Tools/Notes
1
0
Fork 0
mirror of https://github.com/TriliumNext/Notes.git synced 2025-08-09 09:42:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

it errors, but works

Browse Source
This commit is contained in:
perf3ct 2025-03-10 04:28:56 +00:00
parent cf0e9242a0
commit ef6ecdc42d
No known key found for this signature in database
GPG Key ID: 569C4EEC436F5232
5 changed files with 332 additions and 208 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
src/public/app/widgets/llm_chat_panel.ts
View File

@ -133,8 +133,20 @@ export default class LlmChatPanel extends BasicWidget {
try { try {
const useAdvancedContext = this.useAdvancedContextCheckbox.checked; const useAdvancedContext = this.useAdvancedContextCheckbox.checked;
// Setup streaming // Create the message parameters
const source = new EventSource(`./api/llm/messages?sessionId=${this.sessionId}&format=stream`); const messageParams = {
content,
contextNoteId: this.currentNoteId,
useAdvancedContext
};
// First, send the message via POST request
await server.post<any>(`llm/sessions/${this.sessionId}/messages`, messageParams);
// Then set up streaming via EventSource
const streamUrl = `./api/llm/sessions/${this.sessionId}/messages?format=stream&useAdvancedContext=${useAdvancedContext}`;
const source = new EventSource(streamUrl);
let assistantResponse = ''; let assistantResponse = '';
// Handle streaming response // Handle streaming response
@ -171,18 +183,6 @@ export default class LlmChatPanel extends BasicWidget {
toastService.showError('Error connecting to the LLM service. Please try again.'); toastService.showError('Error connecting to the LLM service. Please try again.');
}; };
// Send the actual message
const response = await server.post<any>('llm/messages', {
sessionId: this.sessionId,
content,
contextNoteId: this.currentNoteId,
useAdvancedContext
});
// Handle sources if returned in non-streaming response
if (response && response.sources && response.sources.length > 0) {
this.showSources(response.sources);
}
} catch (error) { } catch (error) {
this.hideLoadingIndicator(); this.hideLoadingIndicator();
toastService.showError('Error sending message: ' + (error as Error).message); toastService.showError('Error sending message: ' + (error as Error).message);

402
src/routes/api/llm.ts
View File

@ -449,26 +449,57 @@ Now, based on the above notes, please answer: ${query}`;
*/ */
async function sendMessage(req: Request, res: Response) { async function sendMessage(req: Request, res: Response) {
try { try {
// Extract the content from the request body // Extract parameters differently based on the request method
const { content, sessionId, useAdvancedContext = false } = req.body || {}; let content, useAdvancedContext, sessionId;
// Validate the content if (req.method === 'POST') {
if (!content || typeof content !== 'string' || content.trim().length === 0) { // For POST requests, get content from the request body
const requestBody = req.body || {};
content = requestBody.content;
useAdvancedContext = requestBody.useAdvancedContext || false;
} else if (req.method === 'GET') {
// For GET (streaming) requests, get format from query params
// The content should have been sent in a previous POST request
useAdvancedContext = req.query.useAdvancedContext === 'true';
content = ''; // We don't need content for GET requests
}
// Get sessionId from URL params since it's part of the route
sessionId = req.params.sessionId;
// Get the Accept header once at the start
const acceptHeader = req.get('Accept');
const isStreamingRequest = acceptHeader && acceptHeader.includes('text/event-stream');
// For GET requests, ensure we have the format=stream parameter
if (req.method === 'GET' && (!req.query.format || req.query.format !== 'stream')) {
throw new Error('Stream format parameter is required for GET requests');
}
// For POST requests, validate the content
if (req.method === 'POST' && (!content || typeof content !== 'string' || content.trim().length === 0)) {
throw new Error('Content cannot be empty'); throw new Error('Content cannot be empty');
} }
// Get or create the session // Get session
let session: ChatSession; if (!sessionId || !sessions.has(sessionId)) {
throw new Error('Session not found');
}
if (sessionId && sessions.has(sessionId)) { const session = sessions.get(sessionId)!;
session = sessions.get(sessionId)!; session.lastActive = new Date();
session.lastActive = new Date();
} else { // For POST requests, store the user message
const result = await createSession(req, res); if (req.method === 'POST' && content) {
if (!result?.id) { // Add message to session
throw new Error('Failed to create a new session'); session.messages.push({
} role: 'user',
session = sessions.get(result.id)!; content,
timestamp: new Date()
});
// Log a preview of the message
log.info(`Processing LLM message: "${content.substring(0, 50)}${content.length > 50 ? '...' : ''}"`);
} }
// Check if AI services are available // Check if AI services are available
@ -495,184 +526,225 @@ async function sendMessage(req: Request, res: Response) {
throw new Error('No AI service is available'); throw new Error('No AI service is available');
} }
// Create user message
const userMessage: Message = {
role: 'user',
content
};
// Add message to session
session.messages.push({
role: 'user',
content,
timestamp: new Date()
});
// Log a preview of the message
log.info(`Processing LLM message: "${content.substring(0, 50)}${content.length > 50 ? '...' : ''}"`);
// Information to return to the client // Information to return to the client
let aiResponse = ''; let aiResponse = '';
let sourceNotes: NoteSource[] = []; let sourceNotes: NoteSource[] = [];
// If Advanced Context is enabled, we use the improved method // For POST requests, we need to process the message
if (useAdvancedContext) { // For GET (streaming) requests, we use the latest user message from the session
// Use the Trilium-specific approach if (req.method === 'POST' || isStreamingRequest) {
const contextNoteId = session.noteContext || null; // Get the latest user message for context
const results = await triliumContextService.processQuery(content, service, contextNoteId); const latestUserMessage = session.messages
.filter(msg => msg.role === 'user')
.pop();
// Get the generated context if (!latestUserMessage && req.method === 'GET') {
const context = results.context; throw new Error('No user message found in session');
sourceNotes = results.notes; }
// Add system message with the context // Use the latest message content for GET requests
const contextMessage: Message = { const messageContent = req.method === 'POST' ? content : latestUserMessage!.content;
role: 'system',
content: context
};
// Format all messages for the AI // If Advanced Context is enabled, we use the improved method
const aiMessages: Message[] = [ if (useAdvancedContext) {
contextMessage, // Use the Trilium-specific approach
...session.messages.slice(-10).map(msg => ({ const contextNoteId = session.noteContext || null;
role: msg.role, const results = await triliumContextService.processQuery(messageContent, service, contextNoteId);
content: msg.content
}))
];
// Configure chat options from session metadata // Get the generated context
const chatOptions: ChatCompletionOptions = { const context = results.context;
temperature: session.metadata.temperature || 0.7, sourceNotes = results.notes;
maxTokens: session.metadata.maxTokens,
model: session.metadata.model
// 'provider' property has been removed as it's not in the ChatCompletionOptions type
};
// Get streaming response if requested // Add system message with the context
const acceptHeader = req.get('Accept'); const contextMessage: Message = {
if (acceptHeader && acceptHeader.includes('text/event-stream')) { role: 'system',
res.setHeader('Content-Type', 'text/event-stream'); content: context
res.setHeader('Cache-Control', 'no-cache'); };
res.setHeader('Connection', 'keep-alive');
let messageContent = ''; // Format all messages for the AI
const aiMessages: Message[] = [
contextMessage,
...session.messages.slice(-10).map(msg => ({
role: msg.role,
content: msg.content
}))
];
// Stream the response // Configure chat options from session metadata
await service.sendChatCompletion( const chatOptions: ChatCompletionOptions = {
aiMessages, temperature: session.metadata.temperature || 0.7,
chatOptions, maxTokens: session.metadata.maxTokens,
(chunk: string) => { model: session.metadata.model,
messageContent += chunk; stream: isStreamingRequest ? true : undefined
res.write(`data: ${JSON.stringify({ content: chunk })}\n\n`); };
// Process based on whether this is a streaming request
if (isStreamingRequest) {
res.setHeader('Content-Type', 'text/event-stream');
res.setHeader('Cache-Control', 'no-cache');
res.setHeader('Connection', 'keep-alive');
let messageContent = '';
// Use the correct method name: generateChatCompletion
const response = await service.generateChatCompletion(aiMessages, chatOptions);
// Handle streaming if the response includes a stream method
if (response.stream) {
await response.stream((chunk: { text: string; done: boolean }) => {
if (chunk.text) {
messageContent += chunk.text;
res.write(`data: ${JSON.stringify({ content: chunk.text })}\n\n`);
}
if (chunk.done) {
// Signal the end of the stream when done
res.write('data: [DONE]\n\n');
res.end();
}
});
} else {
// If no streaming available, send the response as a single chunk
messageContent = response.text;
res.write(`data: ${JSON.stringify({ content: messageContent })}\n\n`);
res.write('data: [DONE]\n\n');
res.end();
} }
// Store the full response for the session
aiResponse = messageContent;
// Store the assistant's response in the session
session.messages.push({
role: 'assistant',
content: aiResponse,
timestamp: new Date()
});
} else {
// Non-streaming approach for POST requests
const response = await service.generateChatCompletion(aiMessages, chatOptions);
aiResponse = response.text; // Extract the text from the response
// Store the assistant's response in the session
session.messages.push({
role: 'assistant',
content: aiResponse,
timestamp: new Date()
});
// Return the response for POST requests
return {
content: aiResponse,
sources: sourceNotes.map(note => ({
noteId: note.noteId,
title: note.title,
similarity: note.similarity,
branchId: note.branchId
}))
};
}
} else {
// Original approach - find relevant notes through direct embedding comparison
const relevantNotes = await findRelevantNotes(
content,
session.noteContext || null,
5
); );
// Close the stream sourceNotes = relevantNotes;
res.write('data: [DONE]\n\n');
res.end();
// Store the full response // Build context from relevant notes
aiResponse = messageContent; const context = buildContextFromNotes(relevantNotes, content);
} else {
// Non-streaming approach
aiResponse = await service.sendChatCompletion(aiMessages, chatOptions);
}
} else {
// Original approach - find relevant notes through direct embedding comparison
const relevantNotes = await findRelevantNotes(
content,
session.noteContext || null,
5
);
sourceNotes = relevantNotes; // Add system message with the context
const contextMessage: Message = {
role: 'system',
content: context
};
// Build context from relevant notes // Format all messages for the AI
const context = buildContextFromNotes(relevantNotes, content); const aiMessages: Message[] = [
contextMessage,
...session.messages.slice(-10).map(msg => ({
role: msg.role,
content: msg.content
}))
];
// Add system message with the context // Configure chat options from session metadata
const contextMessage: Message = { const chatOptions: ChatCompletionOptions = {
role: 'system', temperature: session.metadata.temperature || 0.7,
content: context maxTokens: session.metadata.maxTokens,
}; model: session.metadata.model,
stream: isStreamingRequest ? true : undefined
};
// Format all messages for the AI if (isStreamingRequest) {
const aiMessages: Message[] = [ res.setHeader('Content-Type', 'text/event-stream');
contextMessage, res.setHeader('Cache-Control', 'no-cache');
...session.messages.slice(-10).map(msg => ({ res.setHeader('Connection', 'keep-alive');
role: msg.role,
content: msg.content
}))
];
// Configure chat options from session metadata let messageContent = '';
const chatOptions: ChatCompletionOptions = {
temperature: session.metadata.temperature || 0.7,
maxTokens: session.metadata.maxTokens,
model: session.metadata.model
// 'provider' property has been removed as it's not in the ChatCompletionOptions type
};
// Get streaming response if requested // Use the correct method name: generateChatCompletion
const acceptHeader = req.get('Accept'); const response = await service.generateChatCompletion(aiMessages, chatOptions);
if (acceptHeader && acceptHeader.includes('text/event-stream')) {
res.setHeader('Content-Type', 'text/event-stream');
res.setHeader('Cache-Control', 'no-cache');
res.setHeader('Connection', 'keep-alive');
let messageContent = ''; // Handle streaming if the response includes a stream method
if (response.stream) {
await response.stream((chunk: { text: string; done: boolean }) => {
if (chunk.text) {
messageContent += chunk.text;
res.write(`data: ${JSON.stringify({ content: chunk.text })}\n\n`);
}
// Stream the response if (chunk.done) {
await service.sendChatCompletion( // Signal the end of the stream when done
aiMessages, res.write('data: [DONE]\n\n');
chatOptions, res.end();
(chunk: string) => { }
messageContent += chunk; });
res.write(`data: ${JSON.stringify({ content: chunk })}\n\n`); } else {
// If no streaming available, send the response as a single chunk
messageContent = response.text;
res.write(`data: ${JSON.stringify({ content: messageContent })}\n\n`);
res.write('data: [DONE]\n\n');
res.end();
} }
);
// Close the stream // Store the full response for the session
res.write('data: [DONE]\n\n'); aiResponse = messageContent;
res.end();
// Store the full response // Store the assistant's response in the session
aiResponse = messageContent; session.messages.push({
} else { role: 'assistant',
// Non-streaming approach content: aiResponse,
aiResponse = await service.sendChatCompletion(aiMessages, chatOptions); timestamp: new Date()
});
} else {
// Non-streaming approach for POST requests
const response = await service.generateChatCompletion(aiMessages, chatOptions);
aiResponse = response.text; // Extract the text from the response
// Store the assistant's response in the session
session.messages.push({
role: 'assistant',
content: aiResponse,
timestamp: new Date()
});
// Return the response for POST requests
return {
content: aiResponse,
sources: sourceNotes.map(note => ({
noteId: note.noteId,
title: note.title,
similarity: note.similarity,
branchId: note.branchId
}))
};
}
} }
} }
// Only store the assistant's message if we're not streaming (otherwise we already did)
const acceptHeader = req.get('Accept');
if (!acceptHeader || !acceptHeader.includes('text/event-stream')) {
// Store the assistant's response in the session
session.messages.push({
role: 'assistant',
content: aiResponse,
timestamp: new Date()
});
// Return the response
return {
content: aiResponse,
sources: sourceNotes.map(note => ({
noteId: note.noteId,
title: note.title,
similarity: note.similarity,
branchId: note.branchId
}))
};
} else {
// For streaming responses, we've already sent the data
// But we still need to add the message to the session
session.messages.push({
role: 'assistant',
content: aiResponse,
timestamp: new Date()
});
}
} catch (error: any) { } catch (error: any) {
log.error(`Error sending message to LLM: ${error.message}`); log.error(`Error sending message to LLM: ${error.message}`);
throw new Error(`Failed to send message: ${error.message}`); throw new Error(`Failed to send message: ${error.message}`);

1
src/routes/routes.ts
View File

@ -387,6 +387,7 @@ function register(app: express.Application) {
apiRoute(PATCH, "/api/llm/sessions/:sessionId", llmRoute.updateSession); apiRoute(PATCH, "/api/llm/sessions/:sessionId", llmRoute.updateSession);
apiRoute(DEL, "/api/llm/sessions/:sessionId", llmRoute.deleteSession); apiRoute(DEL, "/api/llm/sessions/:sessionId", llmRoute.deleteSession);
apiRoute(PST, "/api/llm/sessions/:sessionId/messages", llmRoute.sendMessage); apiRoute(PST, "/api/llm/sessions/:sessionId/messages", llmRoute.sendMessage);
route(GET, "/api/llm/sessions/:sessionId/messages", [auth.checkApiAuth, csrfMiddleware], llmRoute.sendMessage, apiResultHandler);
// Ollama API endpoints // Ollama API endpoints
route(PST, "/api/ollama/list-models", [auth.checkApiAuth, csrfMiddleware], ollamaRoute.listModels, apiResultHandler); route(PST, "/api/ollama/list-models", [auth.checkApiAuth, csrfMiddleware], ollamaRoute.listModels, apiResultHandler);

15
src/services/llm/ai_service_manager.ts
View File

@ -40,7 +40,20 @@ export class AIServiceManager {
if (customOrder) { if (customOrder) {
try { try {
const parsed = JSON.parse(customOrder); // Try to parse as JSON first
let parsed;
// Handle both array in JSON format and simple string format
if (customOrder.startsWith('[') && customOrder.endsWith(']')) {
parsed = JSON.parse(customOrder);
} else if (typeof customOrder === 'string') {
// If it's a simple string (like "ollama"), convert to single-item array
parsed = [customOrder];
} else {
// Fallback to default
parsed = defaultOrder;
}
// Validate that all providers are valid // Validate that all providers are valid
if (Array.isArray(parsed) && if (Array.isArray(parsed) &&
parsed.every(p => Object.keys(this.services).includes(p))) { parsed.every(p => Object.keys(this.services).includes(p))) {

94
src/services/llm/trilium_context_service.ts
View File

@ -102,12 +102,13 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
maxTokens: 300 maxTokens: 300
}; };
// Get the response from the LLM // Get the response from the LLM using the correct method name
const response = await llmService.sendTextCompletion(messages, options); const response = await llmService.generateChatCompletion(messages, options);
const responseText = response.text; // Extract the text from the response object
try { try {
// Parse the JSON response // Parse the JSON response
const jsonStr = response.trim().replace(/```json|```/g, '').trim(); const jsonStr = responseText.trim().replace(/```json|```/g, '').trim();
const queries = JSON.parse(jsonStr); const queries = JSON.parse(jsonStr);
if (Array.isArray(queries) && queries.length > 0) { if (Array.isArray(queries) && queries.length > 0) {
@ -117,7 +118,7 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
} }
} catch (parseError) { } catch (parseError) {
// Fallback: if JSON parsing fails, try to extract queries line by line // Fallback: if JSON parsing fails, try to extract queries line by line
const lines = response.split('\n') const lines = responseText.split('\n')
.map((line: string) => line.trim()) .map((line: string) => line.trim())
.filter((line: string) => line.length > 0 && !line.startsWith('```')); .filter((line: string) => line.length > 0 && !line.startsWith('```'));
@ -176,8 +177,8 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
// Process each query // Process each query
for (const query of queries) { for (const query of queries) {
// Get embeddings for this query // Get embeddings for this query using the correct method name
const queryEmbedding = await this.provider.getEmbedding(query); const queryEmbedding = await this.provider.generateEmbeddings(query);
// Find notes similar to this query // Find notes similar to this query
let results; let results;
@ -192,8 +193,8 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
// Search all notes // Search all notes
results = await vectorStore.findSimilarNotes( results = await vectorStore.findSimilarNotes(
queryEmbedding, queryEmbedding,
this.provider.id, this.provider.name, // Use name property instead of id
this.provider.modelId, this.provider.getConfig().model, // Use getConfig().model instead of modelId
Math.min(limit, 5), // Limit per query Math.min(limit, 5), // Limit per query
0.5 // Lower threshold to get more diverse results 0.5 // Lower threshold to get more diverse results
); );
@ -265,8 +266,8 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
for (const noteId of subtreeNoteIds) { for (const noteId of subtreeNoteIds) {
const noteEmbedding = await vectorStore.getEmbeddingForNote( const noteEmbedding = await vectorStore.getEmbeddingForNote(
noteId, noteId,
this.provider.id, this.provider.name, // Use name property instead of id
this.provider.modelId this.provider.getConfig().model // Use getConfig().model instead of modelId
); );
if (noteEmbedding) { if (noteEmbedding) {
@ -338,7 +339,10 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
*/ */
buildContextFromNotes(sources: any[], query: string): string { buildContextFromNotes(sources: any[], query: string): string {
if (!sources || sources.length === 0) { if (!sources || sources.length === 0) {
return ""; // Return a default context instead of empty string
return "I am an AI assistant helping you with your Trilium notes. " +
"I couldn't find any specific notes related to your query, but I'll try to assist you " +
"with general knowledge about Trilium or other topics you're interested in.";
} }
let context = `The following are relevant notes from your knowledge base that may help answer the query: "${query}"\n\n`; let context = `The following are relevant notes from your knowledge base that may help answer the query: "${query}"\n\n`;
@ -382,28 +386,62 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
*/ */
async processQuery(userQuestion: string, llmService: any, contextNoteId: string | null = null) { async processQuery(userQuestion: string, llmService: any, contextNoteId: string | null = null) {
if (!this.initialized) { if (!this.initialized) {
await this.initialize(); try {
await this.initialize();
} catch (error) {
log.error(`Failed to initialize TriliumContextService: ${error}`);
// Return a fallback response if initialization fails
return {
context: "I am an AI assistant helping you with your Trilium notes. " +
"I'll try to assist you with general knowledge about your query.",
notes: [],
queries: [userQuestion]
};
}
} }
// Step 1: Generate search queries try {
const searchQueries = await this.generateSearchQueries(userQuestion, llmService); // Step 1: Generate search queries
log.info(`Generated search queries: ${JSON.stringify(searchQueries)}`); let searchQueries: string[];
try {
searchQueries = await this.generateSearchQueries(userQuestion, llmService);
} catch (error) {
log.error(`Error generating search queries, using fallback: ${error}`);
searchQueries = [userQuestion]; // Fallback to using the original question
}
log.info(`Generated search queries: ${JSON.stringify(searchQueries)}`);
// Step 2: Find relevant notes using those queries // Step 2: Find relevant notes using those queries
const relevantNotes = await this.findRelevantNotesMultiQuery( let relevantNotes: any[] = [];
searchQueries, try {
contextNoteId, relevantNotes = await this.findRelevantNotesMultiQuery(
8 // Get more notes since we're using multiple queries searchQueries,
); contextNoteId,
8 // Get more notes since we're using multiple queries
);
} catch (error) {
log.error(`Error finding relevant notes: ${error}`);
// Continue with empty notes list
}
// Step 3: Build context from the notes // Step 3: Build context from the notes
const context = this.buildContextFromNotes(relevantNotes, userQuestion); const context = this.buildContextFromNotes(relevantNotes, userQuestion);
return { return {
context, context,
notes: relevantNotes, notes: relevantNotes,
queries: searchQueries queries: searchQueries
}; };
} catch (error) {
log.error(`Error in processQuery: ${error}`);
// Return a fallback response if anything fails
return {
context: "I am an AI assistant helping you with your Trilium notes. " +
"I encountered an error while processing your query, but I'll try to assist you anyway.",
notes: [],
queries: [userQuestion]
};
}
} }
} }