ai 3.4.1 → 3.4.3

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # ai
 
+## 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
+
+- e6c7e98: feat (ai/core): add continuationSteps to generateText
+
 ## 3.4.1
 
 ### 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_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.
      */
@@ -1532,7 +1538,17 @@ By default, it's set to 1, which means that only a single LLM call is made.
      */
     maxSteps?: number;
     /**
-     * Optional telemetry configuration (experimental).
+@deprecated Use `experimental_continueSteps` instead.
+     */
+    experimental_continuationSteps?: boolean;
+    /**
+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;
     /**
@@ -1831,7 +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.
      */
@@ -1868,6 +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;
@@ -1895,7 +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_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.
      */
@@ -1532,7 +1538,17 @@ By default, it's set to 1, which means that only a single LLM call is made.
      */
     maxSteps?: number;
     /**
-     * Optional telemetry configuration (experimental).
+@deprecated Use `experimental_continueSteps` instead.
+     */
+    experimental_continuationSteps?: boolean;
+    /**
+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;
     /**
@@ -1831,7 +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.
      */
@@ -1868,6 +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;
@@ -1895,7 +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

@@ -3052,6 +3052,8 @@ async function generateText({
   maxAutomaticRoundtrips = 0,
   maxToolRoundtrips = maxAutomaticRoundtrips,
   maxSteps = maxToolRoundtrips != null ? maxToolRoundtrips + 1 : 1,
+  experimental_continuationSteps,
+  experimental_continueSteps: continueSteps = experimental_continuationSteps != null ? experimental_continuationSteps : false,
   experimental_telemetry: telemetry,
   experimental_providerMetadata: providerMetadata,
   _internal: {
@@ -3095,7 +3097,7 @@ async function generateText({
     }),
     tracer,
     fn: async (span) => {
-      var _a12, _b, _c, _d, _e;
+      var _a12, _b, _c, _d, _e, _f, _g;
       const retry = retryWithExponentialBackoff({ maxRetries });
       const validatedPrompt = validatePrompt({
         system,
@@ -3116,12 +3118,14 @@ async function generateText({
       let currentToolResults = [];
       let stepCount = 0;
       const responseMessages = [];
+      let text = "";
       const steps = [];
       const usage = {
         completionTokens: 0,
         promptTokens: 0,
         totalTokens: 0
       };
+      let stepType = "initial";
       do {
         const currentInputFormat = stepCount === 0 ? validatedPrompt.type : "messages";
         currentModelResponse = await retry(
@@ -3153,7 +3157,7 @@ async function generateText({
             }),
             tracer,
             fn: async (span2) => {
-              var _a13, _b2, _c2, _d2, _e2, _f;
+              var _a13, _b2, _c2, _d2, _e2, _f2;
               const result = await model.doGenerate({
                 mode,
                 ...callSettings,
@@ -3166,7 +3170,7 @@ 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: (_f = (_e2 = result.response) == null ? void 0 : _e2.modelId) != null ? _f : model.modelId
+                modelId: (_f2 = (_e2 = result.response) == null ? void 0 : _e2.modelId) != null ? _f2 : model.modelId
               };
               span2.setAttributes(
                 selectTelemetryAttributes({
@@ -3220,8 +3224,14 @@ async function generateText({
         usage.completionTokens += currentUsage.completionTokens;
         usage.promptTokens += currentUsage.promptTokens;
         usage.totalTokens += currentUsage.totalTokens;
+        if (stepType === "continue") {
+          text += (_b = currentModelResponse.text) != null ? _b : "";
+        } else {
+          text = (_c = currentModelResponse.text) != null ? _c : "";
+        }
         const currentStep = {
-          text: (_b = currentModelResponse.text) != null ? _b : "",
+          stepType,
+          text: (_d = currentModelResponse.text) != null ? _d : "",
           toolCalls: currentToolCalls,
           toolResults: currentToolResults,
           finishReason: currentModelResponse.finishReason,
@@ -3230,29 +3240,55 @@ async function generateText({
           logprobs: currentModelResponse.logprobs,
           response: {
             ...currentModelResponse.response,
-            headers: (_c = currentModelResponse.rawResponse) == null ? void 0 : _c.headers
+            headers: (_e = currentModelResponse.rawResponse) == null ? void 0 : _e.headers
           },
           experimental_providerMetadata: currentModelResponse.providerMetadata
         };
         steps.push(currentStep);
         await (onStepFinish == null ? void 0 : onStepFinish(currentStep));
-        const newResponseMessages = toResponseMessages({
-          text: currentModelResponse.text,
-          toolCalls: currentToolCalls,
-          toolResults: currentToolResults
-        });
-        responseMessages.push(...newResponseMessages);
-        promptMessages.push(
-          ...newResponseMessages.map(
-            (message) => convertToLanguageModelMessage(message, null)
-          )
-        );
-      } while (
-        // there are tool calls:
-        currentToolCalls.length > 0 && // all current tool calls have results:
-        currentToolResults.length === currentToolCalls.length && // the number of steps is less than the maximum:
-        ++stepCount < maxSteps
-      );
+        if (stepType === "continue") {
+          const lastResponseMessage = responseMessages.pop();
+          promptMessages.pop();
+          if (typeof lastResponseMessage.content === "string") {
+            lastResponseMessage.content = text;
+          } else {
+            lastResponseMessage.content.push({
+              text: (_f = currentModelResponse.text) != null ? _f : "",
+              type: "text"
+            });
+          }
+          responseMessages.push(lastResponseMessage);
+          promptMessages.push(
+            convertToLanguageModelMessage(lastResponseMessage, null)
+          );
+        } else {
+          const newResponseMessages = toResponseMessages({
+            text: currentModelResponse.text,
+            toolCalls: currentToolCalls,
+            toolResults: currentToolResults
+          });
+          responseMessages.push(...newResponseMessages);
+          promptMessages.push(
+            ...newResponseMessages.map(
+              (message) => convertToLanguageModelMessage(message, null)
+            )
+          );
+        }
+        if (++stepCount >= maxSteps) {
+          stepType = "done";
+        } else if (continueSteps && currentStep.finishReason === "length" && // only use continue when there are no tool calls:
+        currentToolCalls.length === 0) {
+          stepType = "continue";
+        } 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";
+        }
+      } while (stepType !== "done");
       span.setAttributes(
         selectTelemetryAttributes({
           telemetry,
@@ -3278,10 +3314,7 @@ async function generateText({
         })
       );
       return new DefaultGenerateTextResult({
-        // Always return a string so that the caller doesn't have to check for undefined.
-        // If they need to check if the model did not return any text,
-        // they can check the length of the string:
-        text: (_d = currentModelResponse.text) != null ? _d : "",
+        text,
         toolCalls: currentToolCalls,
         toolResults: currentToolResults,
         finishReason: currentModelResponse.finishReason,
@@ -3289,7 +3322,7 @@ async function generateText({
         warnings: currentModelResponse.warnings,
         response: {
           ...currentModelResponse.response,
-          headers: (_e = currentModelResponse.rawResponse) == null ? void 0 : _e.headers
+          headers: (_g = currentModelResponse.rawResponse) == null ? void 0 : _g.headers
         },
         logprobs: currentModelResponse.logprobs,
         responseMessages,
@@ -3735,6 +3768,7 @@ 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,
@@ -3879,6 +3913,7 @@ async function streamText({
         telemetry,
         startTimestampMs,
         maxSteps,
+        continueSteps,
         startStep,
         promptMessages,
         modelId: model.modelId,
@@ -3902,6 +3937,7 @@ var DefaultStreamTextResult = class {
     telemetry,
     startTimestampMs,
     maxSteps,
+    continueSteps,
     startStep,
     promptMessages,
     modelId,
@@ -3953,7 +3989,9 @@ var DefaultStreamTextResult = class {
         promptTokens: 0,
         completionTokens: 0,
         totalTokens: 0
-      }
+      },
+      stepType,
+      previousStepText = ""
     }) {
       const stepToolCalls = [];
       const stepToolResults = [];
@@ -3966,6 +4004,7 @@ var DefaultStreamTextResult = class {
       let stepProviderMetadata;
       let stepFirstChunk = true;
       let stepText = "";
+      let fullStepText = stepType === "continue" ? previousStepText : "";
       let stepLogProbs;
       let stepResponse = {
         id: generateId3(),
@@ -3999,6 +4038,7 @@ var DefaultStreamTextResult = class {
                 case "text-delta": {
                   controller.enqueue(chunk);
                   stepText += chunk.textDelta;
+                  fullStepText += chunk.textDelta;
                   await (onChunk == null ? void 0 : onChunk({ chunk }));
                   break;
                 }
@@ -4098,6 +4138,7 @@ var DefaultStreamTextResult = class {
                 response: stepResponse
               });
               const stepResult = {
+                stepType,
                 text: stepText,
                 toolCalls: stepToolCalls,
                 toolResults: stepToolResults,
@@ -4116,21 +4157,37 @@ 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({
+              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 (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,
@@ -4147,7 +4204,9 @@ var DefaultStreamTextResult = class {
                   doStreamSpan: doStreamSpan3,
                   currentStep: currentStep + 1,
                   promptMessages: promptMessages2,
-                  usage: combinedUsage
+                  usage: combinedUsage,
+                  stepType: nextStepType,
+                  previousStepText: fullStepText
                 });
                 return;
               }
@@ -4166,7 +4225,7 @@ var DefaultStreamTextResult = class {
                     telemetry,
                     attributes: {
                       "ai.response.finishReason": stepFinishReason,
-                      "ai.response.text": { output: () => stepText },
+                      "ai.response.text": { output: () => fullStepText },
                       "ai.response.toolCalls": {
                         output: () => stepToolCallsJson
                       },
@@ -4174,27 +4233,38 @@ 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);
@@ -4208,7 +4278,7 @@ 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
@@ -4241,7 +4311,8 @@ var DefaultStreamTextResult = class {
       doStreamSpan,
       currentStep: 0,
       promptMessages,
-      usage: void 0
+      usage: void 0,
+      stepType: "initial"
     });
   }
   /**