequest,
|
|
13
14
|
type EmbedEncoding,
|
|
@@ -135,10 +136,15 @@ const contentPartToItem: (part: EmbedContentPart) => WireItem = Match.type<Embed
|
|
|
135
136
|
Match.exhaustive,
|
|
136
137
|
)
|
|
137
138
|
|
|
138
|
-
|
|
139
|
+
// Jina's flat `input[]` can't fuse a multi-part `content[]` into one vector.
|
|
140
|
+
// Bucket 1: silently splitting or dropping parts would change what the vector
|
|
141
|
+
// represents (and the vector count). Reject as `Unsupported` (capability gap),
|
|
142
|
+
// not `InvalidRequest` (wire-shape) - Jina does carry images, just not fused.
|
|
143
|
+
const multiPartContentRejected: AiError.AiError = new AiError.Unsupported({
|
|
139
144
|
provider: "jina",
|
|
140
|
-
|
|
141
|
-
|
|
145
|
+
capability: "multiPartInput",
|
|
146
|
+
reason:
|
|
147
|
+
"Jina treats each input[] entry as one item; multi-part `content[]` can't be fused into one vector. Split into separate `inputs[]` entries.",
|
|
142
148
|
})
|
|
143
149
|
|
|
144
150
|
/**
|
|
@@ -529,25 +535,38 @@ export const layer = (
|
|
|
529
535
|
Effect.map(
|
|
530
536
|
make(cfg),
|
|
531
537
|
(s): EmbeddingModelService => ({
|
|
538
|
+
// Jina emits dense float32 / binary (binary is bit-quantized, packed
|
|
539
|
+
// into int8 bytes on the wire - NOT scalar int8 per dimension; sparse /
|
|
540
|
+
// multivector live on the provider-typed JinaEncoding, off the
|
|
541
|
+
// cross-provider set). Reject a scalar `int8` encoding (bucket 1) so it
|
|
542
|
+
// isn't returned as a mislabeled vector, matching OpenAI / Gemini.
|
|
532
543
|
embed: <E extends EmbedEncoding | undefined = undefined>(
|
|
533
544
|
req: Omit<CommonEmbedRequest, "encoding"> & { readonly encoding?: E },
|
|
534
545
|
) => {
|
|
535
546
|
const task = mapGenericTask(req.task)
|
|
536
|
-
return
|
|
537
|
-
|
|
538
|
-
|
|
539
|
-
|
|
540
|
-
|
|
547
|
+
return assertEncoding(req.encoding, ["float32", "binary"], "jina").pipe(
|
|
548
|
+
Effect.andThen(
|
|
549
|
+
s.embed({
|
|
550
|
+
...req,
|
|
551
|
+
model: req.model as JinaEmbeddingModel,
|
|
552
|
+
...(task !== undefined && { task }),
|
|
553
|
+
} as JinaEmbedRequest),
|
|
554
|
+
),
|
|
555
|
+
) as Effect.Effect<EmbedResponse<E>, AiError.AiError>
|
|
541
556
|
},
|
|
542
557
|
embedMany: <E extends EmbedEncoding | undefined = undefined>(
|
|
543
558
|
req: Omit<CommonEmbedManyRequest, "encoding"> & { readonly encoding?: E },
|
|
544
559
|
) => {
|
|
545
560
|
const task = mapGenericTask(req.task)
|
|
546
|
-
return
|
|
547
|
-
|
|
548
|
-
|
|
549
|
-
|
|
550
|
-
|
|
561
|
+
return assertEncoding(req.encoding, ["float32", "binary"], "jina").pipe(
|
|
562
|
+
Effect.andThen(
|
|
563
|
+
s.embedMany({
|
|
564
|
+
...req,
|
|
565
|
+
model: req.model as JinaEmbeddingModel,
|
|
566
|
+
...(task !== undefined && { task }),
|
|
567
|
+
} as JinaEmbedManyRequest),
|
|
568
|
+
),
|
|
569
|
+
) as Effect.Effect<EmbedManyResponse<E>, AiError.AiError>
|
|
551
570
|
},
|
|
552
571
|
}),
|
|
553
572
|
),
|