mirror of
https://github.com/TriliumNext/Notes.git
synced 2025-07-29 02:52:27 +08:00
set up agentic thinking
This commit is contained in:
perf3ct
parent
1a8ce967d9
commit
f6afb1d963
361
src/services/llm/agent_tools/contextual_thinking_tool.ts
Normal file
361
src/services/llm/agent_tools/contextual_thinking_tool.ts
Normal file
@ -0,0 +1,361 @@
|
||||
/**
|
||||
* Contextual Thinking Tool
|
||||
*
|
||||
* Provides a way for the LLM agent to expose its reasoning process to the user,
|
||||
* showing how it explores knowledge and reaches conclusions. This makes the
|
||||
* agent's thinking more transparent and allows users to understand the context
|
||||
* behind answers.
|
||||
*
|
||||
* Features:
|
||||
* - Capture and structure the agent's thinking steps
|
||||
* - Visualize reasoning chains for complex queries
|
||||
* - Expose confidence levels for different assertions
|
||||
* - Show how different sources of evidence are weighed
|
||||
*/
|
||||
|
||||
import log from '../../log.js';
|
||||
|
||||
/**
|
||||
* Represents a single reasoning step taken by the agent
|
||||
*/
|
||||
export interface ThinkingStep {
|
||||
id: string;
|
||||
content: string;
|
||||
type: 'observation' | 'hypothesis' | 'question' | 'evidence' | 'conclusion';
|
||||
confidence?: number;
|
||||
sources?: string[];
|
||||
parentId?: string;
|
||||
children?: string[];
|
||||
metadata?: Record<string, any>;
|
||||
}
|
||||
|
||||
/**
|
||||
* Contains the full reasoning process
|
||||
*/
|
||||
export interface ThinkingProcess {
|
||||
id: string;
|
||||
query: string;
|
||||
steps: ThinkingStep[];
|
||||
status: 'in_progress' | 'completed';
|
||||
startTime: number;
|
||||
endTime?: number;
|
||||
}
|
||||
|
||||
export class ContextualThinkingTool {
|
||||
private static thinkingCounter = 0;
|
||||
private static stepCounter = 0;
|
||||
private activeProcId?: string;
|
||||
private processes: Record<string, ThinkingProcess> = {};
|
||||
|
||||
/**
|
||||
* Start a new thinking process for a given query
|
||||
*
|
||||
* @param query The user query that initiated the thinking process
|
||||
* @returns The ID of the new thinking process
|
||||
*/
|
||||
startThinking(query: string): string {
|
||||
const id = this.generateProcessId();
|
||||
|
||||
this.processes[id] = {
|
||||
id,
|
||||
query,
|
||||
steps: [],
|
||||
status: 'in_progress',
|
||||
startTime: Date.now()
|
||||
};
|
||||
|
||||
this.activeProcId = id;
|
||||
return id;
|
||||
}
|
||||
|
||||
/**
|
||||
* Add a thinking step to the current active process
|
||||
*
|
||||
* @param content The content of the thinking step
|
||||
* @param type The type of thinking step
|
||||
* @param options Additional options for the step
|
||||
* @returns The ID of the new step
|
||||
*/
|
||||
addThinkingStep(
|
||||
content: string,
|
||||
type: ThinkingStep['type'],
|
||||
options: {
|
||||
confidence?: number;
|
||||
sources?: string[];
|
||||
parentId?: string;
|
||||
metadata?: Record<string, any>;
|
||||
} = {}
|
||||
): string | null {
|
||||
if (!this.activeProcId || !this.processes[this.activeProcId]) {
|
||||
log.error("No active thinking process to add step to");
|
||||
return null;
|
||||
}
|
||||
|
||||
const stepId = this.generateStepId();
|
||||
const step: ThinkingStep = {
|
||||
id: stepId,
|
||||
content,
|
||||
type,
|
||||
...options
|
||||
};
|
||||
|
||||
// Add to parent's children if a parent is specified
|
||||
if (options.parentId) {
|
||||
const parentIdx = this.processes[this.activeProcId].steps.findIndex(
|
||||
s => s.id === options.parentId
|
||||
);
|
||||
|
||||
if (parentIdx >= 0) {
|
||||
const parent = this.processes[this.activeProcId].steps[parentIdx];
|
||||
if (!parent.children) {
|
||||
parent.children = [];
|
||||
}
|
||||
parent.children.push(stepId);
|
||||
this.processes[this.activeProcId].steps[parentIdx] = parent;
|
||||
}
|
||||
}
|
||||
|
||||
this.processes[this.activeProcId].steps.push(step);
|
||||
return stepId;
|
||||
}
|
||||
|
||||
/**
|
||||
* Complete the current thinking process
|
||||
*
|
||||
* @param processId The ID of the process to complete (defaults to active process)
|
||||
* @returns The completed thinking process
|
||||
*/
|
||||
completeThinking(processId?: string): ThinkingProcess | null {
|
||||
const id = processId || this.activeProcId;
|
||||
|
||||
if (!id || !this.processes[id]) {
|
||||
log.error(`Thinking process ${id} not found`);
|
||||
return null;
|
||||
}
|
||||
|
||||
this.processes[id].status = 'completed';
|
||||
this.processes[id].endTime = Date.now();
|
||||
|
||||
if (id === this.activeProcId) {
|
||||
this.activeProcId = undefined;
|
||||
}
|
||||
|
||||
return this.processes[id];
|
||||
}
|
||||
|
||||
/**
|
||||
* Get a thinking process by ID
|
||||
*/
|
||||
getThinkingProcess(processId: string): ThinkingProcess | null {
|
||||
return this.processes[processId] || null;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the active thinking process
|
||||
*/
|
||||
getActiveThinkingProcess(): ThinkingProcess | null {
|
||||
if (!this.activeProcId) return null;
|
||||
return this.processes[this.activeProcId] || null;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a user-friendly HTML representation of the thinking process
|
||||
*
|
||||
* @param processId The ID of the process to visualize
|
||||
* @returns HTML string representing the thinking process
|
||||
*/
|
||||
visualizeThinking(processId: string): string {
|
||||
const process = this.getThinkingProcess(processId);
|
||||
if (!process) {
|
||||
return `<div class="thinking-error">Thinking process ${processId} not found</div>`;
|
||||
}
|
||||
|
||||
let html = `
|
||||
<div class="thinking-process">
|
||||
<div class="thinking-header">
|
||||
<h3>Thinking Process for: "${process.query}"</h3>
|
||||
<div class="thinking-metadata">
|
||||
<span>Status: ${process.status}</span>
|
||||
<span>Steps: ${process.steps.length}</span>
|
||||
<span>Time: ${this.formatDuration(process.startTime, process.endTime || Date.now())}</span>
|
||||
</div>
|
||||
</div>
|
||||
<div class="thinking-steps">
|
||||
`;
|
||||
|
||||
// Find root steps (those without parents)
|
||||
const rootSteps = process.steps.filter(step => !step.parentId);
|
||||
|
||||
// Recursively render the thinking tree
|
||||
for (const rootStep of rootSteps) {
|
||||
html += this.renderStepTree(rootStep, process.steps);
|
||||
}
|
||||
|
||||
html += `
|
||||
</div>
|
||||
</div>
|
||||
`;
|
||||
|
||||
return html;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a concise text representation of the thinking process
|
||||
* that can be displayed inline in the chat for transparency
|
||||
*
|
||||
* @param processId The ID of the process to summarize
|
||||
* @returns Text summary of the reasoning process
|
||||
*/
|
||||
getThinkingSummary(processId?: string): string {
|
||||
const id = processId || this.activeProcId;
|
||||
if (!id || !this.processes[id]) {
|
||||
return "No thinking process available.";
|
||||
}
|
||||
|
||||
const process = this.processes[id];
|
||||
let summary = `Thinking about: "${process.query}"\n\n`;
|
||||
|
||||
// Group steps by type
|
||||
const stepsByType: Record<string, ThinkingStep[]> = {};
|
||||
for (const step of process.steps) {
|
||||
if (!stepsByType[step.type]) {
|
||||
stepsByType[step.type] = [];
|
||||
}
|
||||
stepsByType[step.type].push(step);
|
||||
}
|
||||
|
||||
// Show observations first
|
||||
if (stepsByType['observation'] && stepsByType['observation'].length > 0) {
|
||||
summary += "🔍 Observations:\n";
|
||||
for (const step of stepsByType['observation'].slice(0, 3)) {
|
||||
summary += `- ${step.content}\n`;
|
||||
}
|
||||
if (stepsByType['observation'].length > 3) {
|
||||
summary += `- ...and ${stepsByType['observation'].length - 3} more observations\n`;
|
||||
}
|
||||
summary += "\n";
|
||||
}
|
||||
|
||||
// Show questions the agent asked itself
|
||||
if (stepsByType['question'] && stepsByType['question'].length > 0) {
|
||||
summary += "❓ Questions considered:\n";
|
||||
for (const step of stepsByType['question'].slice(0, 3)) {
|
||||
summary += `- ${step.content}\n`;
|
||||
}
|
||||
if (stepsByType['question'].length > 3) {
|
||||
summary += `- ...and ${stepsByType['question'].length - 3} more questions\n`;
|
||||
}
|
||||
summary += "\n";
|
||||
}
|
||||
|
||||
// Show evidence
|
||||
if (stepsByType['evidence'] && stepsByType['evidence'].length > 0) {
|
||||
summary += "📋 Evidence found:\n";
|
||||
for (const step of stepsByType['evidence'].slice(0, 3)) {
|
||||
summary += `- ${step.content}\n`;
|
||||
}
|
||||
if (stepsByType['evidence'].length > 3) {
|
||||
summary += `- ...and ${stepsByType['evidence'].length - 3} more pieces of evidence\n`;
|
||||
}
|
||||
summary += "\n";
|
||||
}
|
||||
|
||||
// Show conclusions
|
||||
if (stepsByType['conclusion'] && stepsByType['conclusion'].length > 0) {
|
||||
summary += "✅ Conclusions:\n";
|
||||
for (const step of stepsByType['conclusion']) {
|
||||
const confidence = step.confidence ? ` (${Math.round(step.confidence * 100)}% confidence)` : '';
|
||||
summary += `- ${step.content}${confidence}\n`;
|
||||
}
|
||||
}
|
||||
|
||||
return summary;
|
||||
}
|
||||
|
||||
/**
|
||||
* Reset the active thinking process
|
||||
*/
|
||||
resetActiveThinking(): void {
|
||||
this.activeProcId = undefined;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a unique ID for a thinking process
|
||||
*/
|
||||
private generateProcessId(): string {
|
||||
return `thinking_${Date.now()}_${ContextualThinkingTool.thinkingCounter++}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a unique ID for a thinking step
|
||||
*/
|
||||
private generateStepId(): string {
|
||||
return `step_${Date.now()}_${ContextualThinkingTool.stepCounter++}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Format duration between two timestamps
|
||||
*/
|
||||
private formatDuration(start: number, end: number): string {
|
||||
const durationMs = end - start;
|
||||
if (durationMs < 1000) {
|
||||
return `${durationMs}ms`;
|
||||
} else if (durationMs < 60000) {
|
||||
return `${Math.round(durationMs / 1000)}s`;
|
||||
} else {
|
||||
return `${Math.round(durationMs / 60000)}m ${Math.round((durationMs % 60000) / 1000)}s`;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Recursively render a step and its children
|
||||
*/
|
||||
private renderStepTree(step: ThinkingStep, allSteps: ThinkingStep[]): string {
|
||||
const typeIcons: Record<string, string> = {
|
||||
'observation': '🔍',
|
||||
'hypothesis': '🤔',
|
||||
'question': '❓',
|
||||
'evidence': '📋',
|
||||
'conclusion': '✅'
|
||||
};
|
||||
|
||||
const icon = typeIcons[step.type] || '•';
|
||||
const confidenceDisplay = step.confidence !== undefined
|
||||
? `<span class="confidence">${Math.round(step.confidence * 100)}%</span>`
|
||||
: '';
|
||||
|
||||
let html = `
|
||||
<div class="thinking-step thinking-${step.type}">
|
||||
<div class="step-header">
|
||||
<span class="step-icon">${icon}</span>
|
||||
<span class="step-type">${step.type}</span>
|
||||
${confidenceDisplay}
|
||||
</div>
|
||||
<div class="step-content">${step.content}</div>
|
||||
`;
|
||||
|
||||
// Add sources if available
|
||||
if (step.sources && step.sources.length > 0) {
|
||||
html += `<div class="step-sources">Sources: ${step.sources.join(', ')}</div>`;
|
||||
}
|
||||
|
||||
// Recursively render children
|
||||
if (step.children && step.children.length > 0) {
|
||||
html += `<div class="step-children">`;
|
||||
|
||||
for (const childId of step.children) {
|
||||
const childStep = allSteps.find(s => s.id === childId);
|
||||
if (childStep) {
|
||||
html += this.renderStepTree(childStep, allSteps);
|
||||
}
|
||||
}
|
||||
|
||||
html += `</div>`;
|
||||
}
|
||||
|
||||
html += `</div>`;
|
||||
return html;
|
||||
}
|
||||
}
|
||||
|
||||
export default ContextualThinkingTool;
|
||||
130
src/services/llm/agent_tools/index.ts
Normal file
130
src/services/llm/agent_tools/index.ts
Normal file
@ -0,0 +1,130 @@
|
||||
/**
|
||||
* Agent Tools Index
|
||||
*
|
||||
* This file exports all available agent tools for use by the LLM.
|
||||
* Tools are prioritized in order of importance/impact.
|
||||
*/
|
||||
|
||||
import { VectorSearchTool } from './vector_search_tool.js';
|
||||
import { NoteNavigatorTool } from './note_navigator_tool.js';
|
||||
import { QueryDecompositionTool } from './query_decomposition_tool.js';
|
||||
import { ContextualThinkingTool } from './contextual_thinking_tool.js';
|
||||
|
||||
// Import services needed for initialization
|
||||
import SemanticContextService from '../semantic_context_service.js';
|
||||
import aiServiceManager from '../ai_service_manager.js';
|
||||
import log from '../../log.js';
|
||||
|
||||
/**
|
||||
* Manages all agent tools and provides a unified interface for the LLM agent
|
||||
*/
|
||||
export class AgentToolsManager {
|
||||
private vectorSearchTool: VectorSearchTool | null = null;
|
||||
private noteNavigatorTool: NoteNavigatorTool | null = null;
|
||||
private queryDecompositionTool: QueryDecompositionTool | null = null;
|
||||
private contextualThinkingTool: ContextualThinkingTool | null = null;
|
||||
private initialized = false;
|
||||
|
||||
constructor() {
|
||||
// Initialize tools only when requested to avoid circular dependencies
|
||||
}
|
||||
|
||||
async initialize(aiServiceManager: any): Promise<void> {
|
||||
try {
|
||||
if (this.initialized) {
|
||||
return;
|
||||
}
|
||||
|
||||
log.info("Initializing LLM agent tools...");
|
||||
|
||||
// Create tools
|
||||
this.vectorSearchTool = new VectorSearchTool();
|
||||
this.noteNavigatorTool = new NoteNavigatorTool();
|
||||
this.queryDecompositionTool = new QueryDecompositionTool();
|
||||
this.contextualThinkingTool = new ContextualThinkingTool();
|
||||
|
||||
// Get semantic context service and set it in the vector search tool
|
||||
const semanticContext = aiServiceManager.getSemanticContextService();
|
||||
this.vectorSearchTool.setSemanticContext(semanticContext);
|
||||
|
||||
this.initialized = true;
|
||||
log.info("LLM agent tools initialized successfully");
|
||||
} catch (error: any) {
|
||||
log.error(`Failed to initialize LLM agent tools: ${error.message}`);
|
||||
throw new Error(`Agent tools initialization failed: ${error.message}`);
|
||||
}
|
||||
}
|
||||
|
||||
isInitialized(): boolean {
|
||||
return this.initialized;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get all available agent tools
|
||||
* @returns Object containing all initialized tools
|
||||
*/
|
||||
getAllTools() {
|
||||
if (!this.initialized) {
|
||||
throw new Error("Agent tools not initialized. Call initialize() first.");
|
||||
}
|
||||
|
||||
return {
|
||||
vectorSearch: this.vectorSearchTool,
|
||||
noteNavigator: this.noteNavigatorTool,
|
||||
queryDecomposition: this.queryDecompositionTool,
|
||||
contextualThinking: this.contextualThinkingTool
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the vector search tool
|
||||
*/
|
||||
getVectorSearchTool(): VectorSearchTool {
|
||||
if (!this.initialized || !this.vectorSearchTool) {
|
||||
throw new Error("Vector search tool not initialized");
|
||||
}
|
||||
return this.vectorSearchTool;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the note structure navigator tool
|
||||
*/
|
||||
getNoteNavigatorTool(): NoteNavigatorTool {
|
||||
if (!this.initialized || !this.noteNavigatorTool) {
|
||||
throw new Error("Note navigator tool not initialized");
|
||||
}
|
||||
return this.noteNavigatorTool;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the query decomposition tool
|
||||
*/
|
||||
getQueryDecompositionTool(): QueryDecompositionTool {
|
||||
if (!this.initialized || !this.queryDecompositionTool) {
|
||||
throw new Error("Query decomposition tool not initialized");
|
||||
}
|
||||
return this.queryDecompositionTool;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the contextual thinking tool
|
||||
*/
|
||||
getContextualThinkingTool(): ContextualThinkingTool {
|
||||
if (!this.initialized || !this.contextualThinkingTool) {
|
||||
throw new Error("Contextual thinking tool not initialized");
|
||||
}
|
||||
return this.contextualThinkingTool;
|
||||
}
|
||||
}
|
||||
|
||||
// Export a singleton instance
|
||||
const agentTools = new AgentToolsManager();
|
||||
export default agentTools;
|
||||
|
||||
// Also export individual tool classes for direct use if needed
|
||||
export {
|
||||
VectorSearchTool,
|
||||
NoteNavigatorTool,
|
||||
QueryDecompositionTool,
|
||||
ContextualThinkingTool
|
||||
};
|
||||
463
src/services/llm/agent_tools/note_navigator_tool.ts
Normal file
463
src/services/llm/agent_tools/note_navigator_tool.ts
Normal file
@ -0,0 +1,463 @@
|
||||
/**
|
||||
* Note Structure Navigator Tool
|
||||
*
|
||||
* This tool enables the LLM agent to navigate through the hierarchical
|
||||
* structure of notes in the knowledge base. It provides methods for:
|
||||
* - Finding paths between notes
|
||||
* - Exploring parent-child relationships
|
||||
* - Discovering note attributes and metadata
|
||||
* - Understanding the context of a note within the broader structure
|
||||
*
|
||||
* This helps the LLM agent provide more accurate and contextually relevant responses.
|
||||
*/
|
||||
|
||||
import becca from '../../../becca/becca.js';
|
||||
import log from '../../log.js';
|
||||
import type BNote from '../../../becca/entities/bnote.js';
|
||||
import type BAttribute from '../../../becca/entities/battribute.js';
|
||||
|
||||
export interface NoteInfo {
|
||||
noteId: string;
|
||||
title: string;
|
||||
type: string;
|
||||
mime?: string;
|
||||
dateCreated?: string;
|
||||
dateModified?: string;
|
||||
isProtected: boolean;
|
||||
isArchived: boolean;
|
||||
attributeNames: string[];
|
||||
hasChildren: boolean;
|
||||
}
|
||||
|
||||
export interface NotePathInfo {
|
||||
notePath: string[];
|
||||
notePathTitles: string[];
|
||||
}
|
||||
|
||||
export interface NoteHierarchyLevel {
|
||||
noteId: string;
|
||||
title: string;
|
||||
level: number;
|
||||
children?: NoteHierarchyLevel[];
|
||||
}
|
||||
|
||||
export class NoteNavigatorTool {
|
||||
private maxPathLength: number = 20;
|
||||
private maxBreadth: number = 100;
|
||||
private maxDepth: number = 5;
|
||||
|
||||
/**
|
||||
* Get detailed information about a note
|
||||
*/
|
||||
getNoteInfo(noteId: string): NoteInfo | null {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return null;
|
||||
}
|
||||
|
||||
// Get attribute names for this note
|
||||
const attributeNames = note.ownedAttributes
|
||||
.map(attr => attr.name)
|
||||
.filter((value, index, self) => self.indexOf(value) === index); // unique values
|
||||
|
||||
return {
|
||||
noteId: note.noteId,
|
||||
title: note.title,
|
||||
type: note.type,
|
||||
mime: note.mime,
|
||||
dateCreated: note.dateCreated,
|
||||
dateModified: note.dateModified,
|
||||
isProtected: note.isProtected ?? false,
|
||||
isArchived: note.isArchived || false,
|
||||
attributeNames,
|
||||
hasChildren: note.children.length > 0
|
||||
};
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note info: ${error.message}`);
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get all paths to a note from the root
|
||||
*/
|
||||
getNotePathsFromRoot(noteId: string): NotePathInfo[] {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// Get all possible paths to this note
|
||||
const allPaths = note.getAllNotePaths();
|
||||
if (!allPaths || allPaths.length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// Convert path IDs to titles
|
||||
return allPaths.map(path => {
|
||||
const titles = path.map(id => {
|
||||
const pathNote = becca.notes[id];
|
||||
return pathNote ? pathNote.title : id;
|
||||
});
|
||||
|
||||
return {
|
||||
notePath: path,
|
||||
notePathTitles: titles
|
||||
};
|
||||
}).sort((a, b) => a.notePath.length - b.notePath.length); // Sort by path length, shortest first
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note paths: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the parent notes of a given note
|
||||
*/
|
||||
getParentNotes(noteId: string): NoteInfo[] {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note || !note.parents) {
|
||||
return [];
|
||||
}
|
||||
|
||||
return note.parents
|
||||
.map(parent => this.getNoteInfo(parent.noteId))
|
||||
.filter((info): info is NoteInfo => info !== null);
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting parent notes: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the children notes of a given note
|
||||
*/
|
||||
getChildNotes(noteId: string, maxChildren: number = this.maxBreadth): NoteInfo[] {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note || !note.children) {
|
||||
return [];
|
||||
}
|
||||
|
||||
return note.children
|
||||
.slice(0, maxChildren)
|
||||
.map(child => this.getNoteInfo(child.noteId))
|
||||
.filter((info): info is NoteInfo => info !== null);
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting child notes: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get a note's hierarchy (children up to specified depth)
|
||||
* This is useful for the LLM to understand the structure within a note's subtree
|
||||
*/
|
||||
getNoteHierarchy(noteId: string, depth: number = 2): NoteHierarchyLevel | null {
|
||||
if (depth < 0 || depth > this.maxDepth) {
|
||||
depth = this.maxDepth;
|
||||
}
|
||||
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return null;
|
||||
}
|
||||
|
||||
const result: NoteHierarchyLevel = {
|
||||
noteId: note.noteId,
|
||||
title: note.title,
|
||||
level: 0
|
||||
};
|
||||
|
||||
// Recursively get children if depth allows
|
||||
if (depth > 0 && note.children.length > 0) {
|
||||
result.children = note.children
|
||||
.slice(0, this.maxBreadth)
|
||||
.map(child => this._getHierarchyLevel(child.noteId, 1, depth))
|
||||
.filter((node): node is NoteHierarchyLevel => node !== null);
|
||||
}
|
||||
|
||||
return result;
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note hierarchy: ${error.message}`);
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Recursive helper for getNoteHierarchy
|
||||
*/
|
||||
private _getHierarchyLevel(noteId: string, currentLevel: number, maxDepth: number): NoteHierarchyLevel | null {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return null;
|
||||
}
|
||||
|
||||
const result: NoteHierarchyLevel = {
|
||||
noteId: note.noteId,
|
||||
title: note.title,
|
||||
level: currentLevel
|
||||
};
|
||||
|
||||
// Recursively get children if depth allows
|
||||
if (currentLevel < maxDepth && note.children.length > 0) {
|
||||
result.children = note.children
|
||||
.slice(0, this.maxBreadth)
|
||||
.map(child => this._getHierarchyLevel(child.noteId, currentLevel + 1, maxDepth))
|
||||
.filter((node): node is NoteHierarchyLevel => node !== null);
|
||||
}
|
||||
|
||||
return result;
|
||||
} catch (error) {
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get attributes of a note
|
||||
*/
|
||||
getNoteAttributes(noteId: string): BAttribute[] {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return [];
|
||||
}
|
||||
|
||||
return note.ownedAttributes;
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note attributes: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Find the shortest path between two notes
|
||||
*/
|
||||
findPathBetweenNotes(fromNoteId: string, toNoteId: string): NotePathInfo | null {
|
||||
try {
|
||||
if (fromNoteId === toNoteId) {
|
||||
const note = becca.notes[fromNoteId];
|
||||
if (!note) return null;
|
||||
|
||||
return {
|
||||
notePath: [fromNoteId],
|
||||
notePathTitles: [note.title]
|
||||
};
|
||||
}
|
||||
|
||||
// Simple breadth-first search to find shortest path
|
||||
const visited = new Set<string>();
|
||||
const queue: Array<{noteId: string, path: string[], titles: string[]}> = [];
|
||||
|
||||
// Initialize with the starting note
|
||||
const startNote = becca.notes[fromNoteId];
|
||||
if (!startNote) return null;
|
||||
|
||||
queue.push({
|
||||
noteId: fromNoteId,
|
||||
path: [fromNoteId],
|
||||
titles: [startNote.title]
|
||||
});
|
||||
|
||||
visited.add(fromNoteId);
|
||||
|
||||
while (queue.length > 0 && queue[0].path.length <= this.maxPathLength) {
|
||||
const {noteId, path, titles} = queue.shift()!;
|
||||
const note = becca.notes[noteId];
|
||||
|
||||
if (!note) continue;
|
||||
|
||||
// Get IDs of all connected notes (parents and children)
|
||||
const connections: string[] = [
|
||||
...note.parents.map(p => p.noteId),
|
||||
...note.children.map(c => c.noteId)
|
||||
];
|
||||
|
||||
for (const connectedId of connections) {
|
||||
if (visited.has(connectedId)) continue;
|
||||
|
||||
const connectedNote = becca.notes[connectedId];
|
||||
if (!connectedNote) continue;
|
||||
|
||||
const newPath = [...path, connectedId];
|
||||
const newTitles = [...titles, connectedNote.title];
|
||||
|
||||
// Check if we found the target
|
||||
if (connectedId === toNoteId) {
|
||||
return {
|
||||
notePath: newPath,
|
||||
notePathTitles: newTitles
|
||||
};
|
||||
}
|
||||
|
||||
// Continue BFS
|
||||
queue.push({
|
||||
noteId: connectedId,
|
||||
path: newPath,
|
||||
titles: newTitles
|
||||
});
|
||||
|
||||
visited.add(connectedId);
|
||||
}
|
||||
}
|
||||
|
||||
// No path found
|
||||
return null;
|
||||
} catch (error: any) {
|
||||
log.error(`Error finding path between notes: ${error.message}`);
|
||||
return null;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Search for notes by title
|
||||
*/
|
||||
searchNotesByTitle(searchTerm: string, limit: number = 10): NoteInfo[] {
|
||||
try {
|
||||
if (!searchTerm || searchTerm.trim().length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
searchTerm = searchTerm.toLowerCase();
|
||||
const results: NoteInfo[] = [];
|
||||
|
||||
// Simple in-memory search through all notes
|
||||
for (const noteId in becca.notes) {
|
||||
if (results.length >= limit) break;
|
||||
|
||||
const note = becca.notes[noteId];
|
||||
if (!note || note.isDeleted) continue;
|
||||
|
||||
if (note.title.toLowerCase().includes(searchTerm)) {
|
||||
const info = this.getNoteInfo(noteId);
|
||||
if (info) results.push(info);
|
||||
}
|
||||
}
|
||||
|
||||
return results;
|
||||
} catch (error: any) {
|
||||
log.error(`Error searching notes by title: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get clones of a note (if any)
|
||||
*/
|
||||
getNoteClones(noteId: string): NoteInfo[] {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// A note has clones if it has multiple parents
|
||||
if (note.parents.length <= 1) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// Return parent notes, which represent different contexts for this note
|
||||
return this.getParentNotes(noteId);
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note clones: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a readable overview of a note's position in the hierarchy
|
||||
* This is useful for the LLM to understand the context of a note
|
||||
*/
|
||||
getNoteContextDescription(noteId: string): string {
|
||||
try {
|
||||
const note = becca.notes[noteId];
|
||||
if (!note) {
|
||||
return "Note not found.";
|
||||
}
|
||||
|
||||
const paths = this.getNotePathsFromRoot(noteId);
|
||||
if (paths.length === 0) {
|
||||
return `Note "${note.title}" exists but has no path from root.`;
|
||||
}
|
||||
|
||||
let result = "";
|
||||
|
||||
// Basic note info
|
||||
result += `Note: "${note.title}" (${note.type})\n`;
|
||||
|
||||
// Is it cloned?
|
||||
if (paths.length > 1) {
|
||||
result += `This note appears in ${paths.length} different locations:\n`;
|
||||
|
||||
// Show max 3 paths to avoid overwhelming context
|
||||
for (let i = 0; i < Math.min(3, paths.length); i++) {
|
||||
const path = paths[i];
|
||||
result += `${i+1}. ${path.notePathTitles.join(' > ')}\n`;
|
||||
}
|
||||
|
||||
if (paths.length > 3) {
|
||||
result += `... and ${paths.length - 3} more locations\n`;
|
||||
}
|
||||
} else {
|
||||
// Just one path
|
||||
const path = paths[0];
|
||||
result += `Path: ${path.notePathTitles.join(' > ')}\n`;
|
||||
}
|
||||
|
||||
// Children info
|
||||
const children = this.getChildNotes(noteId, 5);
|
||||
if (children.length > 0) {
|
||||
result += `\nContains ${note.children.length} child notes`;
|
||||
if (children.length < note.children.length) {
|
||||
result += ` (showing first ${children.length})`;
|
||||
}
|
||||
result += `:\n`;
|
||||
|
||||
for (const child of children) {
|
||||
result += `- ${child.title} (${child.type})\n`;
|
||||
}
|
||||
|
||||
if (children.length < note.children.length) {
|
||||
result += `... and ${note.children.length - children.length} more\n`;
|
||||
}
|
||||
} else {
|
||||
result += "\nThis note has no child notes.\n";
|
||||
}
|
||||
|
||||
// Attributes summary
|
||||
const attributes = this.getNoteAttributes(noteId);
|
||||
if (attributes.length > 0) {
|
||||
result += `\nNote has ${attributes.length} attributes.\n`;
|
||||
|
||||
// Group attributes by name
|
||||
const attrMap: Record<string, string[]> = {};
|
||||
for (const attr of attributes) {
|
||||
if (!attrMap[attr.name]) {
|
||||
attrMap[attr.name] = [];
|
||||
}
|
||||
attrMap[attr.name].push(attr.value);
|
||||
}
|
||||
|
||||
for (const [name, values] of Object.entries(attrMap)) {
|
||||
if (values.length === 1) {
|
||||
result += `- ${name}: ${values[0]}\n`;
|
||||
} else {
|
||||
result += `- ${name}: ${values.length} values\n`;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note context: ${error.message}`);
|
||||
return "Error generating note context description.";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
export default NoteNavigatorTool;
|
||||
390
src/services/llm/agent_tools/query_decomposition_tool.ts
Normal file
390
src/services/llm/agent_tools/query_decomposition_tool.ts
Normal file
@ -0,0 +1,390 @@
|
||||
/**
|
||||
* Query Decomposition Tool
|
||||
*
|
||||
* This tool helps the LLM agent break down complex user queries into
|
||||
* sub-questions that can be answered individually and then synthesized
|
||||
* into a comprehensive response.
|
||||
*
|
||||
* Features:
|
||||
* - Analyze query complexity
|
||||
* - Extract multiple intents from a single question
|
||||
* - Create a multi-stage research plan
|
||||
* - Track progress through complex information gathering
|
||||
*/
|
||||
|
||||
import log from '../../log.js';
|
||||
|
||||
export interface SubQuery {
|
||||
id: string;
|
||||
text: string;
|
||||
reason: string;
|
||||
isAnswered: boolean;
|
||||
answer?: string;
|
||||
}
|
||||
|
||||
export interface DecomposedQuery {
|
||||
originalQuery: string;
|
||||
subQueries: SubQuery[];
|
||||
status: 'pending' | 'in_progress' | 'completed';
|
||||
complexity: number;
|
||||
}
|
||||
|
||||
export class QueryDecompositionTool {
|
||||
private static queryCounter: number = 0;
|
||||
|
||||
/**
|
||||
* Break down a complex query into smaller, more manageable sub-queries
|
||||
*
|
||||
* @param query The original user query
|
||||
* @param context Optional context about the current note being viewed
|
||||
* @returns A decomposed query object with sub-queries
|
||||
*/
|
||||
decomposeQuery(query: string, context?: string): DecomposedQuery {
|
||||
try {
|
||||
// Assess query complexity to determine if decomposition is needed
|
||||
const complexity = this.assessQueryComplexity(query);
|
||||
|
||||
// For simple queries, just return the original as a single sub-query
|
||||
if (complexity < 3) {
|
||||
return {
|
||||
originalQuery: query,
|
||||
subQueries: [{
|
||||
id: this.generateSubQueryId(),
|
||||
text: query,
|
||||
reason: 'Direct question that can be answered without decomposition',
|
||||
isAnswered: false
|
||||
}],
|
||||
status: 'pending',
|
||||
complexity
|
||||
};
|
||||
}
|
||||
|
||||
// For complex queries, perform decomposition
|
||||
const subQueries = this.createSubQueries(query, context);
|
||||
|
||||
return {
|
||||
originalQuery: query,
|
||||
subQueries,
|
||||
status: 'pending',
|
||||
complexity
|
||||
};
|
||||
} catch (error: any) {
|
||||
log.error(`Error decomposing query: ${error.message}`);
|
||||
|
||||
// Fallback to treating it as a simple query
|
||||
return {
|
||||
originalQuery: query,
|
||||
subQueries: [{
|
||||
id: this.generateSubQueryId(),
|
||||
text: query,
|
||||
reason: 'Error in decomposition, treating as simple query',
|
||||
isAnswered: false
|
||||
}],
|
||||
status: 'pending',
|
||||
complexity: 1
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Update a sub-query with its answer
|
||||
*
|
||||
* @param decomposedQuery The decomposed query object
|
||||
* @param subQueryId The ID of the sub-query to update
|
||||
* @param answer The answer to the sub-query
|
||||
* @returns The updated decomposed query
|
||||
*/
|
||||
updateSubQueryAnswer(
|
||||
decomposedQuery: DecomposedQuery,
|
||||
subQueryId: string,
|
||||
answer: string
|
||||
): DecomposedQuery {
|
||||
const updatedSubQueries = decomposedQuery.subQueries.map(sq => {
|
||||
if (sq.id === subQueryId) {
|
||||
return {
|
||||
...sq,
|
||||
answer,
|
||||
isAnswered: true
|
||||
};
|
||||
}
|
||||
return sq;
|
||||
});
|
||||
|
||||
// Check if all sub-queries are answered
|
||||
const allAnswered = updatedSubQueries.every(sq => sq.isAnswered);
|
||||
|
||||
return {
|
||||
...decomposedQuery,
|
||||
subQueries: updatedSubQueries,
|
||||
status: allAnswered ? 'completed' : 'in_progress'
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* Synthesize all sub-query answers into a comprehensive response
|
||||
*
|
||||
* @param decomposedQuery The decomposed query with all sub-queries answered
|
||||
* @returns A synthesized answer to the original query
|
||||
*/
|
||||
synthesizeAnswer(decomposedQuery: DecomposedQuery): string {
|
||||
try {
|
||||
// Ensure all sub-queries are answered
|
||||
if (!decomposedQuery.subQueries.every(sq => sq.isAnswered)) {
|
||||
return "Cannot synthesize answer - not all sub-queries have been answered.";
|
||||
}
|
||||
|
||||
// For simple queries with just one sub-query, return the answer directly
|
||||
if (decomposedQuery.subQueries.length === 1) {
|
||||
return decomposedQuery.subQueries[0].answer || "";
|
||||
}
|
||||
|
||||
// For complex queries, build a structured response that references each sub-answer
|
||||
let synthesized = `Answer to: "${decomposedQuery.originalQuery}"\n\n`;
|
||||
|
||||
// Group by themes if there are many sub-queries
|
||||
if (decomposedQuery.subQueries.length > 3) {
|
||||
// Here we would ideally group related sub-queries, but for now we'll just present them in order
|
||||
synthesized += "Based on the information gathered:\n\n";
|
||||
|
||||
for (const sq of decomposedQuery.subQueries) {
|
||||
synthesized += `${sq.answer}\n\n`;
|
||||
}
|
||||
} else {
|
||||
// For fewer sub-queries, present each one with its question
|
||||
for (const sq of decomposedQuery.subQueries) {
|
||||
synthesized += `${sq.answer}\n\n`;
|
||||
}
|
||||
}
|
||||
|
||||
return synthesized.trim();
|
||||
} catch (error: any) {
|
||||
log.error(`Error synthesizing answer: ${error.message}`);
|
||||
return "Error synthesizing the final answer.";
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a status report on the progress of answering a complex query
|
||||
*
|
||||
* @param decomposedQuery The decomposed query
|
||||
* @returns A status report string
|
||||
*/
|
||||
getQueryStatus(decomposedQuery: DecomposedQuery): string {
|
||||
const answeredCount = decomposedQuery.subQueries.filter(sq => sq.isAnswered).length;
|
||||
const totalCount = decomposedQuery.subQueries.length;
|
||||
|
||||
let status = `Progress: ${answeredCount}/${totalCount} sub-queries answered\n\n`;
|
||||
|
||||
for (const sq of decomposedQuery.subQueries) {
|
||||
status += `${sq.isAnswered ? '✓' : '○'} ${sq.text}\n`;
|
||||
if (sq.isAnswered) {
|
||||
status += ` Answer: ${this.truncateText(sq.answer || "", 100)}\n`;
|
||||
}
|
||||
}
|
||||
|
||||
return status;
|
||||
}
|
||||
|
||||
/**
|
||||
* Assess the complexity of a query on a scale of 1-10
|
||||
* This helps determine how many sub-queries are needed
|
||||
*
|
||||
* @param query The query to assess
|
||||
* @returns A complexity score from 1-10
|
||||
*/
|
||||
assessQueryComplexity(query: string): number {
|
||||
// Count the number of question marks as a basic indicator
|
||||
const questionMarkCount = (query.match(/\?/g) || []).length;
|
||||
|
||||
// Count potential sub-questions based on question words
|
||||
const questionWords = ['what', 'how', 'why', 'where', 'when', 'who', 'which'];
|
||||
const questionWordMatches = questionWords.map(word => {
|
||||
const regex = new RegExp(`\\b${word}\\b`, 'gi');
|
||||
return (query.match(regex) || []).length;
|
||||
});
|
||||
|
||||
const questionWordCount = questionWordMatches.reduce((sum, count) => sum + count, 0);
|
||||
|
||||
// Look for conjunctions which might join multiple questions
|
||||
const conjunctionCount = (query.match(/\b(and|or|but|as well as)\b/gi) || []).length;
|
||||
|
||||
// Look for complex requirements
|
||||
const comparisonCount = (query.match(/\b(compare|versus|vs|difference|similarities?)\b/gi) || []).length;
|
||||
const analysisCount = (query.match(/\b(analyze|examine|investigate|explore|explain|discuss)\b/gi) || []).length;
|
||||
|
||||
// Calculate base complexity
|
||||
let complexity = 1;
|
||||
|
||||
// Add for multiple questions
|
||||
complexity += Math.min(2, questionMarkCount);
|
||||
|
||||
// Add for question words beyond the first one
|
||||
complexity += Math.min(2, Math.max(0, questionWordCount - 1));
|
||||
|
||||
// Add for conjunctions that might join questions
|
||||
complexity += Math.min(2, conjunctionCount);
|
||||
|
||||
// Add for comparative/analytical requirements
|
||||
complexity += Math.min(2, comparisonCount + analysisCount);
|
||||
|
||||
// Add for overall length/complexity
|
||||
if (query.length > 100) complexity += 1;
|
||||
if (query.length > 200) complexity += 1;
|
||||
|
||||
// Ensure we stay in the 1-10 range
|
||||
return Math.max(1, Math.min(10, complexity));
|
||||
}
|
||||
|
||||
/**
|
||||
* Generate a unique ID for a sub-query
|
||||
*/
|
||||
private generateSubQueryId(): string {
|
||||
return `sq_${Date.now()}_${QueryDecompositionTool.queryCounter++}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Create sub-queries based on the original query and optional context
|
||||
*/
|
||||
private createSubQueries(query: string, context?: string): SubQuery[] {
|
||||
const subQueries: SubQuery[] = [];
|
||||
|
||||
// Simple heuristics for breaking down the query
|
||||
// In a real implementation, this would be much more sophisticated,
|
||||
// using natural language understanding to identify different intents
|
||||
|
||||
// 1. Look for multiple question marks
|
||||
const questionSplit = query.split(/\?/).filter(q => q.trim().length > 0);
|
||||
|
||||
if (questionSplit.length > 1) {
|
||||
// Multiple distinct questions detected
|
||||
for (let i = 0; i < questionSplit.length; i++) {
|
||||
const text = questionSplit[i].trim() + '?';
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text,
|
||||
reason: `Separate question ${i+1} detected in the original query`,
|
||||
isAnswered: false
|
||||
});
|
||||
}
|
||||
return subQueries;
|
||||
}
|
||||
|
||||
// 2. Look for "and", "or", etc. connecting potentially separate questions
|
||||
const conjunctions = [
|
||||
{ regex: /\b(compare|versus|vs\.?|difference between|similarities between)\b/i, label: 'comparison' },
|
||||
{ regex: /\b(list|enumerate)\b/i, label: 'listing' },
|
||||
{ regex: /\b(analyze|examine|investigate|explore)\b/i, label: 'analysis' },
|
||||
{ regex: /\b(explain|why)\b/i, label: 'explanation' },
|
||||
{ regex: /\b(how to|steps to|process of)\b/i, label: 'procedure' }
|
||||
];
|
||||
|
||||
// Check for comparison queries - these often need multiple sub-queries
|
||||
for (const conj of conjunctions) {
|
||||
if (conj.regex.test(query)) {
|
||||
if (conj.label === 'comparison') {
|
||||
// For comparisons, we need to research each item, then compare them
|
||||
const comparisonMatch = query.match(/\b(compare|versus|vs\.?|difference between|similarities between)\s+(.+?)\s+(and|with|to)\s+(.+?)(\?|$)/i);
|
||||
|
||||
if (comparisonMatch) {
|
||||
const item1 = comparisonMatch[2].trim();
|
||||
const item2 = comparisonMatch[4].trim();
|
||||
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: `What are the key characteristics of ${item1}?`,
|
||||
reason: `Need to understand ${item1} for the comparison`,
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: `What are the key characteristics of ${item2}?`,
|
||||
reason: `Need to understand ${item2} for the comparison`,
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: `What are the main differences and similarities between ${item1} and ${item2}?`,
|
||||
reason: 'Direct comparison after understanding each item',
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
return subQueries;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 3. For complex questions without clear separation, create topic-based sub-queries
|
||||
if (query.length > 100) {
|
||||
// Extract potential key topics from the query
|
||||
const words = query.toLowerCase().split(/\W+/).filter(w =>
|
||||
w.length > 3 &&
|
||||
!['what', 'when', 'where', 'which', 'with', 'would', 'could', 'should', 'have', 'this', 'that', 'there', 'their'].includes(w)
|
||||
);
|
||||
|
||||
// Count word frequencies
|
||||
const wordFrequency: Record<string, number> = {};
|
||||
for (const word of words) {
|
||||
wordFrequency[word] = (wordFrequency[word] || 0) + 1;
|
||||
}
|
||||
|
||||
// Get top frequent words
|
||||
const topWords = Object.entries(wordFrequency)
|
||||
.sort((a, b) => b[1] - a[1])
|
||||
.slice(0, 3)
|
||||
.map(entry => entry[0]);
|
||||
|
||||
if (topWords.length > 0) {
|
||||
// Create factual sub-query
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: `What are the key facts about ${topWords.join(' and ')} relevant to this question?`,
|
||||
reason: 'Gathering basic information about main topics',
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
// Create relationship sub-query if multiple top words
|
||||
if (topWords.length > 1) {
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: `How do ${topWords.join(' and ')} relate to each other?`,
|
||||
reason: 'Understanding relationships between key topics',
|
||||
isAnswered: false
|
||||
});
|
||||
}
|
||||
|
||||
// Add the original query as the final synthesizing question
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: query,
|
||||
reason: 'Original question to be answered after gathering information',
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
return subQueries;
|
||||
}
|
||||
}
|
||||
|
||||
// Fallback: If we can't meaningfully decompose, just use the original query
|
||||
subQueries.push({
|
||||
id: this.generateSubQueryId(),
|
||||
text: query,
|
||||
reason: 'Question treated as a single unit',
|
||||
isAnswered: false
|
||||
});
|
||||
|
||||
return subQueries;
|
||||
}
|
||||
|
||||
/**
|
||||
* Truncate text to a maximum length with ellipsis
|
||||
*/
|
||||
private truncateText(text: string, maxLength: number): string {
|
||||
if (text.length <= maxLength) return text;
|
||||
return text.substring(0, maxLength - 3) + '...';
|
||||
}
|
||||
}
|
||||
|
||||
export default QueryDecompositionTool;
|
||||
185
src/services/llm/agent_tools/vector_search_tool.ts
Normal file
185
src/services/llm/agent_tools/vector_search_tool.ts
Normal file
@ -0,0 +1,185 @@
|
||||
/**
|
||||
* Vector Search Tool
|
||||
*
|
||||
* This tool enables the LLM agent to perform semantic vector-based searches
|
||||
* over the content in the notes database. It handles:
|
||||
* - Finding semantically related notes to a query
|
||||
* - Extracting relevant sections from notes
|
||||
* - Providing relevant context for LLM to generate accurate responses
|
||||
*
|
||||
* The tool uses embeddings to find notes with similar semantic meaning,
|
||||
* allowing the LLM to find relevant information even when exact keywords
|
||||
* are not present.
|
||||
*/
|
||||
|
||||
import log from '../../log.js';
|
||||
|
||||
// Define interface for semantic context service to avoid circular imports
|
||||
interface ISemanticContextService {
|
||||
semanticSearch(query: string, options: any): Promise<any[]>;
|
||||
semanticSearchChunks(query: string, options: any): Promise<any[]>;
|
||||
}
|
||||
|
||||
export interface VectorSearchResult {
|
||||
noteId: string;
|
||||
title: string;
|
||||
contentPreview: string;
|
||||
similarity: number;
|
||||
parentId?: string;
|
||||
dateCreated?: string;
|
||||
dateModified?: string;
|
||||
}
|
||||
|
||||
export interface SearchResultItem {
|
||||
noteId: string;
|
||||
noteTitle: string;
|
||||
contentPreview: string;
|
||||
similarity: number;
|
||||
parentId?: string;
|
||||
dateCreated?: string;
|
||||
dateModified?: string;
|
||||
}
|
||||
|
||||
export interface ChunkSearchResultItem {
|
||||
noteId: string;
|
||||
noteTitle: string;
|
||||
chunk: string;
|
||||
similarity: number;
|
||||
parentId?: string;
|
||||
}
|
||||
|
||||
export class VectorSearchTool {
|
||||
private semanticContext: ISemanticContextService | null = null;
|
||||
private maxResults: number = 5;
|
||||
|
||||
constructor() {
|
||||
// The semantic context will be set later via setSemanticContext
|
||||
}
|
||||
|
||||
/**
|
||||
* Set the semantic context service instance
|
||||
*/
|
||||
setSemanticContext(semanticContext: ISemanticContextService): void {
|
||||
this.semanticContext = semanticContext;
|
||||
}
|
||||
|
||||
/**
|
||||
* Search for notes semantically related to a query
|
||||
*/
|
||||
async searchNotes(query: string, options: {
|
||||
parentNoteId?: string,
|
||||
maxResults?: number,
|
||||
similarityThreshold?: number
|
||||
} = {}): Promise<VectorSearchResult[]> {
|
||||
try {
|
||||
if (!this.semanticContext) {
|
||||
throw new Error("Semantic context service not set. Call setSemanticContext() first.");
|
||||
}
|
||||
|
||||
if (!query || query.trim().length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const maxResults = options.maxResults || this.maxResults;
|
||||
const similarityThreshold = options.similarityThreshold || 0.65; // Default threshold
|
||||
const parentNoteId = options.parentNoteId; // Optional filtering by parent
|
||||
|
||||
// Search notes using the semantic context service
|
||||
const results = await this.semanticContext.semanticSearch(query, {
|
||||
maxResults,
|
||||
similarityThreshold,
|
||||
ancestorNoteId: parentNoteId
|
||||
});
|
||||
|
||||
if (!results || results.length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// Transform results to the tool's format
|
||||
return results.map((result: SearchResultItem) => ({
|
||||
noteId: result.noteId,
|
||||
title: result.noteTitle,
|
||||
contentPreview: result.contentPreview,
|
||||
similarity: result.similarity,
|
||||
parentId: result.parentId,
|
||||
dateCreated: result.dateCreated,
|
||||
dateModified: result.dateModified
|
||||
}));
|
||||
} catch (error: any) {
|
||||
log.error(`Error in vector search: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Search for content chunks within notes that are semantically related to a query
|
||||
*/
|
||||
async searchContentChunks(query: string, options: {
|
||||
noteId?: string,
|
||||
maxResults?: number,
|
||||
similarityThreshold?: number
|
||||
} = {}): Promise<VectorSearchResult[]> {
|
||||
try {
|
||||
if (!this.semanticContext) {
|
||||
throw new Error("Semantic context service not set. Call setSemanticContext() first.");
|
||||
}
|
||||
|
||||
if (!query || query.trim().length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const maxResults = options.maxResults || this.maxResults;
|
||||
const similarityThreshold = options.similarityThreshold || 0.70; // Higher threshold for chunks
|
||||
const noteId = options.noteId; // Optional filtering by specific note
|
||||
|
||||
// Search content chunks using the semantic context service
|
||||
const results = await this.semanticContext.semanticSearchChunks(query, {
|
||||
maxResults,
|
||||
similarityThreshold,
|
||||
noteId
|
||||
});
|
||||
|
||||
if (!results || results.length === 0) {
|
||||
return [];
|
||||
}
|
||||
|
||||
// Transform results to the tool's format
|
||||
return results.map((result: ChunkSearchResultItem) => ({
|
||||
noteId: result.noteId,
|
||||
title: result.noteTitle,
|
||||
contentPreview: result.chunk, // Use the chunk content as preview
|
||||
similarity: result.similarity,
|
||||
parentId: result.parentId
|
||||
}));
|
||||
} catch (error: any) {
|
||||
log.error(`Error in content chunk search: ${error.message}`);
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Elaborate on why certain results were returned for a query
|
||||
*/
|
||||
explainResults(query: string, results: VectorSearchResult[]): string {
|
||||
if (!query || !results || results.length === 0) {
|
||||
return "No results to explain.";
|
||||
}
|
||||
|
||||
let explanation = `For query "${query}", I found these semantically related notes:\n\n`;
|
||||
|
||||
results.forEach((result, index) => {
|
||||
explanation += `${index + 1}. "${result.title}" (similarity: ${(result.similarity * 100).toFixed(1)}%)\n`;
|
||||
explanation += ` Preview: ${result.contentPreview.substring(0, 150)}...\n`;
|
||||
|
||||
if (index < results.length - 1) {
|
||||
explanation += "\n";
|
||||
}
|
||||
});
|
||||
|
||||
explanation += "\nThese results were found based on semantic similarity rather than just keyword matching.";
|
||||
|
||||
return explanation;
|
||||
}
|
||||
}
|
||||
|
||||
export default VectorSearchTool;
|
||||
@ -8,6 +8,7 @@ import { ContextExtractor } from './context/index.js';
|
||||
import semanticContextService from './semantic_context_service.js';
|
||||
import indexService from './index_service.js';
|
||||
import { getEmbeddingProvider, getEnabledEmbeddingProviders } from './embeddings/providers.js';
|
||||
import agentTools from './agent_tools/index.js';
|
||||
|
||||
type ServiceProviders = 'openai' | 'anthropic' | 'ollama';
|
||||
|
||||
@ -281,6 +282,128 @@ export class AIServiceManager {
|
||||
getIndexService() {
|
||||
return indexService;
|
||||
}
|
||||
|
||||
/**
|
||||
* Initialize agent tools for enhanced LLM features
|
||||
*/
|
||||
async initializeAgentTools(): Promise<void> {
|
||||
try {
|
||||
await agentTools.initialize(this);
|
||||
log.info("Agent tools initialized successfully");
|
||||
} catch (error: any) {
|
||||
log.error(`Error initializing agent tools: ${error.message}`);
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the agent tools manager
|
||||
* This provides access to all agent tools
|
||||
*/
|
||||
getAgentTools() {
|
||||
return agentTools;
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the vector search tool for semantic similarity search
|
||||
*/
|
||||
getVectorSearchTool() {
|
||||
return agentTools.getVectorSearchTool();
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the note navigator tool for hierarchical exploration
|
||||
*/
|
||||
getNoteNavigatorTool() {
|
||||
return agentTools.getNoteNavigatorTool();
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the query decomposition tool for complex queries
|
||||
*/
|
||||
getQueryDecompositionTool() {
|
||||
return agentTools.getQueryDecompositionTool();
|
||||
}
|
||||
|
||||
/**
|
||||
* Get the contextual thinking tool for transparent reasoning
|
||||
*/
|
||||
getContextualThinkingTool() {
|
||||
return agentTools.getContextualThinkingTool();
|
||||
}
|
||||
|
||||
/**
|
||||
* Get whether AI features are enabled from options
|
||||
*/
|
||||
getAIEnabled(): boolean {
|
||||
return options.getOptionBool('aiEnabled');
|
||||
}
|
||||
|
||||
/**
|
||||
* Set up embeddings provider for AI features
|
||||
*/
|
||||
async setupEmbeddingsProvider(): Promise<void> {
|
||||
try {
|
||||
if (!this.getAIEnabled()) {
|
||||
log.info('AI features are disabled');
|
||||
return;
|
||||
}
|
||||
|
||||
const preferredProvider = options.getOption('embeddingsDefaultProvider') || 'openai';
|
||||
|
||||
// Check if we have enabled providers
|
||||
const enabledProviders = await getEnabledEmbeddingProviders();
|
||||
|
||||
if (enabledProviders.length === 0) {
|
||||
log.info('No embedding providers are enabled');
|
||||
return;
|
||||
}
|
||||
|
||||
// Validate that preferred provider is enabled
|
||||
const isPreferredEnabled = enabledProviders.some(p => p.name === preferredProvider);
|
||||
|
||||
if (!isPreferredEnabled) {
|
||||
log.info(`Preferred provider "${preferredProvider}" is not enabled. Using first available.`);
|
||||
}
|
||||
|
||||
// Initialize embedding providers
|
||||
log.info('Embedding providers initialized successfully');
|
||||
} catch (error: any) {
|
||||
log.error(`Error setting up embedding providers: ${error.message}`);
|
||||
throw error;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Initialize the AI Service
|
||||
*/
|
||||
async initialize(): Promise<void> {
|
||||
try {
|
||||
log.info("Initializing AI service...");
|
||||
|
||||
// Check if AI is enabled in options
|
||||
const isAIEnabled = this.getAIEnabled();
|
||||
|
||||
if (!isAIEnabled) {
|
||||
log.info("AI features are disabled in options");
|
||||
return;
|
||||
}
|
||||
|
||||
// Set up embeddings provider if AI is enabled
|
||||
await this.setupEmbeddingsProvider();
|
||||
|
||||
// Initialize index service
|
||||
await this.getIndexService().initialize();
|
||||
|
||||
// Initialize agent tools with this service manager instance
|
||||
await agentTools.initialize(this);
|
||||
|
||||
this.initialized = true;
|
||||
log.info("AI service initialized successfully");
|
||||
} catch (error: any) {
|
||||
log.error(`Error initializing AI service: ${error.message}`);
|
||||
throw error;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Don't create singleton immediately, use a lazy-loading pattern
|
||||
@ -321,6 +444,26 @@ export default {
|
||||
},
|
||||
getIndexService() {
|
||||
return getInstance().getIndexService();
|
||||
},
|
||||
// Agent tools related methods
|
||||
async initializeAgentTools(): Promise<void> {
|
||||
const manager = getInstance();
|
||||
return manager.initializeAgentTools();
|
||||
},
|
||||
getAgentTools() {
|
||||
return getInstance().getAgentTools();
|
||||
},
|
||||
getVectorSearchTool() {
|
||||
return getInstance().getVectorSearchTool();
|
||||
},
|
||||
getNoteNavigatorTool() {
|
||||
return getInstance().getNoteNavigatorTool();
|
||||
},
|
||||
getQueryDecompositionTool() {
|
||||
return getInstance().getQueryDecompositionTool();
|
||||
},
|
||||
getContextualThinkingTool() {
|
||||
return getInstance().getContextualThinkingTool();
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
@ -6,6 +6,7 @@ import log from "../log.js";
|
||||
import type { Message } from "./ai_interface.js";
|
||||
import { cosineSimilarity } from "./embeddings/index.js";
|
||||
import sanitizeHtml from "sanitize-html";
|
||||
import aiServiceManager from "./ai_service_manager.js";
|
||||
|
||||
/**
|
||||
* TriliumContextService provides intelligent context management for working with large knowledge bases
|
||||
@ -617,6 +618,140 @@ Example: ["exact topic mentioned", "related concept 1", "related concept 2"]`;
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Enhance LLM context with agent tools
|
||||
*
|
||||
* This adds context from agent tools such as:
|
||||
* 1. Vector search results relevant to the query
|
||||
* 2. Note hierarchy information
|
||||
* 3. Query decomposition planning
|
||||
* 4. Contextual thinking visualization
|
||||
*
|
||||
* @param noteId The current note being viewed
|
||||
* @param query The user's query
|
||||
* @param showThinking Whether to include the agent's thinking process
|
||||
* @returns Enhanced context string
|
||||
*/
|
||||
async getAgentToolsContext(noteId: string, query: string, showThinking: boolean = false): Promise<string> {
|
||||
try {
|
||||
const agentTools = aiServiceManager.getAgentTools();
|
||||
let context = "";
|
||||
|
||||
// 1. Get vector search results related to the query
|
||||
try {
|
||||
const vectorSearchTool = agentTools.getVectorSearchTool();
|
||||
const searchResults = await vectorSearchTool.searchNotes(query, {
|
||||
parentNoteId: noteId,
|
||||
maxResults: 5
|
||||
});
|
||||
|
||||
if (searchResults.length > 0) {
|
||||
context += "## Related Information\n\n";
|
||||
for (const result of searchResults) {
|
||||
context += `### ${result.title}\n`;
|
||||
context += `${result.contentPreview}\n\n`;
|
||||
}
|
||||
context += "\n";
|
||||
}
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting vector search context: ${error.message}`);
|
||||
}
|
||||
|
||||
// 2. Get note structure context
|
||||
try {
|
||||
const navigatorTool = agentTools.getNoteNavigatorTool();
|
||||
const noteContext = navigatorTool.getNoteContextDescription(noteId);
|
||||
|
||||
if (noteContext) {
|
||||
context += "## Current Note Context\n\n";
|
||||
context += noteContext + "\n\n";
|
||||
}
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting note structure context: ${error.message}`);
|
||||
}
|
||||
|
||||
// 3. Use query decomposition if it's a complex query
|
||||
try {
|
||||
const decompositionTool = agentTools.getQueryDecompositionTool();
|
||||
const complexity = decompositionTool.assessQueryComplexity(query);
|
||||
|
||||
if (complexity > 5) { // Only for fairly complex queries
|
||||
const decomposed = decompositionTool.decomposeQuery(query);
|
||||
|
||||
if (decomposed.subQueries.length > 1) {
|
||||
context += "## Query Analysis\n\n";
|
||||
context += `This is a complex query (complexity: ${complexity}/10). It can be broken down into:\n\n`;
|
||||
|
||||
for (const sq of decomposed.subQueries) {
|
||||
context += `- ${sq.text}\n Reason: ${sq.reason}\n\n`;
|
||||
}
|
||||
}
|
||||
}
|
||||
} catch (error: any) {
|
||||
log.error(`Error decomposing query: ${error.message}`);
|
||||
}
|
||||
|
||||
// 4. Show thinking process if enabled
|
||||
if (showThinking) {
|
||||
try {
|
||||
const thinkingTool = agentTools.getContextualThinkingTool();
|
||||
const thinkingId = thinkingTool.startThinking(query);
|
||||
|
||||
// Add a thinking step to demonstrate the feature
|
||||
// In a real implementation, the LLM would add these steps
|
||||
thinkingTool.addThinkingStep(
|
||||
"Analyzing the query to understand what information is needed",
|
||||
"observation",
|
||||
{ confidence: 1.0 }
|
||||
);
|
||||
|
||||
// Add sample thinking for the context
|
||||
const parentId = thinkingTool.addThinkingStep(
|
||||
"Looking for related notes in the knowledge base",
|
||||
"hypothesis",
|
||||
{ confidence: 0.9 }
|
||||
);
|
||||
|
||||
if (parentId) {
|
||||
// Use the VectorSearchTool to find relevant notes
|
||||
const vectorSearchTool = aiServiceManager.getVectorSearchTool();
|
||||
const searchResults = await vectorSearchTool.searchNotes(query, {
|
||||
parentNoteId: parentId,
|
||||
maxResults: 5
|
||||
});
|
||||
|
||||
if (searchResults.length > 0) {
|
||||
context += "## Related Information\n\n";
|
||||
for (const result of searchResults) {
|
||||
context += `### ${result.title}\n`;
|
||||
context += `${result.contentPreview}\n\n`;
|
||||
}
|
||||
context += "\n";
|
||||
}
|
||||
}
|
||||
|
||||
thinkingTool.addThinkingStep(
|
||||
"The most relevant information appears to be in the current note and its semantic neighborhood",
|
||||
"conclusion",
|
||||
{ confidence: 0.85 }
|
||||
);
|
||||
|
||||
// Complete the thinking and add it to context
|
||||
thinkingTool.completeThinking(thinkingId);
|
||||
context += "## Thinking Process\n\n";
|
||||
context += thinkingTool.getThinkingSummary(thinkingId) + "\n\n";
|
||||
} catch (error: any) {
|
||||
log.error(`Error generating thinking process: ${error.message}`);
|
||||
}
|
||||
}
|
||||
|
||||
return context;
|
||||
} catch (error: any) {
|
||||
log.error(`Error getting agent tools context: ${error.message}`);
|
||||
return "";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
export default new TriliumContextService();
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user