sascha
/
Notes
mirror of https://github.com/TriliumNext/Notes
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

improve LLM response parsing

pull/1325/head
perf3ct 2025-03-17 21:02:05 +07:00
parent c40c702761
commit 14acd1cd89
No known key found for this signature in database
GPG Key ID: 569C4EEC436F5232
1 changed files with 101 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

132
src/services/llm/trilium_context_service.ts
Unescape Escape View File

@ -29,7 +29,7 @@ class TriliumContextService {
// Configuration
private cacheExpiryMs = 5 * 60 * 1000; // 5 minutes
private metaPrompt = `You are an AI assistant that decides what information needs to be retrieved from a knowledge base to answer the user's question.
private metaPrompt = `You are an AI assistant that decides what information needs to be retrieved from a user's knowledge base called TriliumNext Notes to answer the user's question.
Given the user's question, generate 3-5 specific search queries that would help find relevant information.
Each query should be focused on a different aspect of the question.
Format your answer as a JSON array of strings, with each string being a search query.
@ -127,28 +127,74 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
const responseText = response.text; // Extract the text from the response object
try {
// Parse the JSON response
const jsonStr = responseText.trim().replace(/```json|```/g, '').trim();
const queries = JSON.parse(jsonStr);
// Remove code blocks, quotes, and clean up the response text
let jsonStr = responseText
.replace(/```(?:json)?|```/g, '') // Remove code block markers
.replace(/[\u201C\u201D]/g, '"') // Replace smart quotes with straight quotes
.trim();
// Check if the text might contain a JSON array (has square brackets)
if (jsonStr.includes('[') && jsonStr.includes(']')) {
// Extract just the array part if there's explanatory text
const arrayMatch = jsonStr.match(/\[[\s\S]*\]/);
if (arrayMatch) {
jsonStr = arrayMatch[0];
}
if (Array.isArray(queries) && queries.length > 0) {
return queries;
} else {
throw new Error("Invalid response format");
// Try to parse the JSON
try {
const queries = JSON.parse(jsonStr);
if (Array.isArray(queries) && queries.length > 0) {
return queries.map(q => typeof q === 'string' ? q : String(q)).filter(Boolean);
}
} catch (innerError) {
// If parsing fails, log it and continue to the fallback
log.info(`JSON parse error: ${innerError}. Will use fallback parsing for: ${jsonStr}`);
}
}
} catch (parseError) {
// Fallback: if JSON parsing fails, try to extract queries line by line
// Fallback 1: Try to extract an array manually by splitting on commas between quotes
if (jsonStr.includes('[') && jsonStr.includes(']')) {
const arrayContent = jsonStr.substring(
jsonStr.indexOf('[') + 1,
jsonStr.lastIndexOf(']')
);
// Use regex to match quoted strings, handling escaped quotes
const stringMatches = arrayContent.match(/"((?:\\.|[^"\\])*)"/g);
if (stringMatches && stringMatches.length > 0) {
return stringMatches
.map((m: string) => m.substring(1, m.length - 1)) // Remove surrounding quotes
.filter((s: string) => s.length > 0);
}
}
// Fallback 2: Extract queries line by line
const lines = responseText.split('\n')
.map((line: string) => line.trim())
.filter((line: string) => line.length > 0 && !line.startsWith('```'));
.filter((line: string) =>
line.length > 0 &&
!line.startsWith('```') &&
!line.match(/^\d+\.?\s*$/) && // Skip numbered list markers alone
!line.match(/^\[|\]$/) // Skip lines that are just brackets
);
if (lines.length > 0) {
return lines.map((line: string) => line.replace(/^["'\d\.\-\s]+/, '').trim());
// Remove numbering, quotes and other list markers from each line
return lines.map((line: string) => {
return line
.replace(/^\d+\.?\s*/, '') // Remove numbered list markers (1., 2., etc)
.replace(/^[-*•]\s*/, '') // Remove bullet list markers
.replace(/^["']|["']$/g, '') // Remove surrounding quotes
.trim();
}).filter((s: string) => s.length > 0);
}
// If all else fails, just use the original question
return [userQuestion];
} catch (parseError) {
log.error(`Error parsing search queries: ${parseError}`);
}
// If all else fails, just use the original question
return [userQuestion];
} catch (error: unknown) {
const errorMessage = error instanceof Error ? error.message : String(error);
log.error(`Error generating search queries: ${errorMessage}`);
@ -195,10 +241,14 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
// Set to keep track of note IDs we've seen to avoid duplicates
const seenNoteIds = new Set<string>();
// Log the provider and model being used
log.info(`Searching with embedding provider: ${this.provider.name}, model: ${this.provider.getConfig().model}`);
// Process each query
for (const query of queries) {
// Get embeddings for this query using the correct method name
const queryEmbedding = await this.provider.generateEmbeddings(query);
log.info(`Generated embedding for query: "${query}" (${queryEmbedding.length} dimensions)`);
// Find notes similar to this query
let results;
@ -209,6 +259,7 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
contextNoteId,
Math.min(limit, 5) // Limit per query
);
log.info(`Found ${results.length} notes within branch context for query: "${query}"`);
} else {
// Search all notes
results = await vectorStore.findSimilarNotes(
@ -218,6 +269,7 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
Math.min(limit, 5), // Limit per query
0.5 // Lower threshold to get more diverse results
);
log.info(`Found ${results.length} notes in vector store for query: "${query}"`);
}
// Process results
@ -246,6 +298,8 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
.sort((a, b) => b.similarity - a.similarity)
.slice(0, limit);
log.info(`Total unique relevant notes found across all queries: ${sortedResults.length}`);
// Cache the results
this.recentQueriesCache.set(cacheKey, {
timestamp: Date.now(),
@ -362,22 +416,33 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
"with general knowledge about Trilium or other topics you're interested in.";
}
let context = `I've found some relevant information in your notes that may help answer: "${query}"\n\n`;
// Get provider name to adjust context for different models
const providerId = this.provider?.name || 'default';
// Import the constants dynamically to avoid circular dependencies
const { LLM_CONSTANTS } = await import('../../routes/api/llm.js');
// Get appropriate context size and format based on provider
const maxTotalLength =
providerId === 'openai' ? LLM_CONSTANTS.CONTEXT_WINDOW.OPENAI :
providerId === 'anthropic' ? LLM_CONSTANTS.CONTEXT_WINDOW.ANTHROPIC :
providerId === 'ollama' ? LLM_CONSTANTS.CONTEXT_WINDOW.OLLAMA :
LLM_CONSTANTS.CONTEXT_WINDOW.DEFAULT;
// Use a format appropriate for the model family
// Anthropic has a specific system message format that works better with certain structures
const isAnthropicFormat = providerId === 'anthropic';
// Start with different headers based on provider
let context = isAnthropicFormat
? `I'm your AI assistant helping with your Trilium notes database. For your query: "${query}", I found these relevant notes:\n\n`
: `I've found some relevant information in your notes that may help answer: "${query}"\n\n`;
// Sort sources by similarity if available to prioritize most relevant
if (sources[0] && sources[0].similarity !== undefined) {
sources = [...sources].sort((a, b) => (b.similarity || 0) - (a.similarity || 0));
}
// Get provider name to adjust context for different models
const providerId = this.provider?.name || 'default';
// Get approximate max length based on provider using constants
// Import the constants dynamically to avoid circular dependencies
const { LLM_CONSTANTS } = await import('../../routes/api/llm.js');
const maxTotalLength = providerId === 'ollama' ? LLM_CONSTANTS.CONTEXT_WINDOW.OLLAMA :
providerId === 'openai' ? LLM_CONSTANTS.CONTEXT_WINDOW.OPENAI :
LLM_CONSTANTS.CONTEXT_WINDOW.ANTHROPIC;
// Track total context length to avoid oversized context
let currentLength = context.length;
const maxNoteContentLength = Math.min(LLM_CONSTANTS.CONTENT.MAX_NOTE_CONTENT_LENGTH,
@ -387,7 +452,7 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
// Check if adding this source would exceed our total limit
if (currentLength >= maxTotalLength) return;
// Build source section
// Build source section with formatting appropriate for the provider
let sourceSection = `### ${source.title}\n`;
// Add relationship context if available
@ -429,11 +494,16 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
}
});
// Add clear instructions about how to reference the notes
context += "When referring to information from these notes in your response, please cite them by their titles " +
"(e.g., \"According to your note on [Title]...\") rather than using labels like \"Note 1\" or \"Note 2\".\n\n";
context += "If the information doesn't contain what you need, just say so and use your general knowledge instead.";
// Add provider-specific instructions
if (isAnthropicFormat) {
context += "When you refer to any information from these notes, cite the note title explicitly (e.g., \"According to the note [Title]...\"). " +
"If the provided notes don't answer the query fully, acknowledge that and then use your general knowledge to help.\n\n" +
"Be concise but thorough in your responses.";
} else {
context += "When referring to information from these notes in your response, please cite them by their titles " +
"(e.g., \"According to your note on [Title]...\") rather than using labels like \"Note 1\" or \"Note 2\".\n\n" +
"If the information doesn't contain what you need, just say so and use your general knowledge instead.";
}
return context;
}