ai 3.4.1 → 3.4.2

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # ai
 
+## 3.4.2
+
+### Patch Changes
+
+- e6c7e98: feat (ai/core): add continuationSteps to generateText
+
 ## 3.4.1
 
 ### Patch Changes

package/dist/index.d.mts

@@ -1491,7 +1491,7 @@ 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: continuationSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1531,6 +1531,7 @@ 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;
+    experimental_continuationSteps?: boolean;
     /**
      * Optional telemetry configuration (experimental).
      */

package/dist/index.d.ts

@@ -1491,7 +1491,7 @@ 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: continuationSteps, experimental_telemetry: telemetry, experimental_providerMetadata: providerMetadata, _internal: { generateId, currentDate, }, onStepFinish, ...settings }: CallSettings & Prompt & {
     /**
 The language model to use.
      */
@@ -1531,6 +1531,7 @@ 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;
+    experimental_continuationSteps?: boolean;
     /**
      * Optional telemetry configuration (experimental).
      */

package/dist/index.js

@@ -3052,6 +3052,7 @@ async function generateText({
   maxAutomaticRoundtrips = 0,
   maxToolRoundtrips = maxAutomaticRoundtrips,
   maxSteps = maxToolRoundtrips != null ? maxToolRoundtrips + 1 : 1,
+  experimental_continuationSteps: continuationSteps = false,
   experimental_telemetry: telemetry,
   experimental_providerMetadata: providerMetadata,
   _internal: {
@@ -3095,7 +3096,7 @@ async function generateText({
     }),
     tracer,
     fn: async (span) => {
-      var _a12, _b, _c, _d, _e;
+      var _a12, _b, _c, _d, _e, _f;
       const retry = retryWithExponentialBackoff({ maxRetries });
       const validatedPrompt = validatePrompt({
         system,
@@ -3116,12 +3117,14 @@ 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 +3156,7 @@ 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 +3169,7 @@ 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 +3223,13 @@ 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 : "";
+        }
         const currentStep = {
-          text: (_b = currentModelResponse.text) != null ? _b : "",
+          text: (_d = currentModelResponse.text) != null ? _d : "",
           toolCalls: currentToolCalls,
           toolResults: currentToolResults,
           finishReason: currentModelResponse.finishReason,
@@ -3230,29 +3238,55 @@ 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 === "continuation") {
+          const lastResponseMessage = responseMessages.pop();
+          promptMessages.pop();
+          if (typeof lastResponseMessage.content === "string") {
+            lastResponseMessage.content = text;
+          } else {
+            lastResponseMessage.content.push({
+              text: " " + currentModelResponse.text,
+              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 (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";
+        }
+      } while (stepType !== "done");
       span.setAttributes(
         selectTelemetryAttributes({
           telemetry,
@@ -3278,10 +3312,7 @@ 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 +3320,7 @@ async function generateText({
         warnings: currentModelResponse.warnings,
         response: {
           ...currentModelResponse.response,
-          headers: (_e = currentModelResponse.rawResponse) == null ? void 0 : _e.headers
+          headers: (_f = currentModelResponse.rawResponse) == null ? void 0 : _f.headers
         },
         logprobs: currentModelResponse.logprobs,
         responseMessages,