modelfusion 0.4.0 → 0.4.1

package/composed-function/summarize/summarizeRecursivelyWithTextGenerationAndTokenSplitting.cjs

@@ -9,17 +9,16 @@ const summarizeRecursively_js_1 = require("./summarizeRecursively.cjs");
  * It automatically splits the text into optimal chunks that are small enough to be processed by the model,
  * while leaving enough space for the model to generate text.
  */
-async function summarizeRecursivelyWithTextGenerationAndTokenSplitting({ text, model, prompt, reservedCompletionTokens, join, }, options) {
+async function summarizeRecursivelyWithTextGenerationAndTokenSplitting({ text, model, prompt, tokenLimit = model.contextWindowSize -
+    (model.maxCompletionTokens ?? model.contextWindowSize / 4), join, }, options) {
     const emptyPromptTokens = await model.countPromptTokens(await prompt({ text: "" }));
     return (0, summarizeRecursively_js_1.summarizeRecursively)({
         split: (0, splitRecursively_js_1.splitRecursivelyAtTokenAsSplitFunction)({
             tokenizer: model.tokenizer,
-            maxChunkSize: model.contextWindowSize -
-                reservedCompletionTokens -
-                emptyPromptTokens,
+            maxChunkSize: tokenLimit - emptyPromptTokens,
         }),
         summarize: async (input) => {
-            const { text } = await (0, generateText_js_1.generateText)(model.withMaxCompletionTokens(reservedCompletionTokens), await prompt(input), options);
+            const { text } = await (0, generateText_js_1.generateText)(model, await prompt(input), options);
             return text;
         },
         join,

package/composed-function/summarize/summarizeRecursivelyWithTextGenerationAndTokenSplitting.d.ts

@@ -6,7 +6,7 @@ import { Run } from "../../run/Run.js";
  * It automatically splits the text into optimal chunks that are small enough to be processed by the model,
  * while leaving enough space for the model to generate text.
  */
-export declare function summarizeRecursivelyWithTextGenerationAndTokenSplitting<PROMPT>({ text, model, prompt, reservedCompletionTokens, join, }: {
+export declare function summarizeRecursivelyWithTextGenerationAndTokenSplitting<PROMPT>({ text, model, prompt, tokenLimit, join, }: {
     text: string;
     model: TextGenerationModel<PROMPT, any, any, TextGenerationModelSettings> & {
         contextWindowSize: number;
@@ -16,7 +16,7 @@ export declare function summarizeRecursivelyWithTextGenerationAndTokenSplitting<
     prompt: (input: {
         text: string;
     }) => Promise<PROMPT>;
-    reservedCompletionTokens: number;
+    tokenLimit?: number;
     join?: (texts: Array<string>) => string;
 }, options?: {
     functionId?: string;

package/composed-function/summarize/summarizeRecursivelyWithTextGenerationAndTokenSplitting.js

@@ -6,17 +6,16 @@ import { summarizeRecursively } from "./summarizeRecursively.js";
  * It automatically splits the text into optimal chunks that are small enough to be processed by the model,
  * while leaving enough space for the model to generate text.
  */
-export async function summarizeRecursivelyWithTextGenerationAndTokenSplitting({ text, model, prompt, reservedCompletionTokens, join, }, options) {
+export async function summarizeRecursivelyWithTextGenerationAndTokenSplitting({ text, model, prompt, tokenLimit = model.contextWindowSize -
+    (model.maxCompletionTokens ?? model.contextWindowSize / 4), join, }, options) {
     const emptyPromptTokens = await model.countPromptTokens(await prompt({ text: "" }));
     return summarizeRecursively({
         split: splitRecursivelyAtTokenAsSplitFunction({
             tokenizer: model.tokenizer,
-            maxChunkSize: model.contextWindowSize -
-                reservedCompletionTokens -
-                emptyPromptTokens,
+            maxChunkSize: tokenLimit - emptyPromptTokens,
         }),
         summarize: async (input) => {
-            const { text } = await generateText(model.withMaxCompletionTokens(reservedCompletionTokens), await prompt(input), options);
+            const { text } = await generateText(model, await prompt(input), options);
             return text;
         },
         join,

package/model-function/generate-text/streamText.d.ts

@@ -1,7 +1,7 @@
 import { FunctionOptions } from "../FunctionOptions.js";
+import { CallMetadata } from "../executeCall.js";
 import { DeltaEvent } from "./DeltaEvent.js";
 import { TextGenerationModel, TextGenerationModelSettings } from "./TextGenerationModel.js";
-import { CallMetadata } from "model-function/executeCall.js";
 export declare function streamText<PROMPT, FULL_DELTA, SETTINGS extends TextGenerationModelSettings>(model: TextGenerationModel<PROMPT, unknown, FULL_DELTA, SETTINGS> & {
     generateDeltaStreamResponse: (prompt: PROMPT, options: FunctionOptions<SETTINGS>) => PromiseLike<AsyncIterable<DeltaEvent<FULL_DELTA>>>;
     extractTextDelta: (fullDelta: FULL_DELTA) => string | undefined;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "modelfusion",
   "description": "Build AI applications, chatbots, and agents with JavaScript and TypeScript.",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "author": "Lars Grammel",
   "license": "MIT",
   "keywords": [

package/prompt/chat/trimChatPrompt.cjs

@@ -12,7 +12,8 @@ const validateChatPrompt_js_1 = require("./validateChatPrompt.cjs");
  *
  * @see https://modelfusion.dev/guide/function/generate-text/prompt-mapping#limiting-the-chat-length
  */
-async function trimChatPrompt({ prompt, model, tokenLimit = model.contextWindowSize - model.maxCompletionTokens, }) {
+async function trimChatPrompt({ prompt, model, tokenLimit = model.contextWindowSize -
+    (model.maxCompletionTokens ?? model.contextWindowSize / 4), }) {
     (0, validateChatPrompt_js_1.validateChatPrompt)(prompt);
     const startsWithSystemMessage = "system" in prompt[0];
     const systemMessage = startsWithSystemMessage ? [prompt[0]] : [];

package/prompt/chat/trimChatPrompt.d.ts

@@ -1,4 +1,4 @@
-import { TextGenerationModel } from "model-function/generate-text/TextGenerationModel.js";
+import { TextGenerationModel } from "../../model-function/generate-text/TextGenerationModel.js";
 import { ChatPrompt } from "./ChatPrompt.js";
 /**
  * Keeps only the most recent messages in the prompt, while leaving enough space for the completion.
@@ -14,7 +14,6 @@ export declare function trimChatPrompt({ prompt, model, tokenLimit, }: {
     prompt: ChatPrompt;
     model: TextGenerationModel<ChatPrompt, any, any, any> & {
         contextWindowSize: number;
-        maxCompletionTokens: number;
         countPromptTokens: (prompt: ChatPrompt) => PromiseLike<number>;
     };
     tokenLimit?: number;

package/prompt/chat/trimChatPrompt.js

@@ -9,7 +9,8 @@ import { validateChatPrompt } from "./validateChatPrompt.js";
  *
  * @see https://modelfusion.dev/guide/function/generate-text/prompt-mapping#limiting-the-chat-length
  */
-export async function trimChatPrompt({ prompt, model, tokenLimit = model.contextWindowSize - model.maxCompletionTokens, }) {
+export async function trimChatPrompt({ prompt, model, tokenLimit = model.contextWindowSize -
+    (model.maxCompletionTokens ?? model.contextWindowSize / 4), }) {
     validateChatPrompt(prompt);
     const startsWithSystemMessage = "system" in prompt[0];
     const systemMessage = startsWithSystemMessage ? [prompt[0]] : [];