@dxos/assistant 0.8.1-main.ba2dec9 → 0.8.1-staging.31c3ee1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dxos/assistant",
-  "version": "0.8.1-main.ba2dec9",
+  "version": "0.8.1-staging.31c3ee1",
   "description": "Natural language interface to DXOS",
   "homepage": "https://dxos.org",
   "bugs": "https://github.com/dxos/dxos/issues",
@@ -40,18 +40,18 @@
     "parjs": "^1.3.9",
     "parsimmon": "^1.18.1",
     "ulidx": "^2.3.0",
-    "@dxos/context": "0.8.1-main.ba2dec9",
-    "@dxos/artifact": "0.8.1-main.ba2dec9",
-    "@dxos/async": "0.8.1-main.ba2dec9",
-    "@dxos/debug": "0.8.1-main.ba2dec9",
-    "@dxos/echo-db": "0.8.1-main.ba2dec9",
-    "@dxos/echo-schema": "0.8.1-main.ba2dec9",
-    "@dxos/invariant": "0.8.1-main.ba2dec9",
-    "@dxos/live-object": "0.8.1-main.ba2dec9",
-    "@dxos/keys": "0.8.1-main.ba2dec9",
-    "@dxos/log": "0.8.1-main.ba2dec9",
-    "@dxos/util": "0.8.1-main.ba2dec9",
-    "@dxos/node-std": "0.8.1-main.ba2dec9"
+    "@dxos/artifact": "0.8.1-staging.31c3ee1",
+    "@dxos/async": "0.8.1-staging.31c3ee1",
+    "@dxos/echo-db": "0.8.1-staging.31c3ee1",
+    "@dxos/debug": "0.8.1-staging.31c3ee1",
+    "@dxos/echo-schema": "0.8.1-staging.31c3ee1",
+    "@dxos/invariant": "0.8.1-staging.31c3ee1",
+    "@dxos/context": "0.8.1-staging.31c3ee1",
+    "@dxos/keys": "0.8.1-staging.31c3ee1",
+    "@dxos/live-object": "0.8.1-staging.31c3ee1",
+    "@dxos/log": "0.8.1-staging.31c3ee1",
+    "@dxos/node-std": "0.8.1-staging.31c3ee1",
+    "@dxos/util": "0.8.1-staging.31c3ee1"
   },
   "devDependencies": {
     "@types/parsimmon": "^1.10.9",

package/src/ai-service/client.test.ts

@@ -11,12 +11,12 @@ import { invariant } from '@dxos/invariant';
 import { SpaceId } from '@dxos/keys';
 import { log } from '@dxos/log';
 
-import { DEFAULT_LLM_MODEL } from './defs';
+import { DEFAULT_EDGE_MODEL } from './defs';
 import { AIServiceEdgeClient } from './edge-client';
 import { OllamaClient } from './ollama-client';
 import { MixedStreamParser } from './parser';
 import { ToolTypes } from './types';
-import { AI_SERVICE_ENDPOINT } from '../testing';
+import { createTestOllamaClient, AI_SERVICE_ENDPOINT } from '../testing';
 
 // log.config({ filter: 'debug' });
 
@@ -40,7 +40,7 @@ describe.skip('AI Service Client', () => {
     ]);
 
     const stream = await client.exec({
-      model: DEFAULT_LLM_MODEL,
+      model: DEFAULT_EDGE_MODEL,
       spaceId,
       threadId,
       systemPrompt: 'You are a poet',
@@ -85,7 +85,7 @@ describe.skip('AI Service Client', () => {
 
     {
       const stream1 = await client.exec({
-        model: DEFAULT_LLM_MODEL,
+        model: DEFAULT_EDGE_MODEL,
         spaceId,
         threadId,
         systemPrompt: 'You are a helpful assistant.',
@@ -119,7 +119,7 @@ describe.skip('AI Service Client', () => {
 
     {
       const stream2 = await client.exec({
-        model: DEFAULT_LLM_MODEL,
+        model: DEFAULT_EDGE_MODEL,
         spaceId,
         threadId,
         systemPrompt: 'You are a helpful assistant.',
@@ -157,7 +157,7 @@ describe.skip('AI Service Client', () => {
     ]);
 
     const stream = await client.exec({
-      model: DEFAULT_LLM_MODEL,
+      model: DEFAULT_EDGE_MODEL,
       spaceId,
       threadId,
       tools: [
@@ -179,14 +179,14 @@ describe.skip('AI Service Client', () => {
   });
 });
 
-describe('Ollama Client', () => {
+describe.skip('Ollama Client', () => {
   test('basic', async (ctx) => {
     const isRunning = await OllamaClient.isRunning();
     if (!isRunning) {
       ctx.skip();
     }
 
-    const client = OllamaClient.createClient();
+    const client = createTestOllamaClient();
     const parser = new MixedStreamParser();
 
     const messages = await parser.parse(
@@ -198,7 +198,7 @@ describe('Ollama Client', () => {
       }),
     );
 
-    log.info('messages', { messages });
+    log('messages', { messages });
   });
 
   test('tool calls', async (ctx) => {
@@ -207,7 +207,7 @@ describe('Ollama Client', () => {
       ctx.skip();
     }
 
-    const client = OllamaClient.createClient({
+    const client = createTestOllamaClient({
       tools: [
         defineTool('test', {
           name: 'encrypt',
@@ -221,7 +221,7 @@ describe('Ollama Client', () => {
     });
     const parser = new MixedStreamParser();
     parser.streamEvent.on((event) => {
-      // log.info('event', { event });
+      // log('event', { event });
     });
 
     const messages = await parser.parse(
@@ -233,7 +233,7 @@ describe('Ollama Client', () => {
       }),
     );
 
-    log.info('messages', { messages });
+    log('messages', { messages });
   });
 
   test('text-to-image', async (ctx) => {
@@ -242,7 +242,7 @@ describe('Ollama Client', () => {
       ctx.skip();
     }
 
-    const client = OllamaClient.createClient();
+    const client = createTestOllamaClient();
     const parser = new MixedStreamParser();
 
     const messages = await parser.parse(
@@ -260,6 +260,6 @@ describe('Ollama Client', () => {
       }),
     );
 
-    log.info('messages', { messages });
+    log('messages', { messages });
   });
 });

package/src/ai-service/defs.ts

@@ -5,9 +5,9 @@
 // TODO(burdon): Share with Edge.
 // TODO(burdon): Move to service config retrieved by HTTP?
 
-export const DEFAULT_LLM_MODEL = '@anthropic/claude-3-5-sonnet-20241022';
+export const DEFAULT_EDGE_MODEL = '@anthropic/claude-3-5-sonnet-20241022';
 
-export const DEFAULT_LLM_MODELS = [
+export const DEFAULT_EDGE_MODELS = [
   // AI Gateway.
   // https://developers.cloudflare.com/ai-gateway
   '@anthropic/claude-3-5-haiku-20241022',
@@ -21,4 +21,19 @@ export const DEFAULT_LLM_MODELS = [
   '@ollama/llama-3-1-nemotron-70b-instruct',
   '@ollama/llama-3-1-nemotron-mini-4b-instruct',
   '@ollama/llama-3-2-3b',
-];
+] as const;
+
+// TODO(burdon): Config.
+export const DEFAULT_OLLAMA_ENDPOINT = 'http://localhost:11434';
+
+export const DEFAULT_OLLAMA_MODEL = 'llama3.2:1b';
+
+/**
+ * https://ollama.com/library
+ */
+export const DEFAULT_OLLAMA_MODELS = [
+  //
+  'llama3.2:1b',
+  'llama3:70b',
+  'deepseek-r1:latest',
+] as const;

package/src/ai-service/edge-client.ts

@@ -5,7 +5,7 @@
 import { Schema as S } from 'effect';
 
 import { Message } from '@dxos/artifact';
-import { invariant } from '@dxos/invariant';
+import { assertArgument, invariant } from '@dxos/invariant';
 import { log } from '@dxos/log';
 
 import type { AIServiceClient, GenerationStream } from './service';
@@ -47,7 +47,8 @@ export class AIServiceEdgeClient implements AIServiceClient {
    * Process request and open message stream.
    */
   async exec(request: GenerateRequest): Promise<GenerationStream> {
-    log.info('requesting', { endpoint: this._endpoint });
+    assertArgument(typeof request.model === 'string', 'model is required');
+    log('requesting', { endpoint: this._endpoint });
     const controller = new AbortController();
     const response = await fetch(`${this._endpoint}/generate`, {
       signal: controller.signal,

package/src/ai-service/ollama-client.ts

@@ -2,21 +2,20 @@
 // Copyright 2025 DXOS.org
 //
 
+import { Schema as S } from 'effect';
+
 import { type MessageContentBlock, defineTool, type Tool, ToolResult } from '@dxos/artifact';
-import { ObjectId, S } from '@dxos/echo-schema';
+import { ObjectId } from '@dxos/echo-schema';
 import { invariant } from '@dxos/invariant';
 import { log } from '@dxos/log';
 
+import { DEFAULT_OLLAMA_ENDPOINT } from './defs';
 import { MessageCollector, emitMessageAsEvents } from './message-collector';
 import { type AIServiceClient, type GenerationStream } from './service';
 import { GenerationStreamImpl } from './stream';
 import { ToolTypes, type GenerateRequest, type GenerationStreamEvent } from './types';
 import { isToolUse, runTools } from '../conversation';
 
-// TODO(burdon): Config.
-const DEFAULT_OLLAMA_URL = 'http://localhost:11434';
-const DEFAULT_OLLAMA_MODEL = 'llama3.2:1b';
-
 export type OllamaClientParams = {
   endpoint?: string;
 
@@ -42,23 +41,13 @@ export type OllamaClientParams = {
 };
 
 export class OllamaClient implements AIServiceClient {
-  /**
-   * Create a test client with small local model and no temperature for predictable results.
-   */
-  static createClient(options?: Pick<OllamaClientParams, 'tools'>) {
-    return new OllamaClient({
-      tools: options?.tools,
-      overrides: { model: DEFAULT_OLLAMA_MODEL, temperature: 0 },
-    });
-  }
-
   /**
    * Check if Ollama server is running and accessible.
    * @returns Promise that resolves to true if Ollama is running, false otherwise.
    */
   static async isRunning(): Promise<boolean> {
     try {
-      const response = await fetch(`${DEFAULT_OLLAMA_URL}/api/version`, {
+      const response = await fetch(`${DEFAULT_OLLAMA_ENDPOINT}/api/version`, {
         method: 'GET',
         headers: {
           'Content-Type': 'application/json',
@@ -78,8 +67,8 @@ export class OllamaClient implements AIServiceClient {
   private readonly _temperatureOverride?: number;
   private readonly _maxToolInvocations: number;
 
-  constructor({ endpoint, tools, overrides, maxToolInvocations }: OllamaClientParams) {
-    this._endpoint = endpoint ?? DEFAULT_OLLAMA_URL;
+  constructor({ endpoint, tools, overrides, maxToolInvocations }: OllamaClientParams = {}) {
+    this._endpoint = endpoint ?? DEFAULT_OLLAMA_ENDPOINT;
     this._tools = tools ?? [];
     this._modelOverride = overrides?.model;
     this._temperatureOverride = overrides?.temperature;
@@ -104,7 +93,7 @@ export class OllamaClient implements AIServiceClient {
       messages: [],
       stream: true,
       options: {
-        temperature: this._temperatureOverride ?? 0.7,
+        temperature: this._temperatureOverride,
       },
     };
 
@@ -190,7 +179,7 @@ export class OllamaClient implements AIServiceClient {
       }
     }
 
-    log.info('generate request', { history: request.history, prompt: request.prompt, ollamaRequest });
+    log('generate request', { history: request.history, prompt: request.prompt, ollamaRequest });
 
     const response = await fetch(`${this._endpoint}/api/chat`, {
       method: 'POST',
@@ -201,10 +190,9 @@ export class OllamaClient implements AIServiceClient {
       signal,
     });
 
-    // log.info('ollama response', { response: await response.text() });
-
+    // TODO(burdon): Test if model is not found.
     if (!response.ok) {
-      throw new Error(`Ollama API error: ${response.status} ${response.statusText}`);
+      throw await parseErrorResponse(response);
     }
 
     if (!response.body) {
@@ -253,7 +241,7 @@ export class OllamaClient implements AIServiceClient {
           break;
         }
 
-        log.info('ollama data', { data });
+        log('ollama data', { data });
 
         // Ollama sends chunks of text, we convert them to our event format.
         if (data.message?.content) {
@@ -471,3 +459,41 @@ class OllamaDecoderStream extends TransformStream<string, OllamaResponseData> {
     });
   }
 }
+
+class ModelDoesNotSupportToolsError extends S.TaggedError<ModelDoesNotSupportToolsError>()(
+  'ModelDoesNotSupportToolsError',
+  {
+    model: S.String,
+  },
+) {}
+
+class ModelNotFoundError extends S.TaggedError<ModelNotFoundError>()('ModelNotFoundError', {
+  model: S.String,
+}) {}
+
+const parseOllamaError = (error: string): Error => {
+  {
+    const match = error.match(/^([^']+) does not support tools$/);
+    if (match) {
+      return new ModelDoesNotSupportToolsError({ model: match[1] });
+    }
+  }
+
+  {
+    const match = error.match(/^model "([^"]+)" not found, try pulling it first$/);
+    if (match) {
+      return new ModelNotFoundError({ model: match[1] });
+    }
+  }
+
+  throw new Error(`Ollama API error: ${error}`);
+};
+
+const parseErrorResponse = async (response: Response): Promise<Error> => {
+  if (response.headers.get('content-type')?.includes('application/json')) {
+    const body = await response.json();
+    return parseOllamaError(body.error);
+  }
+
+  throw new Error(`Ollama API error: ${response.status} ${response.statusText} ${await response.text()}`);
+};

package/src/ai-service/parser.ts

@@ -80,9 +80,10 @@ export class MixedStreamParser {
     const messagesCollected: Message[] = [];
 
     for await (const event of stream) {
+      log('streamEvent', { event });
       this.streamEvent.emit(event);
 
-      // log.info('event', { type: event.type, event });
+      // log('event', { type: event.type, event });
       switch (event.type) {
         //
         // Messages.
@@ -143,7 +144,7 @@ export class MixedStreamParser {
             case 'text_delta': {
               const chunks = transformer.transform(event.delta.text);
               for (const chunk of chunks) {
-                // log.info('text_delta', { chunk, current });
+                log('text_delta', { chunk });
 
                 switch (streamBlock?.type) {
                   //
@@ -157,6 +158,7 @@ export class MixedStreamParser {
                         if (stack.length > 0) {
                           const top = stack.pop();
                           invariant(top && top.type === 'tag');
+                          log('pop', { top });
                           top.content.push(streamBlock);
                           streamBlock = top;
                         } else {

package/src/ai-service/stream.test.ts

@@ -2,7 +2,7 @@
 // Copyright 2025 DXOS.org
 //
 
-import { describe, it } from 'vitest';
+import { describe, test } from 'vitest';
 
 import { type Message } from '@dxos/artifact';
 import { log } from '@dxos/log';
@@ -55,7 +55,7 @@ const TEST_BLOCKS = [
 ];
 
 describe('GenerationStream', () => {
-  it('should generate a stream of blocks', async ({ expect }) => {
+  test('should generate a stream of blocks', async ({ expect }) => {
     const stream = createGenerationStream(new Response(createTestSSEStream(TEST_BLOCKS)));
     const events: GenerationStreamEvent[] = [];
     for await (const event of stream) {
@@ -66,7 +66,7 @@ describe('GenerationStream', () => {
   });
 
   for (const splitBy of ['word'] as const) {
-    it(`should emit xml blocks (splitBy: ${splitBy})`, async ({ expect }) => {
+    test(`should emit xml blocks (splitBy: ${splitBy})`, async ({ expect }) => {
       const stream = createGenerationStream(new Response(createTestSSEStream(TEST_BLOCKS, { splitBy })));
       const parser = new MixedStreamParser();
 
@@ -78,7 +78,7 @@ describe('GenerationStream', () => {
       await parser.parse(stream);
 
       for (const block of message?.content ?? []) {
-        log.info('block', { block });
+        log('block', { block });
       }
       expect(message?.content.map((block) => block.type)).to.deep.eq([
         //
@@ -98,4 +98,28 @@ describe('GenerationStream', () => {
       ]);
     });
   }
+
+  test.skip('invalid xml', async ({ expect }) => {
+    const BLOCKS = [
+      //
+      '<cot> thoughts <tool-list> more thoughts </cot> <tool-list/>',
+    ];
+
+    const stream = createGenerationStream(new Response(createTestSSEStream(BLOCKS, { splitBy: 'word' })));
+    const parser = new MixedStreamParser();
+
+    const result = await parser.parse(stream);
+    expect(result[0].content).to.deep.eq([
+      {
+        type: 'text',
+        disposition: 'cot',
+        text: 'thoughts more thoughts',
+      },
+      {
+        type: 'json',
+        disposition: 'tool_list',
+        json: '{}',
+      },
+    ]);
+  });
 });

package/src/ai-service/types.ts

@@ -7,11 +7,11 @@ import { Schema as S } from 'effect';
 import { Tool, Message, type MessageContentBlock, SpaceIdSchema } from '@dxos/artifact';
 import { ObjectId } from '@dxos/echo-schema';
 
-import { DEFAULT_LLM_MODELS } from './defs';
+import { DEFAULT_EDGE_MODELS, DEFAULT_OLLAMA_MODELS } from './defs';
 
 export const createArtifactElement = (id: ObjectId) => `<artifact id=${id} />`;
 
-export const LLMModel = S.Literal(...DEFAULT_LLM_MODELS);
+export const LLMModel = S.Literal(...DEFAULT_EDGE_MODELS, ...DEFAULT_OLLAMA_MODELS);
 
 export type LLMModel = S.Schema.Type<typeof LLMModel>;
 

package/src/bin/chat.ts

@@ -10,7 +10,7 @@ import { ObjectId } from '@dxos/echo-schema';
 import { SpaceId } from '@dxos/keys';
 import { log } from '@dxos/log';
 
-import { AIServiceEdgeClient, ToolTypes, DEFAULT_LLM_MODEL } from '../ai-service';
+import { AIServiceEdgeClient, ToolTypes, DEFAULT_EDGE_MODEL } from '../ai-service';
 import { runLLM } from '../conversation';
 import {
   AI_SERVICE_ENDPOINT,
@@ -49,7 +49,7 @@ while (true) {
   await client.appendMessages([createUserMessage(spaceId, threadId, prompt.message)]);
 
   await runLLM({
-    model: DEFAULT_LLM_MODEL,
+    model: DEFAULT_EDGE_MODEL,
     spaceId,
     threadId,
     system: createSystemPrompt(schemaTypes),
@@ -69,7 +69,7 @@ while (true) {
       onImage: (img) => {
         const path = `/tmp/image-${img.id}.jpeg`;
         writeFileSync(path, Buffer.from(img.source.data, 'base64'));
-        log.info('Saved image', { path });
+        log('Saved image', { path });
         // Print image in iTerm using ANSI escape sequence
         const imageData = img.source.data;
         process.stdout.write('\x1b]1337;File=inline=1:' + imageData + '\x07');

package/src/conversation/conversation.test.ts

@@ -11,7 +11,7 @@ import { SpaceId } from '@dxos/keys';
 import { log } from '@dxos/log';
 
 import { runLLM, type ConversationEvent } from './conversation';
-import { AIServiceEdgeClient, DEFAULT_LLM_MODEL } from '../ai-service';
+import { AIServiceEdgeClient, DEFAULT_EDGE_MODEL } from '../ai-service';
 import { AI_SERVICE_ENDPOINT } from '../testing';
 
 // TODO(burdon): Local live LLM test.
@@ -25,7 +25,7 @@ describe.skip('Conversation tests', () => {
     const threadId = ObjectId.random();
 
     const result = await runLLM({
-      model: DEFAULT_LLM_MODEL,
+      model: DEFAULT_EDGE_MODEL,
       history: [createUserMessage(spaceId, threadId, 'Hello, how are you?')],
       tools: [],
       client,
@@ -56,7 +56,7 @@ describe.skip('Conversation tests', () => {
     const threadId = ObjectId.random();
 
     const result = await runLLM({
-      model: DEFAULT_LLM_MODEL,
+      model: DEFAULT_EDGE_MODEL,
       history: [createUserMessage(spaceId, threadId, 'What is the password? Ask the custodian.')],
       tools: [custodian],
       client,

package/src/conversation/conversation.ts

@@ -175,7 +175,23 @@ export const runTools = async ({ message, tools, extensions }: RunToolsOptions):
   const toolCall = toolCalls[0];
   const tool = tools.find((tool) => tool.name === toolCall.name);
   if (!tool) {
-    throw new Error(`Tool not found: ${toolCall.name}`);
+    const resultMessage = createStatic(Message, {
+      role: 'user',
+      content: [
+        {
+          type: 'tool_result',
+          toolUseId: toolCall.id,
+          content: `Tool not found: ${toolCall.name}`,
+          isError: true,
+        },
+      ],
+    });
+
+    // TODO(burdon): Retry count?
+    return {
+      type: 'continue',
+      message: resultMessage,
+    };
   }
 
   let toolResult: ToolResult;

package/src/cypher/query-executor.test.ts

@@ -12,10 +12,10 @@ import { createTestData } from '../testing';
 test.skip('run', async ({ expect }) => {
   const dataSource = createTestData();
   // for (const node of dataSource.nodes) {
-  //   log.info('node', { id: node.id, label: node.label, name: node.properties.name });
+  //   log('node', { id: node.id, label: node.label, name: node.properties.name });
   // }
   // for (const relationship of dataSource.relationships) {
-  //   log.info('relationship', {
+  //   log('relationship', {
   //     source: relationship.source.id,
   //     target: relationship.target.id,
   //     label: relationship.label,
@@ -43,5 +43,5 @@ test.skip('run', async ({ expect }) => {
     },
   ]);
 
-  log.info('result', { result });
+  log('result', { result });
 });

package/src/cypher/schema.test.ts

@@ -12,6 +12,6 @@ import { Contact, Org, Project, Task } from '../testing';
 test('Org-Project-Task-Contact schema', async ({ expect }) => {
   const jsonSchema = [Org, Project, Task, Contact].map((schema) => toJsonSchema(schema));
   const result = formatJsonSchemaForLLM(jsonSchema);
-  // log.info('result', { result });
+  // log('result', { result });
   expect(result).toHaveLength(8);
 });

package/src/index.ts

@@ -5,3 +5,4 @@
 export * from './ai-service';
 export * from './cypher';
 export * from './conversation';
+export * from './session';

package/src/session/index.ts

@@ -0,0 +1,5 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+export * from './session';