@effect/ai-openai-compat 4.0.0-beta.4 → 4.0.0-beta.41

package/src/OpenAiEmbeddingModel.ts

@@ -0,0 +1,203 @@
+/**
+ * OpenAI Embedding Model implementation.
+ *
+ * Provides an EmbeddingModel implementation for OpenAI-compatible embeddings APIs.
+ *
+ * @since 1.0.0
+ */
+import * as Effect from "effect/Effect"
+import { dual } from "effect/Function"
+import * as Layer from "effect/Layer"
+import * as ServiceMap from "effect/ServiceMap"
+import type { Simplify } from "effect/Types"
+import * as AiError from "effect/unstable/ai/AiError"
+import * as EmbeddingModel from "effect/unstable/ai/EmbeddingModel"
+import * as AiModel from "effect/unstable/ai/Model"
+import type { CreateEmbedding200, CreateEmbeddingRequestJson } from "./OpenAiClient.ts"
+import { OpenAiClient } from "./OpenAiClient.ts"
+
+/**
+ * @since 1.0.0
+ * @category models
+ */
+export type Model = string
+
+/**
+ * Service definition for OpenAI embedding model configuration.
+ *
+ * @since 1.0.0
+ * @category context
+ */
+export class Config extends ServiceMap.Service<
+  Config,
+  Simplify<
+    & Partial<
+      Omit<
+        CreateEmbeddingRequestJson,
+        "input"
+      >
+    >
+    & {
+      readonly [x: string]: unknown
+    }
+  >
+>()("@effect/ai-openai-compat/OpenAiEmbeddingModel/Config") {}
+
+/**
+ * @since 1.0.0
+ * @category constructors
+ */
+export const model = (
+  model: string,
+  options: {
+    readonly dimensions: number
+    readonly config?: Omit<typeof Config.Service, "model" | "dimensions">
+  }
+): AiModel.Model<"openai", EmbeddingModel.EmbeddingModel | EmbeddingModel.Dimensions, OpenAiClient> =>
+  AiModel.make(
+    "openai",
+    model,
+    Layer.merge(
+      layer({
+        model,
+        config: {
+          ...options.config,
+          dimensions: options.dimensions
+        }
+      }),
+      Layer.succeed(EmbeddingModel.Dimensions, options.dimensions)
+    )
+  )
+
+/**
+ * Creates an OpenAI embedding model service.
+ *
+ * @since 1.0.0
+ * @category constructors
+ */
+export const make = Effect.fnUntraced(function*({ model, config: providerConfig }: {
+  readonly model: string
+  readonly config?: Omit<typeof Config.Service, "model"> | undefined
+}): Effect.fn.Return<EmbeddingModel.Service, never, OpenAiClient> {
+  const client = yield* OpenAiClient
+
+  const makeConfig = Effect.gen(function*() {
+    const services = yield* Effect.services<never>()
+    return { model, ...providerConfig, ...services.mapUnsafe.get(Config.key) }
+  })
+
+  return yield* EmbeddingModel.make({
+    embedMany: Effect.fnUntraced(function*({ inputs }) {
+      const config = yield* makeConfig
+      const response = yield* client.createEmbedding({ ...config, input: inputs })
+      return yield* mapProviderResponse(inputs.length, response)
+    })
+  })
+})
+
+/**
+ * Creates a layer for the OpenAI embedding model.
+ *
+ * @since 1.0.0
+ * @category layers
+ */
+export const layer = (options: {
+  readonly model: string
+  readonly config?: Omit<typeof Config.Service, "model"> | undefined
+}): Layer.Layer<EmbeddingModel.EmbeddingModel, never, OpenAiClient> =>
+  Layer.effect(EmbeddingModel.EmbeddingModel, make(options))
+
+/**
+ * Provides config overrides for OpenAI embedding model operations.
+ *
+ * @since 1.0.0
+ * @category configuration
+ */
+export const withConfigOverride: {
+  /**
+   * Provides config overrides for OpenAI embedding model operations.
+   *
+   * @since 1.0.0
+   * @category configuration
+   */
+  (overrides: typeof Config.Service): <A, E, R>(self: Effect.Effect<A, E, R>) => Effect.Effect<A, E, Exclude<R, Config>>
+  /**
+   * Provides config overrides for OpenAI embedding model operations.
+   *
+   * @since 1.0.0
+   * @category configuration
+   */
+  <A, E, R>(self: Effect.Effect<A, E, R>, overrides: typeof Config.Service): Effect.Effect<A, E, Exclude<R, Config>>
+} = dual<
+  /**
+   * Provides config overrides for OpenAI embedding model operations.
+   *
+   * @since 1.0.0
+   * @category configuration
+   */
+  (overrides: typeof Config.Service) => <A, E, R>(self: Effect.Effect<A, E, R>) => Effect.Effect<A, E, Exclude<R, Config>>,
+  /**
+   * Provides config overrides for OpenAI embedding model operations.
+   *
+   * @since 1.0.0
+   * @category configuration
+   */
+  <A, E, R>(self: Effect.Effect<A, E, R>, overrides: typeof Config.Service) => Effect.Effect<A, E, Exclude<R, Config>>
+>(2, (self, overrides) =>
+  Effect.flatMap(
+    Effect.serviceOption(Config),
+    (config) =>
+      Effect.provideService(self, Config, {
+        ...(config._tag === "Some" ? config.value : {}),
+        ...overrides
+      })
+  ))
+
+const mapProviderResponse = (
+  inputLength: number,
+  response: CreateEmbedding200
+): Effect.Effect<EmbeddingModel.ProviderResponse, AiError.AiError> => {
+  if (response.data.length !== inputLength) {
+    return Effect.fail(
+      invalidOutput(`Provider returned ${response.data.length} embeddings but expected ${inputLength}`)
+    )
+  }
+
+  const results = new Array<Array<number>>(inputLength)
+  const seen = new Set<number>()
+
+  for (const entry of response.data) {
+    if (!Number.isInteger(entry.index) || entry.index < 0 || entry.index >= inputLength) {
+      return Effect.fail(invalidOutput(`Provider returned invalid embedding index: ${entry.index}`))
+    }
+    if (seen.has(entry.index)) {
+      return Effect.fail(invalidOutput(`Provider returned duplicate embedding index: ${entry.index}`))
+    }
+    if (!Array.isArray(entry.embedding)) {
+      return Effect.fail(invalidOutput(`Provider returned non-vector embedding at index ${entry.index}`))
+    }
+
+    seen.add(entry.index)
+    results[entry.index] = [...entry.embedding]
+  }
+
+  if (seen.size !== inputLength) {
+    return Effect.fail(
+      invalidOutput(`Provider returned embeddings for ${seen.size} inputs but expected ${inputLength}`)
+    )
+  }
+
+  return Effect.succeed({
+    results,
+    usage: {
+      inputTokens: response.usage?.prompt_tokens
+    }
+  })
+}
+
+const invalidOutput = (description: string): AiError.AiError =>
+  AiError.make({
+    module: "OpenAiEmbeddingModel",
+    method: "embedMany",
+    reason: new AiError.InvalidOutputError({ description })
+  })

package/src/OpenAiError.ts

@@ -52,51 +52,43 @@ export type OpenAiRateLimitMetadata = OpenAiErrorMetadata & {
 }
 
 declare module "effect/unstable/ai/AiError" {
-  export interface RateLimitError {
-    readonly metadata: {
-      readonly openai?: OpenAiRateLimitMetadata | null
-    }
+  export interface RateLimitErrorMetadata {
+    readonly openai?: OpenAiRateLimitMetadata | null
   }
 
-  export interface QuotaExhaustedError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface QuotaExhaustedErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface AuthenticationError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface AuthenticationErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface ContentPolicyError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface ContentPolicyErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface InvalidRequestError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface InvalidRequestErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface InternalProviderError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface InternalProviderErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface InvalidOutputError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface InvalidOutputErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 
-  export interface UnknownError {
-    readonly metadata: {
-      readonly openai?: OpenAiErrorMetadata | null
-    }
+  export interface StructuredOutputErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
+  }
+
+  export interface UnsupportedSchemaErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
+  }
+
+  export interface UnknownErrorMetadata {
+    readonly openai?: OpenAiErrorMetadata | null
   }
 }

package/src/OpenAiLanguageModel.ts

@@ -8,9 +8,10 @@
  */
 import * as DateTime from "effect/DateTime"
 import * as Effect from "effect/Effect"
-import * as Base64 from "effect/encoding/Base64"
+import * as Encoding from "effect/Encoding"
 import { dual } from "effect/Function"
 import * as Layer from "effect/Layer"
+import * as Option from "effect/Option"
 import * as Predicate from "effect/Predicate"
 import * as Redactable from "effect/Redactable"
 import type * as Schema from "effect/Schema"
@@ -101,6 +102,7 @@ export class Config extends ServiceMap.Service<
        * Defaults to `true`.
        */
       readonly strictJsonSchema?: boolean | undefined
+      readonly [x: string]: unknown
     }
   >
 >()("@effect/ai-openai-compat/OpenAiLanguageModel/Config") {}
@@ -316,7 +318,7 @@ export const model = (
   model: string,
   config?: Omit<typeof Config.Service, "model">
 ): AiModel.Model<"openai", LanguageModel.LanguageModel, OpenAiClient> =>
-  AiModel.make("openai", layer({ model, config }))
+  AiModel.make("openai", model, layer({ model, config }))
 
 // TODO
 // /**
@@ -327,7 +329,7 @@ export const model = (
 //   model: string,
 //   config?: Omit<typeof Config.Service, "model">
 // ): AiModel.Model<"openai", LanguageModel.LanguageModel | Tokenizer.Tokenizer, OpenAiClient> =>
-//   AiModel.make("openai", layerWithTokenizer({ model, config }))
+//   AiModel.make("openai", model, layerWithTokenizer({ model, config }))
 
 /**
  * Creates an OpenAI language model service.
@@ -386,6 +388,7 @@ export const make = Effect.fnUntraced(function*({ model, config: providerConfig
   )
 
   return yield* LanguageModel.make({
+    codecTransformer: toCodecOpenAI,
     generateText: Effect.fnUntraced(
       function*(options) {
         const config = yield* makeConfig
@@ -423,10 +426,7 @@ export const make = Effect.fnUntraced(function*({ model, config: providerConfig
           })
         )
     )
-  }).pipe(Effect.provideService(
-    LanguageModel.CurrentCodecTransformer,
-    toCodecOpenAI
-  ))
+  })
 })
 
 /**
@@ -561,7 +561,7 @@ const prepareMessages = Effect.fnUntraced(
                   }
 
                   if (part.data instanceof Uint8Array) {
-                    const base64 = Base64.encode(part.data)
+                    const base64 = Encoding.encodeBase64(part.data)
                     const imageUrl = `data:${mediaType};base64,${base64}`
                     content.push({ type: "input_image", image_url: imageUrl, detail })
                   }
@@ -575,7 +575,7 @@ const prepareMessages = Effect.fnUntraced(
                   }
 
                   if (part.data instanceof Uint8Array) {
-                    const base64 = Base64.encode(part.data)
+                    const base64 = Encoding.encodeBase64(part.data)
                     const fileName = part.fileName ?? `part-${index}.pdf`
                     const fileData = `data:application/pdf;base64,${base64}`
                     content.push({ type: "input_file", filename: fileName, file_data: fileData })
@@ -782,7 +782,7 @@ const buildHttpRequestDetails = (
   method: request.method,
   url: request.url,
   urlParams: Array.from(request.urlParams),
-  hash: request.hash,
+  hash: Option.getOrUndefined(request.hash),
   headers: Redactable.redact(request.headers) as Record<string, string>
 })
 
@@ -1002,11 +1002,13 @@ const makeStreamResponse = Effect.fnUntraced(
           hasToolCalls = hasToolCalls || choice.delta.tool_calls.length > 0
           choice.delta.tool_calls.forEach((deltaTool, indexInChunk) => {
             const toolIndex = deltaTool.index ?? indexInChunk
-            const toolId = deltaTool.id ?? `${event.id}_tool_${toolIndex}`
+            const activeToolCall = activeToolCalls[toolIndex]
+            const toolId = activeToolCall?.id ?? deltaTool.id ?? `${event.id}_tool_${toolIndex}`
             const providerToolName = deltaTool.function?.name
-            const toolName = toolNameMapper.getCustomName(providerToolName ?? "unknown_tool")
+            const toolName = providerToolName !== undefined
+              ? toolNameMapper.getCustomName(providerToolName)
+              : activeToolCall?.name ?? toolNameMapper.getCustomName("unknown_tool")
             const argumentsDelta = deltaTool.function?.arguments ?? ""
-            const activeToolCall = activeToolCalls[toolIndex]
 
             if (Predicate.isUndefined(activeToolCall)) {
               activeToolCalls[toolIndex] = {
@@ -1220,6 +1222,7 @@ const toChatCompletionsRequest = (payload: CreateResponse): CreateResponseReques
   const toolChoice = toChatToolChoice(payload.tool_choice)
 
   return {
+    ...extractCustomRequestProperties(payload),
     model: payload.model ?? "",
     messages: messages.length > 0 ? messages : [{ role: "user", content: "" }],
     ...(payload.temperature !== undefined ? { temperature: payload.temperature } : undefined),
@@ -1231,12 +1234,54 @@ const toChatCompletionsRequest = (payload: CreateResponse): CreateResponseReques
       ? { parallel_tool_calls: payload.parallel_tool_calls }
       : undefined),
     ...(payload.service_tier !== undefined ? { service_tier: payload.service_tier } : undefined),
+    ...(payload.reasoning !== undefined ? { reasoning: payload.reasoning } : undefined),
     ...(responseFormat !== undefined ? { response_format: responseFormat } : undefined),
     ...(tools.length > 0 ? { tools } : undefined),
     ...(toolChoice !== undefined ? { tool_choice: toolChoice } : undefined)
   }
 }
 
+const createResponseKnownProperties = new Set<string>([
+  "metadata",
+  "top_logprobs",
+  "temperature",
+  "top_p",
+  "user",
+  "safety_identifier",
+  "prompt_cache_key",
+  "service_tier",
+  "prompt_cache_retention",
+  "previous_response_id",
+  "model",
+  "reasoning",
+  "background",
+  "max_output_tokens",
+  "max_tool_calls",
+  "text",
+  "tools",
+  "tool_choice",
+  "truncation",
+  "input",
+  "include",
+  "parallel_tool_calls",
+  "store",
+  "instructions",
+  "stream",
+  "conversation",
+  "modalities",
+  "seed"
+])
+
+const extractCustomRequestProperties = (payload: CreateResponse): Record<string, unknown> => {
+  const customProperties: Record<string, unknown> = {}
+  for (const [key, value] of Object.entries(payload)) {
+    if (!createResponseKnownProperties.has(key)) {
+      customProperties[key] = value
+    }
+  }
+  return customProperties
+}
+
 const toChatResponseFormat = (
   format: TextResponseFormatConfiguration | undefined
 ): CreateResponseRequestJson["response_format"] | undefined => {

package/src/index.ts

@@ -14,6 +14,15 @@ export * as OpenAiClient from "./OpenAiClient.ts"
  */
 export * as OpenAiConfig from "./OpenAiConfig.ts"
 
+/**
+ * OpenAI Embedding Model implementation.
+ *
+ * Provides an EmbeddingModel implementation for OpenAI-compatible embeddings APIs.
+ *
+ * @since 1.0.0
+ */
+export * as OpenAiEmbeddingModel from "./OpenAiEmbeddingModel.ts"
+
 /**
  * @since 1.0.0
  */

package/src/internal/errors.ts

@@ -153,12 +153,12 @@ export const parseRateLimitHeaders = (headers: Record<string, string>) => {
   let retryAfter: Duration.Duration | undefined
   if (retryAfterRaw !== undefined) {
     const parsed = Number.parse(retryAfterRaw)
-    if (parsed !== undefined) {
-      retryAfter = Duration.seconds(parsed)
+    if (Option.isSome(parsed)) {
+      retryAfter = Duration.seconds(parsed.value)
     }
   }
   const remainingRaw = headers["x-ratelimit-remaining-requests"]
-  const remaining = remainingRaw !== undefined ? Number.parse(remainingRaw) ?? null : null
+  const remaining = remainingRaw !== undefined ? Option.getOrNull(Number.parse(remainingRaw)) : null
   return {
     retryAfter,
     limit: headers["x-ratelimit-limit-requests"] ?? null,
@@ -175,7 +175,7 @@ export const buildHttpRequestDetails = (
   method: request.method,
   url: request.url,
   urlParams: Array.from(request.urlParams),
-  hash: request.hash,
+  hash: Option.getOrUndefined(request.hash),
   headers: Redactable.redact(request.headers) as Record<string, string>
 })