@ai-sdk/openai 2.0.139 → 2.0.141

package/dist/internal/index.d.mts

@@ -168,6 +168,8 @@ type OpenAITranscriptionModelId = 'whisper-1' | 'gpt-4o-mini-transcribe' | 'gpt-
 declare const openAITranscriptionProviderOptions: _ai_sdk_provider_utils.LazyValidator<{
     include?: string[] | undefined;
     language?: string | undefined;
+    languages?: string[] | undefined;
+    keywords?: string[] | undefined;
     prompt?: string | undefined;
     temperature?: number | undefined;
     timestampGranularities?: ("word" | "segment")[] | undefined;

package/dist/internal/index.d.ts

@@ -168,6 +168,8 @@ type OpenAITranscriptionModelId = 'whisper-1' | 'gpt-4o-mini-transcribe' | 'gpt-
 declare const openAITranscriptionProviderOptions: _ai_sdk_provider_utils.LazyValidator<{
     include?: string[] | undefined;
     language?: string | undefined;
+    languages?: string[] | undefined;
+    keywords?: string[] | undefined;
     prompt?: string | undefined;
     temperature?: number | undefined;
     timestampGranularities?: ("word" | "segment")[] | undefined;

package/dist/internal/index.js

@@ -129,7 +129,18 @@ function getGptVersion(modelId) {
 var import_provider = require("@ai-sdk/provider");
 var import_provider_utils2 = require("@ai-sdk/provider-utils");
 function serializeToolCallArguments(input) {
-  return JSON.stringify(
+  if (typeof input === "string") {
+    try {
+      const parsed = (0, import_provider_utils2.secureJsonParse)(input);
+      if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) {
+        return "{}";
+      }
+      return input;
+    } catch (e) {
+      return "{}";
+    }
+  }
+  return (0, import_provider_utils2.stringifyWithBigIntAndCycles)(
     typeof input === "object" && input !== null && !Array.isArray(input) ? input : {}
   );
 }
@@ -352,7 +363,7 @@ function convertToOpenAIChatMessages({
             case "content":
             case "json":
             case "error-json":
-              contentValue = JSON.stringify(output.value);
+              contentValue = (0, import_provider_utils2.stringifyWithBigIntAndCycles)(output.value);
               break;
           }
           messages.push({
@@ -392,10 +403,13 @@ function getResponseMetadata({
 }
 
 // src/chat/map-openai-finish-reason.ts
-function mapOpenAIFinishReason(finishReason) {
+function mapOpenAIFinishReason({
+  finishReason,
+  hasRefusal = false
+}) {
   switch (finishReason) {
     case "stop":
-      return "stop";
+      return hasRefusal ? "content-filter" : "stop";
     case "length":
       return "length";
     case "content_filter":
@@ -422,6 +436,8 @@ var openaiChatResponseSchema = (0, import_provider_utils3.lazyValidator)(
           message: import_v42.z.object({
             role: import_v42.z.literal("assistant").nullish(),
             content: import_v42.z.string().nullish(),
+            // set instead of `content` when the model refuses to answer:
+            refusal: import_v42.z.string().nullish(),
             audio: import_v42.z.object({
               transcript: import_v42.z.string().nullish()
             }).nullish(),
@@ -460,6 +476,20 @@ var openaiChatResponseSchema = (0, import_provider_utils3.lazyValidator)(
                   })
                 )
               })
+            ).nullish(),
+            // logprobs of the refusal tokens (set instead of `content` when
+            // the model refuses to answer):
+            refusal: import_v42.z.array(
+              import_v42.z.object({
+                token: import_v42.z.string(),
+                logprob: import_v42.z.number(),
+                top_logprobs: import_v42.z.array(
+                  import_v42.z.object({
+                    token: import_v42.z.string(),
+                    logprob: import_v42.z.number()
+                  })
+                )
+              })
             ).nullish()
           }).nullish(),
           finish_reason: import_v42.z.string().nullish()
@@ -494,6 +524,8 @@ var openaiChatChunkSchema = (0, import_provider_utils3.lazyValidator)(
             delta: import_v42.z.object({
               role: import_v42.z.enum(["assistant"]).nullish(),
               content: import_v42.z.string().nullish(),
+              // set instead of `content` when the model refuses to answer:
+              refusal: import_v42.z.string().nullish(),
               tool_calls: import_v42.z.array(
                 import_v42.z.object({
                   index: import_v42.z.number(),
@@ -529,6 +561,20 @@ var openaiChatChunkSchema = (0, import_provider_utils3.lazyValidator)(
                     })
                   )
                 })
+              ).nullish(),
+              // logprobs of the refusal tokens (set instead of `content` when
+              // the model refuses to answer):
+              refusal: import_v42.z.array(
+                import_v42.z.object({
+                  token: import_v42.z.string(),
+                  logprob: import_v42.z.number(),
+                  top_logprobs: import_v42.z.array(
+                    import_v42.z.object({
+                      token: import_v42.z.string(),
+                      logprob: import_v42.z.number()
+                    })
+                  )
+                })
               ).nullish()
             }).nullish(),
             finish_reason: import_v42.z.string().nullish(),
@@ -960,7 +1006,7 @@ var OpenAIChatLanguageModel = class {
     };
   }
   async doGenerate(options) {
-    var _a2, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o;
+    var _a2, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m, _n, _o, _p;
     const { args: body, warnings } = await this.getArgs(options);
     const {
       responseHeaders,
@@ -992,6 +1038,11 @@ var OpenAIChatLanguageModel = class {
     if (text != null && text.length > 0) {
       content.push({ type: "text", text });
     }
+    const refusal = choice.message.refusal;
+    const hasRefusal = refusal != null && refusal.length > 0;
+    if (hasRefusal) {
+      content.push({ type: "text", text: refusal });
+    }
     for (const toolCall of (_b = choice.message.tool_calls) != null ? _b : []) {
       content.push({
         type: "tool-call",
@@ -1023,18 +1074,26 @@ var OpenAIChatLanguageModel = class {
         cacheWriteTokens: promptTokenDetails.cache_write_tokens
       };
     }
-    if (((_g = choice.logprobs) == null ? void 0 : _g.content) != null) {
-      providerMetadata.openai.logprobs = choice.logprobs.content;
+    const contentLogprobs = (_g = choice.logprobs) == null ? void 0 : _g.content;
+    const refusalLogprobs = (_h = choice.logprobs) == null ? void 0 : _h.refusal;
+    if (contentLogprobs != null || refusalLogprobs != null) {
+      providerMetadata.openai.logprobs = [
+        ...contentLogprobs != null ? contentLogprobs : [],
+        ...refusalLogprobs != null ? refusalLogprobs : []
+      ];
     }
     return {
       content,
-      finishReason: mapOpenAIFinishReason(choice.finish_reason),
+      finishReason: mapOpenAIFinishReason({
+        finishReason: choice.finish_reason,
+        hasRefusal
+      }),
       usage: {
-        inputTokens: (_i = (_h = response.usage) == null ? void 0 : _h.prompt_tokens) != null ? _i : void 0,
-        outputTokens: (_k = (_j = response.usage) == null ? void 0 : _j.completion_tokens) != null ? _k : void 0,
-        totalTokens: (_m = (_l = response.usage) == null ? void 0 : _l.total_tokens) != null ? _m : void 0,
-        reasoningTokens: (_n = completionTokenDetails == null ? void 0 : completionTokenDetails.reasoning_tokens) != null ? _n : void 0,
-        cachedInputTokens: (_o = promptTokenDetails == null ? void 0 : promptTokenDetails.cached_tokens) != null ? _o : void 0
+        inputTokens: (_j = (_i = response.usage) == null ? void 0 : _i.prompt_tokens) != null ? _j : void 0,
+        outputTokens: (_l = (_k = response.usage) == null ? void 0 : _k.completion_tokens) != null ? _l : void 0,
+        totalTokens: (_n = (_m = response.usage) == null ? void 0 : _m.total_tokens) != null ? _n : void 0,
+        reasoningTokens: (_o = completionTokenDetails == null ? void 0 : completionTokenDetails.reasoning_tokens) != null ? _o : void 0,
+        cachedInputTokens: (_p = promptTokenDetails == null ? void 0 : promptTokenDetails.cached_tokens) != null ? _p : void 0
       },
       request: { body },
       response: {
@@ -1078,6 +1137,10 @@ var OpenAIChatLanguageModel = class {
     };
     let metadataExtracted = false;
     let isActiveText = false;
+    let hasRefusal = false;
+    let rawFinishReason = void 0;
+    let hasLogprobs = false;
+    const streamedLogprobs = [];
     const providerMetadata = { openai: {} };
     return {
       stream: response.pipeThrough(
@@ -1090,18 +1153,20 @@ var OpenAIChatLanguageModel = class {
             controller.enqueue({ type: "stream-start", warnings });
           },
           transform(chunk, controller) {
-            var _a2, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l, _m;
+            var _a2, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l;
             if (options.includeRawChunks) {
               controller.enqueue({ type: "raw", rawValue: chunk.rawValue });
             }
             if (!chunk.success) {
               finishReason = "error";
+              rawFinishReason = void 0;
               controller.enqueue({ type: "error", error: chunk.error });
               return;
             }
             const value = chunk.value;
             if ("error" in value) {
               finishReason = "error";
+              rawFinishReason = void 0;
               controller.enqueue({ type: "error", error: value.error });
               return;
             }
@@ -1135,16 +1200,35 @@ var OpenAIChatLanguageModel = class {
             }
             const choice = value.choices[0];
             if ((choice == null ? void 0 : choice.finish_reason) != null) {
-              finishReason = mapOpenAIFinishReason(choice.finish_reason);
+              rawFinishReason = choice.finish_reason;
             }
-            if (((_m = choice == null ? void 0 : choice.logprobs) == null ? void 0 : _m.content) != null) {
-              providerMetadata.openai.logprobs = choice.logprobs.content;
+            if ((choice == null ? void 0 : choice.logprobs) != null) {
+              const contentLogprobs = choice.logprobs.content;
+              const refusalLogprobs = choice.logprobs.refusal;
+              if (contentLogprobs != null || refusalLogprobs != null) {
+                hasLogprobs = true;
+                streamedLogprobs.push(
+                  ...contentLogprobs != null ? contentLogprobs : [],
+                  ...refusalLogprobs != null ? refusalLogprobs : []
+                );
+              }
             }
             if ((choice == null ? void 0 : choice.delta) == null) {
               return;
             }
             const delta = choice.delta;
-            if (delta.content != null) {
+            const textDeltas = [];
+            if (delta.content != null && delta.content.length > 0) {
+              textDeltas.push(delta.content);
+            }
+            if (delta.refusal != null && delta.refusal.length > 0) {
+              hasRefusal = true;
+              textDeltas.push(delta.refusal);
+            }
+            if (textDeltas.length === 0 && delta.content === "") {
+              textDeltas.push(delta.content);
+            }
+            for (const textDelta of textDeltas) {
               if (!isActiveText) {
                 controller.enqueue({ type: "text-start", id: "0" });
                 isActiveText = true;
@@ -1152,7 +1236,7 @@ var OpenAIChatLanguageModel = class {
               controller.enqueue({
                 type: "text-delta",
                 id: "0",
-                delta: delta.content
+                delta: textDelta
               });
             }
             if (delta.tool_calls != null) {
@@ -1177,6 +1261,15 @@ var OpenAIChatLanguageModel = class {
               controller.enqueue({ type: "text-end", id: "0" });
             }
             toolCallTracker.flush();
+            if (rawFinishReason != null) {
+              finishReason = mapOpenAIFinishReason({
+                finishReason: rawFinishReason,
+                hasRefusal
+              });
+            }
+            if (hasLogprobs) {
+              providerMetadata.openai.logprobs = streamedLogprobs;
+            }
             controller.enqueue({
               type: "finish",
               finishReason,
@@ -2024,6 +2117,16 @@ var openAITranscriptionProviderOptions = (0, import_provider_utils16.lazyValidat
        * The language of the input audio in ISO-639-1 format.
        */
       language: import_v411.z.string().optional(),
+      /**
+       * Expected languages for REST transcription, e.g. with gpt-transcribe.
+       * Use this instead of language. The API validates model support.
+       */
+      languages: import_v411.z.array(import_v411.z.string()).optional(),
+      /**
+       * Vocabulary hints for REST transcription, e.g. names or technical terms.
+       * The API validates model support.
+       */
+      keywords: import_v411.z.array(import_v411.z.string()).optional(),
       /**
        * An optional text to guide the model's style or continue a previous audio segment.
        */
@@ -2135,6 +2238,8 @@ var OpenAITranscriptionModel = class {
       const transcriptionModelOptions = {
         include: openAIOptions.include,
         language: openAIOptions.language,
+        languages: openAIOptions.languages,
+        keywords: openAIOptions.keywords,
         prompt: openAIOptions.prompt,
         // https://platform.openai.com/docs/api-reference/audio/createTranscription#audio_createtranscription-response_format
         // prefer verbose_json to get segments for models that support it
@@ -2898,7 +3003,7 @@ async function convertToOpenAIResponsesInput({
                 type: "function_call",
                 call_id: part.toolCallId,
                 name: part.toolName,
-                arguments: JSON.stringify(part.input),
+                arguments: serializeToolCallArguments2(part.input),
                 ...isAsync != null && { async: isAsync },
                 id: canUseItemId ? id : void 0
               });
@@ -2960,7 +3065,7 @@ async function convertToOpenAIResponsesInput({
                   } else if (reasoningMessage !== void 0) {
                     warnings.push({
                       type: "other",
-                      message: `Cannot append empty reasoning part to existing reasoning sequence. Skipping reasoning part: ${JSON.stringify(part)}.`
+                      message: `Cannot append empty reasoning part to existing reasoning sequence. Skipping reasoning part: ${(0, import_provider_utils23.stringifyWithBigIntAndCycles)(part)}.`
                     });
                   }
                   if (reasoningMessage === void 0) {
@@ -2981,7 +3086,7 @@ async function convertToOpenAIResponsesInput({
               } else {
                 warnings.push({
                   type: "other",
-                  message: `Non-OpenAI reasoning parts are not supported. Skipping reasoning part: ${JSON.stringify(part)}.`
+                  message: `Non-OpenAI reasoning parts are not supported. Skipping reasoning part: ${(0, import_provider_utils23.stringifyWithBigIntAndCycles)(part)}.`
                 });
               }
               break;
@@ -3022,10 +3127,10 @@ async function convertToOpenAIResponsesInput({
               break;
             case "json":
             case "error-json":
-              contentValue = promptCacheBreakpoint == null ? JSON.stringify(output.value) : [
+              contentValue = promptCacheBreakpoint == null ? (0, import_provider_utils23.stringifyWithBigIntAndCycles)(output.value) : [
                 {
                   type: "input_text",
-                  text: JSON.stringify(output.value),
+                  text: (0, import_provider_utils23.stringifyWithBigIntAndCycles)(output.value),
                   prompt_cache_breakpoint: promptCacheBreakpoint
                 }
               ];
@@ -3098,16 +3203,29 @@ var openaiResponsesReasoningProviderOptionsSchema = (0, import_provider_utils23.
     })
   )
 );
+function serializeToolCallArguments2(input) {
+  if (typeof input === "string") {
+    try {
+      (0, import_provider_utils23.secureJsonParse)(input);
+      return input;
+    } catch (e) {
+      return (0, import_provider_utils23.stringifyWithBigIntAndCycles)(input);
+    }
+  }
+  return (0, import_provider_utils23.stringifyWithBigIntAndCycles)(input === void 0 ? {} : input);
+}
 
 // src/responses/map-openai-responses-finish-reason.ts
 function mapOpenAIResponseFinishReason({
   finishReason,
-  hasFunctionCall
+  hasFunctionCall,
+  hasRefusal = false
 }) {
   switch (finishReason) {
     case void 0:
     case null:
-      return hasFunctionCall ? "tool-calls" : "stop";
+      if (hasFunctionCall) return "tool-calls";
+      return hasRefusal ? "content-filter" : "stop";
     case "max_output_tokens":
       return "length";
     case "content_filter":
@@ -3140,6 +3258,20 @@ var openaiResponsesChunkSchema = (0, import_provider_utils24.lazyValidator)(
           })
         ).nullish()
       }),
+      import_v417.z.object({
+        type: import_v417.z.literal("response.refusal.delta"),
+        item_id: import_v417.z.string(),
+        output_index: import_v417.z.number().nullish(),
+        content_index: import_v417.z.number().nullish(),
+        delta: import_v417.z.string()
+      }),
+      import_v417.z.object({
+        type: import_v417.z.literal("response.refusal.done"),
+        item_id: import_v417.z.string(),
+        output_index: import_v417.z.number().nullish(),
+        content_index: import_v417.z.number().nullish(),
+        refusal: import_v417.z.string()
+      }),
       import_v417.z.object({
         type: import_v417.z.enum(["response.completed", "response.incomplete"]),
         response: import_v417.z.object({
@@ -3426,57 +3558,64 @@ var openaiResponsesResponseSchema = (0, import_provider_utils24.lazyValidator)(
             role: import_v417.z.literal("assistant"),
             id: import_v417.z.string(),
             phase: import_v417.z.enum(["commentary", "final_answer"]).nullish(),
+            // the API returns either text or a refusal as content parts:
             content: import_v417.z.array(
-              import_v417.z.object({
-                type: import_v417.z.literal("output_text"),
-                text: import_v417.z.string(),
-                logprobs: import_v417.z.array(
-                  import_v417.z.object({
-                    token: import_v417.z.string(),
-                    logprob: import_v417.z.number(),
-                    top_logprobs: import_v417.z.array(
-                      import_v417.z.object({
-                        token: import_v417.z.string(),
-                        logprob: import_v417.z.number()
-                      })
-                    )
-                  })
-                ).nullish(),
-                annotations: import_v417.z.array(
-                  import_v417.z.discriminatedUnion("type", [
-                    import_v417.z.object({
-                      type: import_v417.z.literal("url_citation"),
-                      start_index: import_v417.z.number(),
-                      end_index: import_v417.z.number(),
-                      url: import_v417.z.string(),
-                      title: import_v417.z.string()
-                    }),
-                    import_v417.z.object({
-                      type: import_v417.z.literal("file_citation"),
-                      file_id: import_v417.z.string(),
-                      filename: import_v417.z.string().nullish(),
-                      index: import_v417.z.number().nullish(),
-                      start_index: import_v417.z.number().nullish(),
-                      end_index: import_v417.z.number().nullish(),
-                      quote: import_v417.z.string().nullish()
-                    }),
-                    import_v417.z.object({
-                      type: import_v417.z.literal("container_file_citation"),
-                      container_id: import_v417.z.string(),
-                      file_id: import_v417.z.string(),
-                      filename: import_v417.z.string().nullish(),
-                      start_index: import_v417.z.number().nullish(),
-                      end_index: import_v417.z.number().nullish(),
-                      index: import_v417.z.number().nullish()
-                    }),
+              import_v417.z.discriminatedUnion("type", [
+                import_v417.z.object({
+                  type: import_v417.z.literal("output_text"),
+                  text: import_v417.z.string(),
+                  logprobs: import_v417.z.array(
                     import_v417.z.object({
-                      type: import_v417.z.literal("file_path"),
-                      file_id: import_v417.z.string(),
-                      index: import_v417.z.number().nullish()
+                      token: import_v417.z.string(),
+                      logprob: import_v417.z.number(),
+                      top_logprobs: import_v417.z.array(
+                        import_v417.z.object({
+                          token: import_v417.z.string(),
+                          logprob: import_v417.z.number()
+                        })
+                      )
                     })
-                  ])
-                )
-              })
+                  ).nullish(),
+                  annotations: import_v417.z.array(
+                    import_v417.z.discriminatedUnion("type", [
+                      import_v417.z.object({
+                        type: import_v417.z.literal("url_citation"),
+                        start_index: import_v417.z.number(),
+                        end_index: import_v417.z.number(),
+                        url: import_v417.z.string(),
+                        title: import_v417.z.string()
+                      }),
+                      import_v417.z.object({
+                        type: import_v417.z.literal("file_citation"),
+                        file_id: import_v417.z.string(),
+                        filename: import_v417.z.string().nullish(),
+                        index: import_v417.z.number().nullish(),
+                        start_index: import_v417.z.number().nullish(),
+                        end_index: import_v417.z.number().nullish(),
+                        quote: import_v417.z.string().nullish()
+                      }),
+                      import_v417.z.object({
+                        type: import_v417.z.literal("container_file_citation"),
+                        container_id: import_v417.z.string(),
+                        file_id: import_v417.z.string(),
+                        filename: import_v417.z.string().nullish(),
+                        start_index: import_v417.z.number().nullish(),
+                        end_index: import_v417.z.number().nullish(),
+                        index: import_v417.z.number().nullish()
+                      }),
+                      import_v417.z.object({
+                        type: import_v417.z.literal("file_path"),
+                        file_id: import_v417.z.string(),
+                        index: import_v417.z.number().nullish()
+                      })
+                    ])
+                  )
+                }),
+                import_v417.z.object({
+                  type: import_v417.z.literal("refusal"),
+                  refusal: import_v417.z.string()
+                })
+              ])
             )
           }),
           import_v417.z.object({
@@ -4320,6 +4459,7 @@ var OpenAIResponsesLanguageModel = class {
     const content = [];
     const logprobs = [];
     let hasFunctionCall = false;
+    let hasRefusal = false;
     for (const part of response.output) {
       switch (part.type) {
         case "reasoning": {
@@ -4377,6 +4517,20 @@ var OpenAIResponsesLanguageModel = class {
         }
         case "message": {
           for (const contentPart of part.content) {
+            if (contentPart.type === "refusal") {
+              hasRefusal = true;
+              content.push({
+                type: "text",
+                text: contentPart.refusal,
+                providerMetadata: {
+                  [providerKey]: {
+                    itemId: part.id,
+                    ...part.phase != null && { phase: part.phase }
+                  }
+                }
+              });
+              continue;
+            }
             if (((_c = (_b = options.providerOptions) == null ? void 0 : _b.openai) == null ? void 0 : _c.logprobs) && contentPart.logprobs) {
               logprobs.push(contentPart.logprobs);
             }
@@ -4577,7 +4731,8 @@ var OpenAIResponsesLanguageModel = class {
       content,
       finishReason: mapOpenAIResponseFinishReason({
         finishReason: (_y = response.incomplete_details) == null ? void 0 : _y.reason,
-        hasFunctionCall
+        hasFunctionCall,
+        hasRefusal
       }),
       usage: {
         inputTokens: usage.input_tokens,
@@ -4636,6 +4791,7 @@ var OpenAIResponsesLanguageModel = class {
     const ongoingAnnotations = [];
     let activeMessagePhase;
     let hasFunctionCall = false;
+    let hasRefusal = false;
     const activeReasoning = {};
     let serviceTier;
     let reasoningContext;
@@ -4977,6 +5133,15 @@ var OpenAIResponsesLanguageModel = class {
               if (((_h = (_g = options.providerOptions) == null ? void 0 : _g.openai) == null ? void 0 : _h.logprobs) && value.logprobs) {
                 logprobs.push(value.logprobs);
               }
+            } else if (isRefusalDeltaChunk(value)) {
+              hasRefusal = true;
+              controller.enqueue({
+                type: "text-delta",
+                id: value.item_id,
+                delta: value.delta
+              });
+            } else if (isRefusalDoneChunk(value)) {
+              hasRefusal = true;
             } else if (value.type === "response.reasoning_summary_part.added") {
               if (value.summary_index > 0) {
                 const activeReasoningPart = activeReasoning[value.item_id];
@@ -5033,7 +5198,8 @@ var OpenAIResponsesLanguageModel = class {
             } else if (isResponseFinishedChunk(value)) {
               finishReason = mapOpenAIResponseFinishReason({
                 finishReason: (_k = value.response.incomplete_details) == null ? void 0 : _k.reason,
-                hasFunctionCall
+                hasFunctionCall,
+                hasRefusal
               });
               usage.inputTokens = value.response.usage.input_tokens;
               usage.outputTokens = value.response.usage.output_tokens;
@@ -5110,6 +5276,12 @@ var OpenAIResponsesLanguageModel = class {
     };
   }
 };
+function isRefusalDeltaChunk(chunk) {
+  return chunk.type === "response.refusal.delta";
+}
+function isRefusalDoneChunk(chunk) {
+  return chunk.type === "response.refusal.done";
+}
 function isTextDeltaChunk(chunk) {
   return chunk.type === "response.output_text.delta";
 }