modelfusion 0.93.1 → 0.95.0

package/README.md

@@ -581,6 +581,7 @@ const textStream = await streamText(
 | Anthropic       | ✅          | ✅                 | ✅          |
 | Llama 2         | ✅          | ✅                 | ✅          |
 | ChatML          | ✅          | ✅                 | ✅          |
+| NeuralChat      | ✅          | ✅                 | ✅          |
 | Alpaca          | ✅          | ✅                 | ❌          |
 | Vicuna          | ❌          | ❌                 | ✅          |
 | Generic Text    | ✅          | ✅                 | ✅          |

package/model-function/generate-text/prompt-template/ChatMLPromptTemplate.cjs

@@ -4,11 +4,11 @@ exports.chat = exports.instruction = exports.text = void 0;
 const ChatPrompt_js_1 = require("./ChatPrompt.cjs");
 const START_SEGMENT = "<|im_start|>";
 const END_SEGMENT = "<|im_end|>";
-function chatMLStart(role) {
+function segmentStart(role) {
     return `${START_SEGMENT}${role}\n`;
 }
-function chatMLSegment(role, text) {
-    return text == null ? "" : `${chatMLStart(role)}${text}${END_SEGMENT}\n`;
+function segment(role, text) {
+    return text == null ? "" : `${segmentStart(role)}${text}${END_SEGMENT}\n`;
 }
 /**
  * Formats a text prompt using the ChatML format.
@@ -18,7 +18,7 @@ function text() {
         stopSequences: [END_SEGMENT],
         format(prompt) {
             // prompt and then prefix start of assistant response:
-            return chatMLSegment("user", prompt) + chatMLStart("assistant");
+            return segment("user", prompt) + segmentStart("assistant");
         },
     };
 }
@@ -40,9 +40,9 @@ function instruction() {
     return {
         stopSequences: [END_SEGMENT],
         format(prompt) {
-            return (chatMLSegment("system", prompt.system) +
-                chatMLSegment("user", prompt.instruction) +
-                chatMLStart("assistant") +
+            return (segment("system", prompt.system) +
+                segment("user", prompt.instruction) +
+                segmentStart("assistant") +
                 (prompt.responsePrefix ?? ""));
         },
     };
@@ -65,15 +65,15 @@ function chat() {
     return {
         format(prompt) {
             (0, ChatPrompt_js_1.validateChatPrompt)(prompt);
-            let text = prompt.system != null ? chatMLSegment("system", prompt.system) : "";
+            let text = prompt.system != null ? segment("system", prompt.system) : "";
             for (const { role, content } of prompt.messages) {
                 switch (role) {
                     case "user": {
-                        text += chatMLSegment("user", content);
+                        text += segment("user", content);
                         break;
                     }
                     case "assistant": {
-                        text += chatMLSegment("assistant", content);
+                        text += segment("assistant", content);
                         break;
                     }
                     default: {
@@ -83,7 +83,7 @@ function chat() {
                 }
             }
             // prefix start of assistant response:
-            text += chatMLStart("assistant");
+            text += segmentStart("assistant");
             return text;
         },
         stopSequences: [END_SEGMENT],

package/model-function/generate-text/prompt-template/ChatMLPromptTemplate.js

@@ -1,11 +1,11 @@
 import { validateChatPrompt } from "./ChatPrompt.js";
 const START_SEGMENT = "<|im_start|>";
 const END_SEGMENT = "<|im_end|>";
-function chatMLStart(role) {
+function segmentStart(role) {
     return `${START_SEGMENT}${role}\n`;
 }
-function chatMLSegment(role, text) {
-    return text == null ? "" : `${chatMLStart(role)}${text}${END_SEGMENT}\n`;
+function segment(role, text) {
+    return text == null ? "" : `${segmentStart(role)}${text}${END_SEGMENT}\n`;
 }
 /**
  * Formats a text prompt using the ChatML format.
@@ -15,7 +15,7 @@ export function text() {
         stopSequences: [END_SEGMENT],
         format(prompt) {
             // prompt and then prefix start of assistant response:
-            return chatMLSegment("user", prompt) + chatMLStart("assistant");
+            return segment("user", prompt) + segmentStart("assistant");
         },
     };
 }
@@ -36,9 +36,9 @@ export function instruction() {
     return {
         stopSequences: [END_SEGMENT],
         format(prompt) {
-            return (chatMLSegment("system", prompt.system) +
-                chatMLSegment("user", prompt.instruction) +
-                chatMLStart("assistant") +
+            return (segment("system", prompt.system) +
+                segment("user", prompt.instruction) +
+                segmentStart("assistant") +
                 (prompt.responsePrefix ?? ""));
         },
     };
@@ -60,15 +60,15 @@ export function chat() {
     return {
         format(prompt) {
             validateChatPrompt(prompt);
-            let text = prompt.system != null ? chatMLSegment("system", prompt.system) : "";
+            let text = prompt.system != null ? segment("system", prompt.system) : "";
             for (const { role, content } of prompt.messages) {
                 switch (role) {
                     case "user": {
-                        text += chatMLSegment("user", content);
+                        text += segment("user", content);
                         break;
                     }
                     case "assistant": {
-                        text += chatMLSegment("assistant", content);
+                        text += segment("assistant", content);
                         break;
                     }
                     default: {
@@ -78,7 +78,7 @@ export function chat() {
                 }
             }
             // prefix start of assistant response:
-            text += chatMLStart("assistant");
+            text += segmentStart("assistant");
             return text;
         },
         stopSequences: [END_SEGMENT],

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.cjs

@@ -0,0 +1,81 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.chat = exports.instruction = exports.text = void 0;
+const ChatPrompt_js_1 = require("./ChatPrompt.cjs");
+const roleNames = {
+    system: "System",
+    user: "User",
+    assistant: "Assistant",
+};
+function segmentStart(role) {
+    return `### ${roleNames[role]}:\n`;
+}
+function segment(role, text) {
+    return text == null ? "" : `${segmentStart(role)}${text}\n`;
+}
+/**
+ * Formats a text prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+function text() {
+    return {
+        stopSequences: [],
+        format(prompt) {
+            // prompt and then prefix start of assistant response:
+            return segment("user", prompt) + segmentStart("assistant");
+        },
+    };
+}
+exports.text = text;
+/**
+ * Formats an instruction prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+const instruction = () => ({
+    stopSequences: [],
+    format(prompt) {
+        return (segment("system", prompt.system) +
+            segment("user", prompt.instruction) +
+            segmentStart("assistant") +
+            (prompt.responsePrefix ?? ""));
+    },
+});
+exports.instruction = instruction;
+/**
+ * Formats a chat prompt as a basic text prompt.
+ *
+ * @param user The label of the user in the chat. Default to "user".
+ * @param assistant The label of the assistant in the chat. Default to "assistant".
+ * @param system The label of the system in the chat. Optional, defaults to no prefix.
+ */
+function chat() {
+    return {
+        format(prompt) {
+            (0, ChatPrompt_js_1.validateChatPrompt)(prompt);
+            let text = prompt.system != null ? segment("system", prompt.system) : "";
+            for (const { role, content } of prompt.messages) {
+                switch (role) {
+                    case "user": {
+                        text += segment("user", content);
+                        break;
+                    }
+                    case "assistant": {
+                        text += segment("assistant", content);
+                        break;
+                    }
+                    default: {
+                        const _exhaustiveCheck = role;
+                        throw new Error(`Unsupported role: ${_exhaustiveCheck}`);
+                    }
+                }
+            }
+            // prefix start of assistant response:
+            text += segmentStart("assistant");
+            return text;
+        },
+        stopSequences: [`\n${roleNames.user}:`],
+    };
+}
+exports.chat = chat;

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.d.ts

@@ -0,0 +1,23 @@
+import { TextGenerationPromptTemplate } from "../TextGenerationPromptTemplate.js";
+import { TextChatPrompt } from "./ChatPrompt.js";
+import { TextInstructionPrompt } from "./InstructionPrompt.js";
+/**
+ * Formats a text prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+export declare function text(): TextGenerationPromptTemplate<string, string>;
+/**
+ * Formats an instruction prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+export declare const instruction: () => TextGenerationPromptTemplate<TextInstructionPrompt, string>;
+/**
+ * Formats a chat prompt as a basic text prompt.
+ *
+ * @param user The label of the user in the chat. Default to "user".
+ * @param assistant The label of the assistant in the chat. Default to "assistant".
+ * @param system The label of the system in the chat. Optional, defaults to no prefix.
+ */
+export declare function chat(): TextGenerationPromptTemplate<TextChatPrompt, string>;

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.js

@@ -0,0 +1,75 @@
+import { validateChatPrompt } from "./ChatPrompt.js";
+const roleNames = {
+    system: "System",
+    user: "User",
+    assistant: "Assistant",
+};
+function segmentStart(role) {
+    return `### ${roleNames[role]}:\n`;
+}
+function segment(role, text) {
+    return text == null ? "" : `${segmentStart(role)}${text}\n`;
+}
+/**
+ * Formats a text prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+export function text() {
+    return {
+        stopSequences: [],
+        format(prompt) {
+            // prompt and then prefix start of assistant response:
+            return segment("user", prompt) + segmentStart("assistant");
+        },
+    };
+}
+/**
+ * Formats an instruction prompt as a neural chat prompt.
+ *
+ * @see https://huggingface.co/Intel/neural-chat-7b-v3-1#prompt-template
+ */
+export const instruction = () => ({
+    stopSequences: [],
+    format(prompt) {
+        return (segment("system", prompt.system) +
+            segment("user", prompt.instruction) +
+            segmentStart("assistant") +
+            (prompt.responsePrefix ?? ""));
+    },
+});
+/**
+ * Formats a chat prompt as a basic text prompt.
+ *
+ * @param user The label of the user in the chat. Default to "user".
+ * @param assistant The label of the assistant in the chat. Default to "assistant".
+ * @param system The label of the system in the chat. Optional, defaults to no prefix.
+ */
+export function chat() {
+    return {
+        format(prompt) {
+            validateChatPrompt(prompt);
+            let text = prompt.system != null ? segment("system", prompt.system) : "";
+            for (const { role, content } of prompt.messages) {
+                switch (role) {
+                    case "user": {
+                        text += segment("user", content);
+                        break;
+                    }
+                    case "assistant": {
+                        text += segment("assistant", content);
+                        break;
+                    }
+                    default: {
+                        const _exhaustiveCheck = role;
+                        throw new Error(`Unsupported role: ${_exhaustiveCheck}`);
+                    }
+                }
+            }
+            // prefix start of assistant response:
+            text += segmentStart("assistant");
+            return text;
+        },
+        stopSequences: [`\n${roleNames.user}:`],
+    };
+}

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.test.cjs

@@ -0,0 +1,49 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+const NeuralChatPromptTemplate_js_1 = require("./NeuralChatPromptTemplate.cjs");
+describe("text prompt", () => {
+    it("should format prompt", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.text)().format("prompt");
+        expect(prompt).toMatchSnapshot();
+    });
+});
+describe("instruction prompt", () => {
+    it("should format prompt with instruction", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.instruction)().format({
+            instruction: "instruction",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with system and instruction", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.instruction)().format({
+            system: "system",
+            instruction: "instruction",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with instruction and response prefix", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.instruction)().format({
+            instruction: "instruction",
+            responsePrefix: "response prefix",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+});
+describe("chat prompt", () => {
+    it("should format prompt with user message", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.chat)().format({
+            messages: [{ role: "user", content: "user message" }],
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with user-assistant-user messages", () => {
+        const prompt = (0, NeuralChatPromptTemplate_js_1.chat)().format({
+            messages: [
+                { role: "user", content: "1st user message" },
+                { role: "assistant", content: "assistant message" },
+                { role: "user", content: "2nd user message" },
+            ],
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+});

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.test.d.ts

@@ -0,0 +1 @@
+export {};

package/model-function/generate-text/prompt-template/NeuralChatPromptTemplate.test.js

@@ -0,0 +1,47 @@
+import { chat, instruction, text } from "./NeuralChatPromptTemplate.js";
+describe("text prompt", () => {
+    it("should format prompt", () => {
+        const prompt = text().format("prompt");
+        expect(prompt).toMatchSnapshot();
+    });
+});
+describe("instruction prompt", () => {
+    it("should format prompt with instruction", () => {
+        const prompt = instruction().format({
+            instruction: "instruction",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with system and instruction", () => {
+        const prompt = instruction().format({
+            system: "system",
+            instruction: "instruction",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with instruction and response prefix", () => {
+        const prompt = instruction().format({
+            instruction: "instruction",
+            responsePrefix: "response prefix",
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+});
+describe("chat prompt", () => {
+    it("should format prompt with user message", () => {
+        const prompt = chat().format({
+            messages: [{ role: "user", content: "user message" }],
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+    it("should format prompt with user-assistant-user messages", () => {
+        const prompt = chat().format({
+            messages: [
+                { role: "user", content: "1st user message" },
+                { role: "assistant", content: "assistant message" },
+                { role: "user", content: "2nd user message" },
+            ],
+        });
+        expect(prompt).toMatchSnapshot();
+    });
+});

package/model-function/generate-text/prompt-template/TextPromptTemplate.cjs

@@ -20,9 +20,9 @@ const instruction = () => ({
         if (prompt.system != null) {
             text += `${prompt.system}\n\n`;
         }
-        text += prompt.instruction;
+        text += `${prompt.instruction}\n\n`;
         if (prompt.responsePrefix != null) {
-            text += `\n\n${prompt.responsePrefix}`;
+            text += prompt.responsePrefix;
         }
         return text;
     },

package/model-function/generate-text/prompt-template/TextPromptTemplate.js

@@ -16,9 +16,9 @@ export const instruction = () => ({
         if (prompt.system != null) {
             text += `${prompt.system}\n\n`;
         }
-        text += prompt.instruction;
+        text += `${prompt.instruction}\n\n`;
         if (prompt.responsePrefix != null) {
-            text += `\n\n${prompt.responsePrefix}`;
+            text += prompt.responsePrefix;
         }
         return text;
     },

package/model-function/generate-text/prompt-template/index.cjs

@@ -26,14 +26,15 @@ var __exportStar = (this && this.__exportStar) || function(m, exports) {
     for (var p in m) if (p !== "default" && !Object.prototype.hasOwnProperty.call(exports, p)) __createBinding(exports, m, p);
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.VicunaPrompt = exports.TextPrompt = exports.Llama2Prompt = exports.ChatMLPrompt = exports.AlpacaPrompt = void 0;
+exports.VicunaPrompt = exports.TextPrompt = exports.NeuralChatPrompt = exports.Llama2Prompt = exports.ChatMLPrompt = exports.AlpacaPrompt = void 0;
 exports.AlpacaPrompt = __importStar(require("./AlpacaPromptTemplate.cjs"));
 exports.ChatMLPrompt = __importStar(require("./ChatMLPromptTemplate.cjs"));
 __exportStar(require("./ChatPrompt.cjs"), exports);
 __exportStar(require("./Content.cjs"), exports);
 __exportStar(require("./InstructionPrompt.cjs"), exports);
-exports.Llama2Prompt = __importStar(require("./Llama2PromptTemplate.cjs"));
 __exportStar(require("./InvalidPromptError.cjs"), exports);
+exports.Llama2Prompt = __importStar(require("./Llama2PromptTemplate.cjs"));
+exports.NeuralChatPrompt = __importStar(require("./NeuralChatPromptTemplate.cjs"));
 exports.TextPrompt = __importStar(require("./TextPromptTemplate.cjs"));
 exports.VicunaPrompt = __importStar(require("./VicunaPromptTemplate.cjs"));
 __exportStar(require("./trimChatPrompt.cjs"), exports);

package/model-function/generate-text/prompt-template/index.d.ts

@@ -3,8 +3,9 @@ export * as ChatMLPrompt from "./ChatMLPromptTemplate.js";
 export * from "./ChatPrompt.js";
 export * from "./Content.js";
 export * from "./InstructionPrompt.js";
-export * as Llama2Prompt from "./Llama2PromptTemplate.js";
 export * from "./InvalidPromptError.js";
+export * as Llama2Prompt from "./Llama2PromptTemplate.js";
+export * as NeuralChatPrompt from "./NeuralChatPromptTemplate.js";
 export * as TextPrompt from "./TextPromptTemplate.js";
 export * as VicunaPrompt from "./VicunaPromptTemplate.js";
 export * from "./trimChatPrompt.js";

package/model-function/generate-text/prompt-template/index.js

@@ -3,8 +3,9 @@ export * as ChatMLPrompt from "./ChatMLPromptTemplate.js";
 export * from "./ChatPrompt.js";
 export * from "./Content.js";
 export * from "./InstructionPrompt.js";
-export * as Llama2Prompt from "./Llama2PromptTemplate.js";
 export * from "./InvalidPromptError.js";
+export * as Llama2Prompt from "./Llama2PromptTemplate.js";
+export * as NeuralChatPrompt from "./NeuralChatPromptTemplate.js";
 export * as TextPrompt from "./TextPromptTemplate.js";
 export * as VicunaPrompt from "./VicunaPromptTemplate.js";
 export * from "./trimChatPrompt.js";

package/model-provider/llamacpp/LlamaCppTextGenerationModel.cjs

@@ -57,6 +57,7 @@ class LlamaCppTextGenerationModel extends AbstractModel_js_1.AbstractModel {
             "maxCompletionTokens",
             "stopSequences",
             "contextWindowSize",
+            "cachePrompt",
             "temperature",
             "topK",
             "topP",
@@ -197,13 +198,14 @@ const llamaCppTextStreamingResponseSchema = new ZodSchema_js_1.ZodSchema(zod_1.z
     }),
     llamaCppTextGenerationResponseSchema,
 ]));
-async function callLlamaCppTextGenerationAPI({ api = new LlamaCppApiConfiguration_js_1.LlamaCppApiConfiguration(), abortSignal, responseFormat, prompt, temperature, topK, topP, nPredict, nKeep, stop, tfsZ, typicalP, repeatPenalty, repeatLastN, penalizeNl, mirostat, mirostatTau, mirostatEta, seed, ignoreEos, logitBias, }) {
+async function callLlamaCppTextGenerationAPI({ api = new LlamaCppApiConfiguration_js_1.LlamaCppApiConfiguration(), abortSignal, responseFormat, prompt, cachePrompt, temperature, topK, topP, nPredict, nKeep, stop, tfsZ, typicalP, repeatPenalty, repeatLastN, penalizeNl, mirostat, mirostatTau, mirostatEta, seed, ignoreEos, logitBias, }) {
     return (0, postToApi_js_1.postJsonToApi)({
         url: api.assembleUrl(`/completion`),
         headers: api.headers,
         body: {
             stream: responseFormat.stream,
             prompt: prompt.text,
+            cache_prompt: cachePrompt,
             temperature,
             top_k: topK,
             top_p: topP,

package/model-provider/llamacpp/LlamaCppTextGenerationModel.d.ts

@@ -15,6 +15,10 @@ export interface LlamaCppTextGenerationModelSettings<CONTEXT_WINDOW_SIZE extends
      * Llama.cpp server.
      */
     contextWindowSize?: CONTEXT_WINDOW_SIZE;
+    /**
+     * Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+     */
+    cachePrompt?: boolean;
     temperature?: number;
     topK?: number;
     topP?: number;

package/model-provider/llamacpp/LlamaCppTextGenerationModel.js

@@ -54,6 +54,7 @@ export class LlamaCppTextGenerationModel extends AbstractModel {
             "maxCompletionTokens",
             "stopSequences",
             "contextWindowSize",
+            "cachePrompt",
             "temperature",
             "topK",
             "topP",
@@ -193,13 +194,14 @@ const llamaCppTextStreamingResponseSchema = new ZodSchema(z.discriminatedUnion("
     }),
     llamaCppTextGenerationResponseSchema,
 ]));
-async function callLlamaCppTextGenerationAPI({ api = new LlamaCppApiConfiguration(), abortSignal, responseFormat, prompt, temperature, topK, topP, nPredict, nKeep, stop, tfsZ, typicalP, repeatPenalty, repeatLastN, penalizeNl, mirostat, mirostatTau, mirostatEta, seed, ignoreEos, logitBias, }) {
+async function callLlamaCppTextGenerationAPI({ api = new LlamaCppApiConfiguration(), abortSignal, responseFormat, prompt, cachePrompt, temperature, topK, topP, nPredict, nKeep, stop, tfsZ, typicalP, repeatPenalty, repeatLastN, penalizeNl, mirostat, mirostatTau, mirostatEta, seed, ignoreEos, logitBias, }) {
     return postJsonToApi({
         url: api.assembleUrl(`/completion`),
         headers: api.headers,
         body: {
             stream: responseFormat.stream,
             prompt: prompt.text,
+            cache_prompt: cachePrompt,
             temperature,
             top_k: topK,
             top_p: topP,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "modelfusion",
   "description": "The TypeScript library for building multi-modal AI applications.",
-  "version": "0.93.1",
+  "version": "0.95.0",
   "author": "Lars Grammel",
   "license": "MIT",
   "keywords": [