Create short-candles-drive.md

.changeset/short-candles-drive.md

@@ -0,0 +1,8 @@
+---
+"llamaindex": patch
+"@llamaindex/ollama": patch
+---
+
+feat: support ollama tool call
+
+Note that `OllamaEmbedding` now is not the subclass of `Ollama`.

e2e/fixtures/llm/ollama.ts

@@ -0,0 +1,3 @@
+import { OpenAI } from "./openai.js";
+
+export class Ollama extends OpenAI {}

e2e/node/ollama.e2e.ts

@@ -0,0 +1,35 @@
+import { Ollama } from "@llamaindex/ollama";
+import assert from "node:assert";
+import { test } from "node:test";
+import { getWeatherTool } from "./fixtures/tools.js";
+import { mockLLMEvent } from "./utils.js";
+
+await test("ollama", async (t) => {
+  await mockLLMEvent(t, "ollama");
+  await t.test("ollama function call", async (t) => {
+    const llm = new Ollama({
+      model: "llama3.2",
+    });
+    const chatResponse = await llm.chat({
+      messages: [
+        {
+          role: "user",
+          content: "What is the weather in Paris?",
+        },
+      ],
+      tools: [getWeatherTool],
+    });
+    if (
+      chatResponse.message.options &&
+      "toolCall" in chatResponse.message.options
+    ) {
+      assert.equal(chatResponse.message.options.toolCall.length, 1);
+      assert.equal(
+        chatResponse.message.options.toolCall[0]!.name,
+        getWeatherTool.metadata.name,
+      );
+    } else {
+      throw new Error("Expected tool calls in response");
+    }
+  });
+});

e2e/node/snapshot/ollama.snap

@@ -0,0 +1,5 @@
+{
+  "llmEventStart": [],
+  "llmEventEnd": [],
+  "llmEventStream": []
+}

packages/core/src/llms/base.ts

@@ -1,5 +1,4 @@
-import { streamConverter } from "../utils";
-import { extractText } from "../utils/llms";
+import { extractText, streamConverter } from "../utils";
 import type {
   ChatResponse,
   ChatResponseChunk,

packages/llamaindex/src/embeddings/OllamaEmbedding.ts

@@ -1,7 +1 @@
-import type { BaseEmbedding } from "@llamaindex/core/embeddings";
-import { Ollama } from "@llamaindex/ollama";
-
-/**
- * OllamaEmbedding is an alias for Ollama that implements the BaseEmbedding interface.
- */
-export class OllamaEmbedding extends Ollama implements BaseEmbedding {}
+export { OllamaEmbedding } from "@llamaindex/ollama";

packages/providers/ollama/src/index.ts

@@ -1,16 +1,20 @@
 import { BaseEmbedding } from "@llamaindex/core/embeddings";
-import type {
-  ChatResponse,
-  ChatResponseChunk,
-  CompletionResponse,
-  LLM,
-  LLMChatParamsNonStreaming,
-  LLMChatParamsStreaming,
-  LLMCompletionParamsNonStreaming,
-  LLMCompletionParamsStreaming,
-  LLMMetadata,
+import {
+  ToolCallLLM,
+  type BaseTool,
+  type ChatResponse,
+  type ChatResponseChunk,
+  type CompletionResponse,
+  type LLMChatParamsNonStreaming,
+  type LLMChatParamsStreaming,
+  type LLMCompletionParamsNonStreaming,
+  type LLMCompletionParamsStreaming,
+  type LLMMetadata,
+  type ToolCallLLMMessageOptions,
 } from "@llamaindex/core/llms";
 import { extractText, streamConverter } from "@llamaindex/core/utils";
+import { randomUUID } from "@llamaindex/env";
+import type { ChatRequest, GenerateRequest, Tool } from "ollama";
 import {
   Ollama as OllamaBase,
   type Config,
@@ -38,7 +42,8 @@ export type OllamaParams = {
   options?: Partial<Options>;
 };
 
-export class Ollama extends BaseEmbedding implements LLM {
+export class Ollama extends ToolCallLLM {
+  supportToolCall: boolean = true;
   public readonly ollama: OllamaBase;
 
   // https://ollama.ai/library
@@ -78,12 +83,16 @@ export class Ollama extends BaseEmbedding implements LLM {
   chat(
     params: LLMChatParamsStreaming,
   ): Promise<AsyncIterable<ChatResponseChunk>>;
-  chat(params: LLMChatParamsNonStreaming): Promise<ChatResponse>;
+  chat(
+    params: LLMChatParamsNonStreaming,
+  ): Promise<ChatResponse<ToolCallLLMMessageOptions>>;
   async chat(
     params: LLMChatParamsNonStreaming | LLMChatParamsStreaming,
-  ): Promise<ChatResponse | AsyncIterable<ChatResponseChunk>> {
-    const { messages, stream } = params;
-    const payload = {
+  ): Promise<
+    ChatResponse<ToolCallLLMMessageOptions> | AsyncIterable<ChatResponseChunk>
+  > {
+    const { messages, stream, tools } = params;
+    const payload: ChatRequest = {
       model: this.model,
       messages: messages.map((message) => ({
         role: message.role,
@@ -94,11 +103,30 @@ export class Ollama extends BaseEmbedding implements LLM {
         ...this.options,
       },
     };
+    if (tools) {
+      payload.tools = tools.map((tool) => Ollama.toTool(tool));
+    }
     if (!stream) {
       const chatResponse = await this.ollama.chat({
         ...payload,
         stream: false,
       });
+      if (chatResponse.message.tool_calls) {
+        return {
+          message: {
+            role: "assistant",
+            content: chatResponse.message.content,
+            options: {
+              toolCall: chatResponse.message.tool_calls.map((toolCall) => ({
+                name: toolCall.function.name,
+                input: toolCall.function.arguments,
+                id: randomUUID(),
+              })),
+            },
+          },
+          raw: chatResponse,
+        };
+      }
 
       return {
         message: {
@@ -126,7 +154,7 @@ export class Ollama extends BaseEmbedding implements LLM {
     params: LLMCompletionParamsStreaming | LLMCompletionParamsNonStreaming,
   ): Promise<CompletionResponse | AsyncIterable<CompletionResponse>> {
     const { prompt, stream } = params;
-    const payload = {
+    const payload: GenerateRequest = {
       model: this.model,
       prompt: extractText(prompt),
       stream: !!stream,
@@ -152,15 +180,39 @@ export class Ollama extends BaseEmbedding implements LLM {
     }
   }
 
-  private async getEmbedding(prompt: string): Promise<number[]> {
-    const payload = {
-      model: this.model,
-      prompt,
-      options: {
-        ...this.options,
+  static toTool(tool: BaseTool): Tool {
+    return {
+      type: "function",
+      function: {
+        name: tool.metadata.name,
+        description: tool.metadata.description,
+        parameters: {
+          type: tool.metadata.parameters?.type,
+          required: tool.metadata.parameters?.required,
+          properties: tool.metadata.parameters?.properties,
+        },
       },
     };
-    const response = await this.ollama.embeddings({
+  }
+}
+
+export class OllamaEmbedding extends BaseEmbedding {
+  private readonly llm: Ollama;
+
+  constructor(params: OllamaParams) {
+    super();
+    this.llm = new Ollama(params);
+  }
+
+  private async getEmbedding(prompt: string): Promise<number[]> {
+    const payload = {
+      model: this.llm.model,
+      prompt,
+      options: {
+        ...this.llm.options,
+      },
+    };
+    const response = await this.llm.ollama.embeddings({
       ...payload,
     });
     return response.embedding;