ai 3.4.2 → 3.4.4

package/CHANGELOG.md

@@ -1,5 +1,19 @@
 # ai
 
+## 3.4.4
+
+### Patch Changes
+
+- 4db074b: fix (ai/core): correct whitespace in generateText continueSteps
+- 1297e1b: fix (ai/core): correct whitespace in streamText continueSteps
+
+## 3.4.3
+
+### Patch Changes
+
+- b270ae3: feat (ai/core): streamText continueSteps (experimental)
+- b270ae3: chore (ai/core): rename generateText continuationSteps to continueSteps
+
 ## 3.4.2
 
 ### Patch Changes

package/dist/index.d.mts

@@ -1362,6 +1362,12 @@ type StepResult<TOOLS extends Record<string, CoreTool>> = {
   results that can be fully encapsulated in the provider.
      */
     readonly experimental_providerMetadata: ProviderMetadata | undefined;
+    /**
+  The type of step that this result is for. The first step is always
+  an "initial" step, and subsequent steps are either "continue" steps
+  or "tool-result" steps.
+     */
+    readonly stepType: 'initial' | 'continue' | 'tool-result';
 };
 
 /**
@@ -1491,7 +1497,7 @@ If set and supported by the model, calls will generate deterministic results.
 @returns
 A result object that contains the generated text, the results of the tool calls, and additional information.
  */
-declare function generateText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxAutomaticRoundtrips, maxToolRoundtrips, maxSteps, experimental_continuationSteps: continuationSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
+declare function generateText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxAutomaticRoundtrips, maxToolRoundtrips, maxSteps, experimental_continuationSteps, experimental_continueSteps: continueSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1531,9 +1537,18 @@ A maximum number is required to prevent infinite loops in the case of misconfigu
 By default, it's set to 1, which means that only a single LLM call is made.
      */
     maxSteps?: number;
+    /**
+@deprecated Use `experimental_continueSteps` instead.
+     */
     experimental_continuationSteps?: boolean;
     /**
-     * Optional telemetry configuration (experimental).
+When enabled, the model will perform additional steps if the finish reason is "length" (experimental).
+
+By default, it's set to false.
+     */
+    experimental_continueSteps?: boolean;
+    /**
+Optional telemetry configuration (experimental).
      */
     experimental_telemetry?: TelemetrySettings;
     /**
@@ -1832,7 +1847,7 @@ If set and supported by the model, calls will generate deterministic results.
 @return
 A result object for accessing different stream types and additional information.
  */
-declare function streamText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxToolRoundtrips, maxSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, experimental_toolCallStreaming: toolCallStreaming, onChunk, onFinish, onStepFinish, _internal: { now, generateId, currentDate, }, ...settings }: CallSettings & Prompt & {
+declare function streamText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxToolRoundtrips, maxSteps, experimental_continueSteps: continueSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, experimental_toolCallStreaming: toolCallStreaming, onChunk, onFinish, onStepFinish, _internal: { now, generateId, currentDate, }, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1869,6 +1884,12 @@ By default, it's set to 1, which means that only a single LLM call is made.
  */
     maxSteps?: number;
     /**
+When enabled, the model will perform additional steps if the finish reason is "length" (experimental).
+
+By default, it's set to false.
+     */
+    experimental_continueSteps?: boolean;
+    /**
 Optional telemetry configuration (experimental).
      */
     experimental_telemetry?: TelemetrySettings;
@@ -1896,7 +1917,7 @@ Callback that is called when the LLM response and all request tool executions
 
 The usage is the combined usage of all steps.
      */
-    onFinish?: (event: StepResult<TOOLS> & {
+    onFinish?: (event: Omit<StepResult<TOOLS>, 'stepType'> & {
         /**
 Details for all steps.
        */

package/dist/index.d.ts

@@ -1362,6 +1362,12 @@ type StepResult<TOOLS extends Record<string, CoreTool>> = {
   results that can be fully encapsulated in the provider.
      */
     readonly experimental_providerMetadata: ProviderMetadata | undefined;
+    /**
+  The type of step that this result is for. The first step is always
+  an "initial" step, and subsequent steps are either "continue" steps
+  or "tool-result" steps.
+     */
+    readonly stepType: 'initial' | 'continue' | 'tool-result';
 };
 
 /**
@@ -1491,7 +1497,7 @@ If set and supported by the model, calls will generate deterministic results.
 @returns
 A result object that contains the generated text, the results of the tool calls, and additional information.
  */
-declare function generateText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxAutomaticRoundtrips, maxToolRoundtrips, maxSteps, experimental_continuationSteps: continuationSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
+declare function generateText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxAutomaticRoundtrips, maxToolRoundtrips, maxSteps, experimental_continuationSteps, experimental_continueSteps: continueSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1531,9 +1537,18 @@ A maximum number is required to prevent infinite loops in the case of misconfigu
 By default, it's set to 1, which means that only a single LLM call is made.
      */
     maxSteps?: number;
+    /**
+@deprecated Use `experimental_continueSteps` instead.
+     */
     experimental_continuationSteps?: boolean;
     /**
-     * Optional telemetry configuration (experimental).
+When enabled, the model will perform additional steps if the finish reason is "length" (experimental).
+
+By default, it's set to false.
+     */
+    experimental_continueSteps?: boolean;
+    /**
+Optional telemetry configuration (experimental).
      */
     experimental_telemetry?: TelemetrySettings;
     /**
@@ -1832,7 +1847,7 @@ If set and supported by the model, calls will generate deterministic results.
 @return
 A result object for accessing different stream types and additional information.
  */
-declare function streamText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxToolRoundtrips, maxSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, experimental_toolCallStreaming: toolCallStreaming, onChunk, onFinish, onStepFinish, _internal: { now, generateId, currentDate, }, ...settings }: CallSettings & Prompt & {
+declare function streamText<TOOLS extends Record<string, CoreTool>>({ model, tools, toolChoice, system, prompt, messages, maxRetries, abortSignal, headers, maxToolRoundtrips, maxSteps, experimental_continueSteps: continueSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, experimental_toolCallStreaming: toolCallStreaming, onChunk, onFinish, onStepFinish, _internal: { now, generateId, currentDate, }, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1869,6 +1884,12 @@ By default, it's set to 1, which means that only a single LLM call is made.
  */
     maxSteps?: number;
     /**
+When enabled, the model will perform additional steps if the finish reason is "length" (experimental).
+
+By default, it's set to false.
+     */
+    experimental_continueSteps?: boolean;
+    /**
 Optional telemetry configuration (experimental).
      */
     experimental_telemetry?: TelemetrySettings;
@@ -1896,7 +1917,7 @@ Callback that is called when the LLM response and all request tool executions
 
 The usage is the combined usage of all steps.
      */
-    onFinish?: (event: StepResult<TOOLS> & {
+    onFinish?: (event: Omit<StepResult<TOOLS>, 'stepType'> & {
         /**
 Details for all steps.
        */

package/dist/index.js

@@ -2977,6 +2977,19 @@ function prepareToolsAndToolChoice({
   };
 }
 
+// core/util/split-on-last-whitespace.ts
+var lastWhitespaceRegexp = /^([\s\S]*?)(\s+)(\S*)$/;
+function splitOnLastWhitespace(text) {
+  const match = text.match(lastWhitespaceRegexp);
+  return match ? { prefix: match[1], whitespace: match[2], suffix: match[3] } : void 0;
+}
+
+// core/util/remove-text-after-last-whitespace.ts
+function removeTextAfterLastWhitespace(text) {
+  const match = splitOnLastWhitespace(text);
+  return match ? match.prefix + match.whitespace : text;
+}
+
 // core/generate-text/parse-tool-call.ts
 var import_provider_utils8 = require("@ai-sdk/provider-utils");
 var import_ui_utils4 = require("@ai-sdk/ui-utils");
@@ -3052,7 +3065,8 @@ async function generateText({
   maxAutomaticRoundtrips = 0,
   maxToolRoundtrips = maxAutomaticRoundtrips,
   maxSteps = maxToolRoundtrips != null ? maxToolRoundtrips + 1 : 1,
-  experimental_continuationSteps: continuationSteps = false,
+  experimental_continuationSteps,
+  experimental_continueSteps: continueSteps = experimental_continuationSteps != null ? experimental_continuationSteps : false,
   experimental_telemetry: telemetry,
   experimental_providerMetadata: providerMetadata,
   _internal: {
@@ -3096,7 +3110,7 @@ async function generateText({
     }),
     tracer,
     fn: async (span) => {
-      var _a12, _b, _c, _d, _e, _f;
+      var _a12, _b, _c, _d, _e;
       const retry = retryWithExponentialBackoff({ maxRetries });
       const validatedPrompt = validatePrompt({
         system,
@@ -3156,7 +3170,7 @@ async function generateText({
             }),
             tracer,
             fn: async (span2) => {
-              var _a13, _b2, _c2, _d2, _e2, _f2;
+              var _a13, _b2, _c2, _d2, _e2, _f;
               const result = await model.doGenerate({
                 mode,
                 ...callSettings,
@@ -3169,7 +3183,7 @@ async function generateText({
               const responseData = {
                 id: (_b2 = (_a13 = result.response) == null ? void 0 : _a13.id) != null ? _b2 : generateId3(),
                 timestamp: (_d2 = (_c2 = result.response) == null ? void 0 : _c2.timestamp) != null ? _d2 : currentDate(),
-                modelId: (_f2 = (_e2 = result.response) == null ? void 0 : _e2.modelId) != null ? _f2 : model.modelId
+                modelId: (_f = (_e2 = result.response) == null ? void 0 : _e2.modelId) != null ? _f : model.modelId
               };
               span2.setAttributes(
                 selectTelemetryAttributes({
@@ -3223,13 +3237,24 @@ async function generateText({
         usage.completionTokens += currentUsage.completionTokens;
         usage.promptTokens += currentUsage.promptTokens;
         usage.totalTokens += currentUsage.totalTokens;
-        if (stepType === "continuation") {
-          text += " " + ((_b = currentModelResponse.text) != null ? _b : "");
-        } else {
-          text = (_c = currentModelResponse.text) != null ? _c : "";
+        let nextStepType = "done";
+        if (++stepCount < maxSteps) {
+          if (continueSteps && currentModelResponse.finishReason === "length" && // only use continue when there are no tool calls:
+          currentToolCalls.length === 0) {
+            nextStepType = "continue";
+          } else if (
+            // there are tool calls:
+            currentToolCalls.length > 0 && // all current tool calls have results:
+            currentToolResults.length === currentToolCalls.length
+          ) {
+            nextStepType = "tool-result";
+          }
         }
+        const stepText = nextStepType === "continue" ? removeTextAfterLastWhitespace((_b = currentModelResponse.text) != null ? _b : "") : (_c = currentModelResponse.text) != null ? _c : "";
+        text = nextStepType === "continue" || stepType === "continue" ? text + stepText : stepText;
         const currentStep = {
-          text: (_d = currentModelResponse.text) != null ? _d : "",
+          stepType,
+          text: stepText,
           toolCalls: currentToolCalls,
           toolResults: currentToolResults,
           finishReason: currentModelResponse.finishReason,
@@ -3238,20 +3263,20 @@ async function generateText({
           logprobs: currentModelResponse.logprobs,
           response: {
             ...currentModelResponse.response,
-            headers: (_e = currentModelResponse.rawResponse) == null ? void 0 : _e.headers
+            headers: (_d = currentModelResponse.rawResponse) == null ? void 0 : _d.headers
           },
           experimental_providerMetadata: currentModelResponse.providerMetadata
         };
         steps.push(currentStep);
         await (onStepFinish == null ? void 0 : onStepFinish(currentStep));
-        if (stepType === "continuation") {
+        if (stepType === "continue") {
           const lastResponseMessage = responseMessages.pop();
           promptMessages.pop();
           if (typeof lastResponseMessage.content === "string") {
             lastResponseMessage.content = text;
           } else {
             lastResponseMessage.content.push({
-              text: " " + currentModelResponse.text,
+              text: stepText,
               type: "text"
             });
           }
@@ -3259,6 +3284,18 @@ async function generateText({
           promptMessages.push(
             convertToLanguageModelMessage(lastResponseMessage, null)
           );
+        } else if (nextStepType === "continue") {
+          const newResponseMessages = toResponseMessages({
+            text,
+            toolCalls: currentToolCalls,
+            toolResults: currentToolResults
+          });
+          responseMessages.push(...newResponseMessages);
+          promptMessages.push(
+            ...newResponseMessages.map(
+              (message) => convertToLanguageModelMessage(message, null)
+            )
+          );
         } else {
           const newResponseMessages = toResponseMessages({
             text: currentModelResponse.text,
@@ -3272,20 +3309,7 @@ async function generateText({
             )
           );
         }
-        if (++stepCount >= maxSteps) {
-          stepType = "done";
-        } else if (continuationSteps === true && currentStep.finishReason === "length" && // only use continuation when there are no tool calls:
-        currentToolCalls.length === 0) {
-          stepType = "continuation";
-        } else if (
-          // there are tool calls:
-          currentToolCalls.length > 0 && // all current tool calls have results:
-          currentToolResults.length === currentToolCalls.length
-        ) {
-          stepType = "tool-result";
-        } else {
-          stepType = "done";
-        }
+        stepType = nextStepType;
       } while (stepType !== "done");
       span.setAttributes(
         selectTelemetryAttributes({
@@ -3320,7 +3344,7 @@ async function generateText({
         warnings: currentModelResponse.warnings,
         response: {
           ...currentModelResponse.response,
-          headers: (_f = currentModelResponse.rawResponse) == null ? void 0 : _f.headers
+          headers: (_e = currentModelResponse.rawResponse) == null ? void 0 : _e.headers
         },
         logprobs: currentModelResponse.logprobs,
         responseMessages,
@@ -3766,6 +3790,7 @@ async function streamText({
   headers,
   maxToolRoundtrips = 0,
   maxSteps = maxToolRoundtrips != null ? maxToolRoundtrips + 1 : 1,
+  experimental_continueSteps: continueSteps = false,
   experimental_telemetry: telemetry,
   experimental_providerMetadata: providerMetadata,
   experimental_toolCallStreaming: toolCallStreaming = false,
@@ -3910,6 +3935,7 @@ async function streamText({
         telemetry,
         startTimestampMs,
         maxSteps,
+        continueSteps,
         startStep,
         promptMessages,
         modelId: model.modelId,
@@ -3933,6 +3959,7 @@ var DefaultStreamTextResult = class {
     telemetry,
     startTimestampMs,
     maxSteps,
+    continueSteps,
     startStep,
     promptMessages,
     modelId,
@@ -3984,7 +4011,9 @@ var DefaultStreamTextResult = class {
         promptTokens: 0,
         completionTokens: 0,
         totalTokens: 0
-      }
+      },
+      stepType,
+      previousStepText = ""
     }) {
       const stepToolCalls = [];
       const stepToolResults = [];
@@ -3997,12 +4026,25 @@ var DefaultStreamTextResult = class {
       let stepProviderMetadata;
       let stepFirstChunk = true;
       let stepText = "";
+      let fullStepText = stepType === "continue" ? previousStepText : "";
       let stepLogProbs;
       let stepResponse = {
         id: generateId3(),
         timestamp: currentDate(),
         modelId
       };
+      let chunkBuffer = "";
+      let chunkTextPublished = false;
+      async function publishTextChunk({
+        controller,
+        chunk
+      }) {
+        controller.enqueue(chunk);
+        stepText += chunk.textDelta;
+        fullStepText += chunk.textDelta;
+        chunkTextPublished = true;
+        await (onChunk == null ? void 0 : onChunk({ chunk }));
+      }
       addStream(
         stream2.pipeThrough(
           new TransformStream({
@@ -4028,9 +4070,22 @@ var DefaultStreamTextResult = class {
               const chunkType = chunk.type;
               switch (chunkType) {
                 case "text-delta": {
-                  controller.enqueue(chunk);
-                  stepText += chunk.textDelta;
-                  await (onChunk == null ? void 0 : onChunk({ chunk }));
+                  if (continueSteps) {
+                    chunkBuffer += chunk.textDelta;
+                    const split = splitOnLastWhitespace(chunkBuffer);
+                    if (split != null) {
+                      chunkBuffer = split.suffix;
+                      await publishTextChunk({
+                        controller,
+                        chunk: {
+                          type: "text-delta",
+                          textDelta: split.prefix + split.whitespace
+                        }
+                      });
+                    }
+                  } else {
+                    await publishTextChunk({ controller, chunk });
+                  }
                   break;
                 }
                 case "tool-call": {
@@ -4086,6 +4141,30 @@ var DefaultStreamTextResult = class {
             // invoke onFinish callback and resolve toolResults promise when the stream is about to close:
             async flush(controller) {
               const stepToolCallsJson = stepToolCalls.length > 0 ? JSON.stringify(stepToolCalls) : void 0;
+              let nextStepType = "done";
+              if (currentStep + 1 < maxSteps) {
+                if (continueSteps && stepFinishReason === "length" && // only use continue when there are no tool calls:
+                stepToolCalls.length === 0) {
+                  nextStepType = "continue";
+                } else if (
+                  // there are tool calls:
+                  stepToolCalls.length > 0 && // all current tool calls have results:
+                  stepToolResults.length === stepToolCalls.length
+                ) {
+                  nextStepType = "tool-result";
+                }
+              }
+              if (continueSteps && chunkBuffer.length > 0 && (nextStepType !== "continue" || // when the next step is a regular step, publish the buffer
+              stepType === "continue" && !chunkTextPublished)) {
+                await publishTextChunk({
+                  controller,
+                  chunk: {
+                    type: "text-delta",
+                    textDelta: chunkBuffer
+                  }
+                });
+                chunkBuffer = "";
+              }
               try {
                 doStreamSpan2.setAttributes(
                   selectTelemetryAttributes({
@@ -4129,6 +4208,7 @@ var DefaultStreamTextResult = class {
                 response: stepResponse
               });
               const stepResult = {
+                stepType,
                 text: stepText,
                 toolCalls: stepToolCalls,
                 toolResults: stepToolResults,
@@ -4147,21 +4227,24 @@ var DefaultStreamTextResult = class {
                 completionTokens: usage.completionTokens + stepUsage.completionTokens,
                 totalTokens: usage.totalTokens + stepUsage.totalTokens
               };
-              if (
-                // there are tool calls:
-                stepToolCalls.length > 0 && // all current tool calls have results:
-                stepToolResults.length === stepToolCalls.length && // the number of steps is less than the maximum:
-                currentStep + 1 < maxSteps
-              ) {
-                promptMessages2.push(
-                  ...toResponseMessages({
+              if (nextStepType !== "done") {
+                if (stepType === "continue") {
+                  const lastPromptMessage = promptMessages2[promptMessages2.length - 1];
+                  lastPromptMessage.content.push({
                     text: stepText,
-                    toolCalls: stepToolCalls,
-                    toolResults: stepToolResults
-                  }).map(
-                    (message) => convertToLanguageModelMessage(message, null)
-                  )
-                );
+                    type: "text"
+                  });
+                } else {
+                  promptMessages2.push(
+                    ...toResponseMessages({
+                      text: stepText,
+                      toolCalls: stepToolCalls,
+                      toolResults: stepToolResults
+                    }).map(
+                      (message) => convertToLanguageModelMessage(message, null)
+                    )
+                  );
+                }
                 const {
                   result,
                   doStreamSpan: doStreamSpan3,
@@ -4178,7 +4261,9 @@ var DefaultStreamTextResult = class {
                   doStreamSpan: doStreamSpan3,
                   currentStep: currentStep + 1,
                   promptMessages: promptMessages2,
-                  usage: combinedUsage
+                  usage: combinedUsage,
+                  stepType: nextStepType,
+                  previousStepText: fullStepText
                 });
                 return;
               }
@@ -4197,7 +4282,7 @@ var DefaultStreamTextResult = class {
                     telemetry,
                     attributes: {
                       "ai.response.finishReason": stepFinishReason,
-                      "ai.response.text": { output: () => stepText },
+                      "ai.response.text": { output: () => fullStepText },
                       "ai.response.toolCalls": {
                         output: () => stepToolCallsJson
                       },
@@ -4205,27 +4290,38 @@ var DefaultStreamTextResult = class {
                       "ai.usage.completionTokens": combinedUsage.completionTokens,
                       // deprecated
                       "ai.finishReason": stepFinishReason,
-                      "ai.result.text": { output: () => stepText },
+                      "ai.result.text": { output: () => fullStepText },
                       "ai.result.toolCalls": {
                         output: () => stepToolCallsJson
                       }
                     }
                   })
                 );
-                const responseMessages = stepResults.reduce(
-                  (responseMessages2, step) => [
+                const responseMessages = stepResults.reduce((responseMessages2, step) => {
+                  if (step.stepType === "continue") {
+                    const lastResponseMessage = responseMessages2.pop();
+                    if (typeof lastResponseMessage.content === "string") {
+                      lastResponseMessage.content += step.text;
+                    } else {
+                      lastResponseMessage.content.push({
+                        text: step.text,
+                        type: "text"
+                      });
+                    }
+                    return [...responseMessages2, lastResponseMessage];
+                  }
+                  return [
                     ...responseMessages2,
                     ...toResponseMessages({
                       text: step.text,
                       toolCalls: step.toolCalls,
                       toolResults: step.toolResults
                     })
-                  ],
-                  []
-                );
+                  ];
+                }, []);
                 resolveUsage(combinedUsage);
                 resolveFinishReason(stepFinishReason);
-                resolveText(stepText);
+                resolveText(fullStepText);
                 resolveToolCalls(stepToolCalls);
                 resolveProviderMetadata(stepProviderMetadata);
                 resolveToolResults(stepToolResults);
@@ -4239,7 +4335,7 @@ var DefaultStreamTextResult = class {
                   finishReason: stepFinishReason,
                   logprobs: stepLogProbs,
                   usage: combinedUsage,
-                  text: stepText,
+                  text: fullStepText,
                   toolCalls: stepToolCalls,
                   // The tool results are inferred as a never[] type, because they are
                   // optional and the execute method with an inferred result type is
@@ -4272,7 +4368,8 @@ var DefaultStreamTextResult = class {
       doStreamSpan,
       currentStep: 0,
       promptMessages,
-      usage: void 0
+      usage: void 0,
+      stepType: "initial"
     });
   }
   /**