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

package/src/OpenAiError.ts

@@ -57,51 +57,43 @@ 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 * as Schema from "effect/Schema"
@@ -23,6 +24,7 @@ import * as AiError from "effect/unstable/ai/AiError"
 import * as IdGenerator from "effect/unstable/ai/IdGenerator"
 import * as LanguageModel from "effect/unstable/ai/LanguageModel"
 import * as AiModel from "effect/unstable/ai/Model"
+import { toCodecOpenAI } from "effect/unstable/ai/OpenAiStructuredOutput"
 import type * as Prompt from "effect/unstable/ai/Prompt"
 import type * as Response from "effect/unstable/ai/Response"
 import * as Tool from "effect/unstable/ai/Tool"
@@ -319,7 +321,7 @@ export const model = (
   model: (string & {}) | Model,
   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
 // /**
@@ -330,7 +332,7 @@ export const model = (
 //   model: (string & {}) | Model,
 //   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.
@@ -369,7 +371,7 @@ export const make = Effect.fnUntraced(function*({ model, config: providerConfig
         options,
         toolNameMapper
       })
-      const responseFormat = prepareResponseFormat({
+      const responseFormat = yield* prepareResponseFormat({
         config,
         options
       })
@@ -381,14 +383,16 @@ export const make = Effect.fnUntraced(function*({ model, config: providerConfig
           verbosity: config.text?.verbosity ?? null,
           format: responseFormat
         },
-        ...(Predicate.isNotUndefined(tools) ? { tools } : undefined),
-        ...(Predicate.isNotUndefined(toolChoice) ? { tool_choice: toolChoice } : undefined)
+        ...(tools ? { tools } : undefined),
+        ...(toolChoice ? { tool_choice: toolChoice } : undefined),
+        ...(options.previousResponseId ? { previous_response_id: options.previousResponseId } : undefined)
       }
       return request
     }
   )
 
   return yield* LanguageModel.make({
+    codecTransformer: toCodecOpenAI,
     generateText: Effect.fnUntraced(
       function*(options) {
         const config = yield* makeConfig
@@ -547,16 +551,17 @@ const prepareMessages = Effect.fnUntraced(
     if (config.store === false && capabilities.isReasoningModel) {
       include.add("reasoning.encrypted_content")
     }
-    if (Predicate.isNotUndefined(codeInterpreterTool)) {
+    if (codeInterpreterTool) {
       include.add("code_interpreter_call.outputs")
     }
-    if (Predicate.isNotUndefined(webSearchTool) || Predicate.isNotUndefined(webSearchPreviewTool)) {
+    if (webSearchTool || webSearchPreviewTool) {
       include.add("web_search_call.action.sources")
     }
 
     const messages: Array<typeof Generated.InputItem.Encoded> = []
+    const prompt = options.incrementalPrompt ?? options.prompt
 
-    for (const message of options.prompt.content) {
+    for (const message of prompt.content) {
       switch (message.role) {
         case "system": {
           messages.push({
@@ -592,7 +597,7 @@ 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 })
                   }
@@ -606,7 +611,7 @@ 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 })
@@ -921,7 +926,7 @@ 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>
 })
 
@@ -1036,10 +1041,11 @@ const makeResponse = Effect.fnUntraced(
 
         case "function_call": {
           hasToolCalls = true
+
           const toolName = part.name
-          const toolParams = part.arguments
-          const params = yield* Effect.try({
-            try: () => Tool.unsafeSecureJsonParse(toolParams),
+
+          const toolParams = yield* Effect.try({
+            try: () => Tool.unsafeSecureJsonParse(part.arguments),
             catch: (cause) =>
               AiError.make({
                 module: "OpenAiLanguageModel",
@@ -1051,6 +1057,9 @@ const makeResponse = Effect.fnUntraced(
                 })
               })
           })
+
+          const params = yield* transformToolCallParams(options.tools, part.name, toolParams)
+
           parts.push({
             type: "tool-call",
             id: part.call_id,
@@ -1370,16 +1379,42 @@ const makeStreamResponse = Effect.fnUntraced(
     // Track annotations for current message to include in text-end metadata
     const activeAnnotations: Array<typeof Generated.Annotation.Encoded> = []
 
+    type ReasoningSummaryPartStatus = "active" | "can-conclude" | "concluded"
+    type ReasoningPart = {
+      encryptedContent: string | undefined
+      summaryParts: Record<number, ReasoningSummaryPartStatus>
+    }
+
     // Track active reasoning items with state machine for proper concluding logic
-    const activeReasoning: Record<string, {
-      readonly encryptedContent: string | undefined
-      readonly summaryParts: Record<number, "active" | "can-conclude" | "concluded">
-    }> = {}
+    const activeReasoning: Record<string, ReasoningPart> = {}
+
+    const getOrCreateReasoningPart = (
+      itemId: string,
+      encryptedContent?: string | null
+    ): ReasoningPart => {
+      const activePart = activeReasoning[itemId]
+      if (Predicate.isNotUndefined(activePart)) {
+        if (Predicate.isNotNullish(encryptedContent)) {
+          activePart.encryptedContent = encryptedContent
+        }
+        return activePart
+      }
+
+      const reasoningPart: ReasoningPart = {
+        encryptedContent: Predicate.isNotNullish(encryptedContent) ? encryptedContent : undefined,
+        summaryParts: {}
+      }
+      activeReasoning[itemId] = reasoningPart
+      return reasoningPart
+    }
 
     // Track active tool calls with optional provider-specific state
     const activeToolCalls: Record<number, {
       readonly id: string
       readonly name: string
+      readonly functionCall?: {
+        emitted: boolean
+      }
       readonly applyPatch?: {
         hasDiff: boolean
         endEmitted: boolean
@@ -1529,7 +1564,8 @@ const makeStreamResponse = Effect.fnUntraced(
               case "function_call": {
                 activeToolCalls[event.output_index] = {
                   id: event.item.call_id,
-                  name: event.item.name
+                  name: event.item.name,
+                  functionCall: { emitted: false }
                 }
                 parts.push({
                   type: "tool-params-start",
@@ -1572,21 +1608,20 @@ const makeStreamResponse = Effect.fnUntraced(
               }
 
               case "reasoning": {
-                const encryptedContent = event.item.encrypted_content ?? undefined
-                activeReasoning[event.item.id] = {
-                  encryptedContent,
-                  summaryParts: { 0: "active" }
-                }
-                parts.push({
-                  type: "reasoning-start",
-                  id: `${event.item.id}:0`,
-                  metadata: {
-                    openai: {
-                      ...makeItemIdMetadata(event.item.id),
-                      ...makeEncryptedContentMetadata(event.item.encrypted_content)
+                const reasoningPart = getOrCreateReasoningPart(event.item.id, event.item.encrypted_content)
+                if (Predicate.isUndefined(reasoningPart.summaryParts[0])) {
+                  reasoningPart.summaryParts[0] = "active"
+                  parts.push({
+                    type: "reasoning-start",
+                    id: `${event.item.id}:0`,
+                    metadata: {
+                      openai: {
+                        ...makeItemIdMetadata(event.item.id),
+                        ...makeEncryptedContentMetadata(reasoningPart.encryptedContent)
+                      }
                     }
-                  }
-                })
+                  })
+                }
                 break
               }
 
@@ -1729,12 +1764,20 @@ const makeStreamResponse = Effect.fnUntraced(
               }
 
               case "function_call": {
+                const toolCall = activeToolCalls[event.output_index]
+                if (Predicate.isNotUndefined(toolCall?.functionCall?.emitted) && toolCall.functionCall.emitted) {
+                  delete activeToolCalls[event.output_index]
+                  break
+                }
                 delete activeToolCalls[event.output_index]
+
                 hasToolCalls = true
+
                 const toolName = event.item.name
-                const toolParams = event.item.arguments
-                const params = yield* Effect.try({
-                  try: () => Tool.unsafeSecureJsonParse(toolParams),
+                const toolArgs = event.item.arguments
+
+                const toolParams = yield* Effect.try({
+                  try: () => Tool.unsafeSecureJsonParse(toolArgs),
                   catch: (cause) =>
                     AiError.make({
                       module: "OpenAiLanguageModel",
@@ -1746,10 +1789,14 @@ const makeStreamResponse = Effect.fnUntraced(
                       })
                     })
                 })
+
+                const params = yield* transformToolCallParams(options.tools, toolName, toolParams)
+
                 parts.push({
                   type: "tool-params-end",
                   id: event.item.call_id
                 })
+
                 parts.push({
                   type: "tool-call",
                   id: event.item.call_id,
@@ -1757,6 +1804,7 @@ const makeStreamResponse = Effect.fnUntraced(
                   params,
                   metadata: { openai: { ...makeItemIdMetadata(event.item.id) } }
                 })
+
                 break
               }
 
@@ -1862,7 +1910,7 @@ const makeStreamResponse = Effect.fnUntraced(
               }
 
               case "reasoning": {
-                const reasoningPart = activeReasoning[event.item.id]
+                const reasoningPart = getOrCreateReasoningPart(event.item.id, event.item.encrypted_content)
                 for (const [summaryIndex, status] of Object.entries(reasoningPart.summaryParts)) {
                   if (status === "active" || status === "can-conclude") {
                     parts.push({
@@ -1871,7 +1919,7 @@ const makeStreamResponse = Effect.fnUntraced(
                       metadata: {
                         openai: {
                           ...makeItemIdMetadata(event.item.id),
-                          ...makeEncryptedContentMetadata(event.item.encrypted_content)
+                          ...makeEncryptedContentMetadata(reasoningPart.encryptedContent)
                         }
                       }
                     })
@@ -2006,6 +2054,48 @@ const makeStreamResponse = Effect.fnUntraced(
             break
           }
 
+          case "response.function_call_arguments.done": {
+            const toolCall = activeToolCalls[event.output_index]
+            if (
+              Predicate.isNotUndefined(toolCall?.functionCall) &&
+              !toolCall.functionCall.emitted
+            ) {
+              hasToolCalls = true
+
+              const toolParams = yield* Effect.try({
+                try: () => Tool.unsafeSecureJsonParse(event.arguments),
+                catch: (cause) =>
+                  AiError.make({
+                    module: "OpenAiLanguageModel",
+                    method: "makeStreamResponse",
+                    reason: new AiError.ToolParameterValidationError({
+                      toolName: toolCall.name,
+                      toolParams: {},
+                      description: `Failed securely JSON parse tool parameters: ${cause}`
+                    })
+                  })
+              })
+
+              const params = yield* transformToolCallParams(options.tools, toolCall.name, toolParams)
+
+              parts.push({
+                type: "tool-params-end",
+                id: toolCall.id
+              })
+
+              parts.push({
+                type: "tool-call",
+                id: toolCall.id,
+                name: toolCall.name,
+                params,
+                metadata: { openai: { ...makeItemIdMetadata(event.item_id) } }
+              })
+
+              toolCall.functionCall.emitted = true
+            }
+            break
+          }
+
           case "response.apply_patch_call_operation_diff.delta": {
             const toolCall = activeToolCalls[event.output_index]
             if (Predicate.isNotUndefined(toolCall?.applyPatch)) {
@@ -2095,28 +2185,28 @@ const makeStreamResponse = Effect.fnUntraced(
           }
 
           case "response.reasoning_summary_part.added": {
-            // The first reasoning start is pushed in the `response.output_item.added` block
+            const reasoningPart = getOrCreateReasoningPart(event.item_id)
             if (event.summary_index > 0) {
-              const reasoningPart = activeReasoning[event.item_id]
-              if (Predicate.isNotUndefined(reasoningPart)) {
-                // Conclude all can-conclude parts before starting new one
-                for (const [summaryIndex, status] of Object.entries(reasoningPart.summaryParts)) {
-                  if (status === "can-conclude") {
-                    parts.push({
-                      type: "reasoning-end",
-                      id: `${event.item_id}:${summaryIndex}`,
-                      metadata: {
-                        openai: {
-                          ...makeItemIdMetadata(event.item_id),
-                          ...makeEncryptedContentMetadata(reasoningPart.encryptedContent)
-                        }
+              // Conclude all can-conclude parts before starting new one
+              for (const [summaryIndex, status] of Object.entries(reasoningPart.summaryParts)) {
+                if (status === "can-conclude") {
+                  parts.push({
+                    type: "reasoning-end",
+                    id: `${event.item_id}:${summaryIndex}`,
+                    metadata: {
+                      openai: {
+                        ...makeItemIdMetadata(event.item_id),
+                        ...makeEncryptedContentMetadata(reasoningPart.encryptedContent)
                       }
-                    })
-                    reasoningPart.summaryParts[Number(summaryIndex)] = "concluded"
-                  }
+                    }
+                  })
+                  reasoningPart.summaryParts[Number(summaryIndex)] = "concluded"
                 }
-                reasoningPart.summaryParts[event.summary_index] = "active"
               }
+            }
+
+            if (Predicate.isUndefined(reasoningPart.summaryParts[event.summary_index])) {
+              reasoningPart.summaryParts[event.summary_index] = "active"
               parts.push({
                 type: "reasoning-start",
                 id: `${event.item_id}:${event.summary_index}`,
@@ -2142,6 +2232,7 @@ const makeStreamResponse = Effect.fnUntraced(
           }
 
           case "response.reasoning_summary_part.done": {
+            const reasoningPart = getOrCreateReasoningPart(event.item_id)
             // When OpenAI stores message data, we can immediately conclude the
             // reasoning part given that we do not need the encrypted content
             if (config.store === true) {
@@ -2151,11 +2242,11 @@ const makeStreamResponse = Effect.fnUntraced(
                 metadata: { openai: { ...makeItemIdMetadata(event.item_id) } }
               })
               // Mark the summary part concluded
-              activeReasoning[event.item_id].summaryParts[event.summary_index] = "concluded"
+              reasoningPart.summaryParts[event.summary_index] = "concluded"
             } else {
               // Mark the summary part as can-conclude given we still need a
               // final summary part with the encrypted content
-              activeReasoning[event.item_id].summaryParts[event.summary_index] = "can-conclude"
+              reasoningPart.summaryParts[event.summary_index] = "can-conclude"
             }
             break
           }
@@ -2281,12 +2372,14 @@ const prepareTools = Effect.fnUntraced(function*<Tools extends ReadonlyArray<Too
   for (const tool of allowedTools) {
     if (Tool.isUserDefined(tool)) {
       const strict = Tool.getStrictMode(tool) ?? config.strictJsonSchema ?? true
+      const description = Tool.getDescription(tool)
+      const parameters = yield* tryJsonSchema(tool.parametersSchema, "prepareTools")
       tools.push({
         type: "function",
         name: tool.name,
-        description: Tool.getDescription(tool) ?? null,
-        parameters: Tool.getJsonSchema(tool) as { readonly [x: string]: Schema.Json },
-        strict
+        parameters,
+        strict,
+        ...(Predicate.isNotUndefined(description) ? { description } : undefined)
       })
     }
 
@@ -2480,23 +2573,45 @@ const makeItemIdMetadata = (itemId: string | undefined) => Predicate.isNotUndefi
 const makeEncryptedContentMetadata = (encryptedContent: string | null | undefined) =>
   Predicate.isNotNullish(encryptedContent) ? { encryptedContent } : undefined
 
-const prepareResponseFormat = ({ config, options }: {
+const unsupportedSchemaError = (error: unknown, method: string): AiError.AiError =>
+  AiError.make({
+    module: "OpenAiLanguageModel",
+    method,
+    reason: new AiError.UnsupportedSchemaError({
+      description: error instanceof Error ? error.message : String(error)
+    })
+  })
+
+const tryCodecTransform = <S extends Schema.Top>(schema: S, method: string) =>
+  Effect.try({
+    try: () => toCodecOpenAI(schema),
+    catch: (error) => unsupportedSchemaError(error, method)
+  })
+
+const tryJsonSchema = <S extends Schema.Top>(schema: S, method: string) =>
+  Effect.try({
+    try: () => Tool.getJsonSchemaFromSchema(schema, { transformer: toCodecOpenAI }),
+    catch: (error) => unsupportedSchemaError(error, method)
+  })
+
+const prepareResponseFormat = Effect.fnUntraced(function*({ config, options }: {
   readonly config: typeof Config.Service
   readonly options: LanguageModel.ProviderOptions
-}): typeof Generated.TextResponseFormatConfiguration.Encoded => {
+}): Effect.fn.Return<typeof Generated.TextResponseFormatConfiguration.Encoded, AiError.AiError> {
   if (options.responseFormat.type === "json") {
     const name = options.responseFormat.objectName
     const schema = options.responseFormat.schema
+    const jsonSchema = yield* tryJsonSchema(schema, "prepareResponseFormat")
     return {
       type: "json_schema",
       name,
       description: AST.resolveDescription(schema.ast) ?? "Response with a JSON object",
-      schema: Tool.getJsonSchemaFromSchema(schema) as any,
+      schema: jsonSchema,
       strict: config.strictJsonSchema ?? true
     }
   }
   return { type: "text" }
-}
+})
 
 interface ModelCapabilities {
   readonly isReasoningModel: boolean
@@ -2606,3 +2721,40 @@ const getUsage = (usage: Generated.ResponseUsage | null | undefined): Response.U
     }
   }
 }
+
+const transformToolCallParams = Effect.fnUntraced(function*<Tools extends ReadonlyArray<Tool.Any>>(
+  tools: Tools,
+  toolName: string,
+  toolParams: unknown
+): Effect.fn.Return<unknown, AiError.AiError> {
+  const tool = tools.find((tool) => tool.name === toolName)
+
+  if (Predicate.isUndefined(tool)) {
+    return yield* AiError.make({
+      module: "OpenAiLanguageModel",
+      method: "makeResponse",
+      reason: new AiError.ToolNotFoundError({
+        toolName,
+        availableTools: tools.map((tool) => tool.name)
+      })
+    })
+  }
+
+  const { codec } = yield* tryCodecTransform(tool.parametersSchema, "makeResponse")
+
+  const transform = Schema.decodeEffect(codec)
+
+  return yield* (
+    transform(toolParams) as Effect.Effect<unknown, Schema.SchemaError>
+  ).pipe(Effect.mapError((error) =>
+    AiError.make({
+      module: "OpenAiLanguageModel",
+      method: "makeResponse",
+      reason: new AiError.ToolParameterValidationError({
+        toolName,
+        toolParams,
+        description: error.issue.toString()
+      })
+    })
+  ))
+})

package/src/index.ts

@@ -24,6 +24,15 @@ export * as OpenAiClient from "./OpenAiClient.ts"
  */
 export * as OpenAiConfig from "./OpenAiConfig.ts"
 
+/**
+ * OpenAI Embedding Model implementation.
+ *
+ * Provides an EmbeddingModel implementation for OpenAI's embeddings API.
+ *
+ * @since 1.0.0
+ */
+export * as OpenAiEmbeddingModel from "./OpenAiEmbeddingModel.ts"
+
 /**
  * OpenAI error metadata augmentation.
  *

package/src/internal/errors.ts

@@ -161,12 +161,14 @@ export const parseRateLimitHeaders = (headers: Record<string, string>) => {
   let retryAfter: Duration.Duration | undefined
   if (Predicate.isNotUndefined(retryAfterRaw)) {
     const parsed = Number.parse(retryAfterRaw)
-    if (Predicate.isNotUndefined(parsed)) {
-      retryAfter = Duration.seconds(parsed)
+    if (Option.isSome(parsed)) {
+      retryAfter = Duration.seconds(parsed.value)
     }
   }
   const remainingRaw = headers["x-ratelimit-remaining-requests"]
-  const remaining = Predicate.isNotUndefined(remainingRaw) ? Number.parse(remainingRaw) ?? null : null
+  const remaining = Predicate.isNotUndefined(remainingRaw)
+    ? Option.getOrNull(Number.parse(remainingRaw))
+    : null
   return {
     retryAfter,
     limit: headers["x-ratelimit-limit-requests"] ?? null,
@@ -187,7 +189,7 @@ 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>
 })