feat: integrate unified RAG pipeline and bump version to 2.60.0

This commit is contained in:
g1nation
2026-05-04 11:00:01 +09:00
parent 0515dd625d
commit 445d530b63
16 changed files with 2178 additions and 112 deletions
+130
View File
@@ -0,0 +1,130 @@
/**
* ============================================================
* Context Budget Manager (컨텍스트 예산 관리)
*
* 시스템 프롬프트의 토큰 예산을 관리하여
* 로컬 모델의 context window를 효율적으로 활용합니다.
* ============================================================
*/
import { RetrievalChunk, ContextBudgetConfig } from './types';
const DEFAULT_BUDGET: ContextBudgetConfig = {
totalBudget: 8000, // ~32K context 중 retrieval에 할당
retrievalRatio: 0.4, // 40%
minChunks: 2,
maxChunks: 12
};
/**
* 토큰 수를 대략 추정합니다 (문자 수 / 4).
* 한국어는 글자당 토큰이 더 많으므로 보정합니다.
*/
export function estimateTokens(text: string): number {
// 한국어 비율 추정
const koreanChars = (text.match(/[가-힣]/g) || []).length;
const totalChars = text.length;
const koreanRatio = totalChars > 0 ? koreanChars / totalChars : 0;
// 한국어는 글자당 ~1.5 토큰, 영어는 ~0.25 토큰
const koreanTokens = koreanChars * 1.5;
const otherTokens = (totalChars - koreanChars) * 0.25;
return Math.ceil(koreanTokens + otherTokens);
}
/**
* 검색 결과 청크들을 토큰 예산 내에서 선택합니다.
*
* 선택 전략:
* 1. 스코어 내림차순 정렬
* 2. 중복 제거 (같은 filePath를 가진 청크)
* 3. 토큰 예산 내에서 순서대로 선택
* 4. 최소 청크 수 보장
*/
export function selectWithinBudget(
chunks: RetrievalChunk[],
config: Partial<ContextBudgetConfig> = {}
): { selected: RetrievalChunk[]; dropped: RetrievalChunk[]; tokensUsed: number } {
const cfg = { ...DEFAULT_BUDGET, ...config };
const budget = Math.floor(cfg.totalBudget * cfg.retrievalRatio);
// 1. Sort by score descending
const sorted = [...chunks].sort((a, b) => b.score - a.score);
// 2. Deduplicate by filePath
const seen = new Set<string>();
const deduped = sorted.filter((chunk) => {
const key = chunk.metadata.filePath || chunk.id;
if (seen.has(key)) return false;
seen.add(key);
return true;
});
// 3. Select within budget
const selected: RetrievalChunk[] = [];
const dropped: RetrievalChunk[] = [];
let tokensUsed = 0;
for (const chunk of deduped) {
const chunkTokens = chunk.tokenEstimate || estimateTokens(chunk.content);
if (selected.length >= cfg.maxChunks) {
dropped.push(chunk);
continue;
}
if (tokensUsed + chunkTokens > budget && selected.length >= cfg.minChunks) {
dropped.push(chunk);
continue;
}
selected.push(chunk);
tokensUsed += chunkTokens;
}
return { selected, dropped, tokensUsed };
}
/**
* 선택된 청크들을 하나의 컨텍스트 문자열로 조립합니다.
* 소스별로 그룹화하여 가독성을 높입니다.
*/
export function assembleContext(chunks: RetrievalChunk[]): string {
if (chunks.length === 0) return '';
const sourceLabels: Record<string, string> = {
'brain-trace': '📚 Second Brain Knowledge',
'brain-memory': '📚 Brain Knowledge',
'long-term-memory': '🧠 Long-Term Memory (사용자 규칙/결정)',
'project-memory': '📂 Project Memory (프로젝트 컨텍스트)',
'procedural-memory': '📋 Procedural Memory (반복 절차)',
'episodic-memory': '📖 Episodic Memory (과거 대화 흐름)',
'project-scan': '🔍 Project Scan',
'recent-knowledge': '📄 Recent Project Knowledge'
};
// Group by source
const groups = new Map<string, RetrievalChunk[]>();
for (const chunk of chunks) {
const key = chunk.source;
if (!groups.has(key)) groups.set(key, []);
groups.get(key)!.push(chunk);
}
const sections: string[] = [];
for (const [source, groupChunks] of groups) {
const label = sourceLabels[source] || source;
const items = groupChunks
.map((c) => `- ${c.title}: ${c.content}`)
.join('\n');
sections.push(`### ${label}\n${items}`);
}
return [
'[MEMORY CONTEXT]',
'Review this layered memory before preparing the answer. Use it only when relevant, and prefer the current user request when there is conflict.',
'',
sections.join('\n\n')
].join('\n');
}
+299
View File
@@ -0,0 +1,299 @@
/**
* ============================================================
* RetrievalOrchestrator — Unified RAG Pipeline
*
* Astra의 모든 검색 소스를 통합 관리하는 오케스트레이터입니다.
*
* 검색 흐름:
* ① Query Planning — 의도 분류 + 검색 전략 결정
* ② Parallel Search — Brain + Memory + Project + Episode 동시 검색
* ③ Result Fusion — 통합 스코어링 + 중복 제거
* ④ Context Budget — 토큰 예산 내에서 최종 선택
* ============================================================
*/
import * as fs from 'fs';
import * as path from 'path';
import { BrainProfile } from '../config';
import { findBrainFiles, summarizeText } from '../utils';
import { MemoryManager } from '../memory';
import { RetrievalChunk, RetrievalResult, ContextBudgetConfig } from './types';
import { tokenize, expandQuery, scoreTfIdf, extractBestExcerpt } from './scoring';
import { selectWithinBudget, assembleContext, estimateTokens } from './contextBudget';
export { tokenize, expandQuery, scoreTfIdf, extractBestExcerpt } from './scoring';
export { selectWithinBudget, assembleContext, estimateTokens } from './contextBudget';
export * from './types';
interface RetrievalOptions {
brain: BrainProfile;
memoryManager: MemoryManager;
workspacePath?: string;
chatHistory?: Array<{ role: string; content: string }>;
contextBudget?: Partial<ContextBudgetConfig>;
brainFileLimit?: number;
includeRawConversations?: boolean;
}
export class RetrievalOrchestrator {
/**
* 통합 검색을 수행합니다.
* 모든 소스에서 검색 → TF-IDF 스코어링 → 중복 제거 → 예산 내 선택
*/
public retrieve(query: string, options: RetrievalOptions): RetrievalResult {
const fusionLog: string[] = [];
const allChunks: RetrievalChunk[] = [];
const queryTokens = tokenize(query);
const expandedTokens = expandQuery(queryTokens);
fusionLog.push(`Query tokens: [${queryTokens.slice(0, 10).join(', ')}]`);
fusionLog.push(`Expanded tokens: [${expandedTokens.slice(0, 15).join(', ')}]`);
// ── ① Brain File Search (TF-IDF enhanced) ──
const brainChunks = this.searchBrainFiles(
query,
expandedTokens,
options.brain,
options.brainFileLimit || 8,
options.includeRawConversations || false
);
allChunks.push(...brainChunks);
fusionLog.push(`Brain search: ${brainChunks.length} chunks found`);
// ── ② Memory Layers ──
const memoryChunks = this.searchMemoryLayers(
query,
options.memoryManager,
options.chatHistory || [],
options.workspacePath
);
allChunks.push(...memoryChunks);
fusionLog.push(`Memory search: ${memoryChunks.length} chunks found`);
// ── ③ Result Fusion — normalize scores across sources ──
this.normalizeScores(allChunks);
fusionLog.push(`Total chunks before budget: ${allChunks.length}`);
// ── ④ Context Budget Selection ──
const { selected, dropped, tokensUsed } = selectWithinBudget(
allChunks,
options.contextBudget
);
fusionLog.push(`Selected: ${selected.length}, Dropped: ${dropped.length}, Tokens: ${tokensUsed}`);
return {
query,
totalChunks: allChunks.length,
selectedChunks: selected,
droppedChunks: dropped,
totalTokensUsed: tokensUsed,
contextBudget: options.contextBudget?.totalBudget || 8000,
fusionLog
};
}
/**
* 검색 결과를 최종 컨텍스트 문자열로 변환합니다.
*/
public buildContextString(result: RetrievalResult): string {
return assembleContext(result.selectedChunks);
}
// ─── Brain File Search ───
private searchBrainFiles(
query: string,
expandedTokens: string[],
brain: BrainProfile,
limit: number,
includeRaw: boolean
): RetrievalChunk[] {
try {
const allFiles = findBrainFiles(brain.localBrainPath)
.filter((file) => includeRaw || !this.isRawConversation(path.relative(brain.localBrainPath, file)));
if (allFiles.length === 0) return [];
// Read all files for TF-IDF
const documents = allFiles.map((file) => {
let content = '';
let lastModified = 0;
try {
content = fs.readFileSync(file, 'utf8');
lastModified = fs.statSync(file).mtimeMs;
} catch { /* skip */ }
return {
title: path.basename(file, '.md'),
content,
lastModified,
filePath: file,
relativePath: path.relative(brain.localBrainPath, file)
};
});
// TF-IDF scoring
const scored = scoreTfIdf(expandedTokens, documents);
return scored
.filter((s) => s.score > 0)
.sort((a, b) => b.score - a.score)
.slice(0, limit)
.map((scored) => {
const doc = documents[scored.index];
const excerpt = extractBestExcerpt(doc.content, expandedTokens, 400);
return {
id: `brain-${scored.index}`,
source: 'brain-memory' as const,
title: doc.relativePath,
content: summarizeText(excerpt, 400),
score: scored.score,
tokenEstimate: estimateTokens(excerpt),
metadata: {
filePath: doc.filePath,
category: this.inferCategory(doc.relativePath),
isProjectEvidence: this.isProjectEvidence(doc.relativePath, doc.content),
lastUpdated: doc.lastModified
}
};
});
} catch {
return [];
}
}
// ─── Memory Layer Search ───
private searchMemoryLayers(
query: string,
memoryManager: MemoryManager,
chatHistory: Array<{ role: string; content: string }>,
workspacePath?: string
): RetrievalChunk[] {
const chunks: RetrievalChunk[] = [];
// Long-Term Memory
const ltm = memoryManager.getLongTermMemory();
const ltmContext = ltm.buildContext(query);
if (ltmContext) {
chunks.push({
id: 'ltm-context',
source: 'long-term-memory',
title: ltmContext.label,
content: ltmContext.content,
score: ltmContext.relevance,
tokenEstimate: estimateTokens(ltmContext.content),
metadata: { category: 'long-term' }
});
}
// Project Memory
if (workspacePath) {
const pm = memoryManager.getProjectMemory(workspacePath);
const pmContext = pm.buildContext(query);
if (pmContext) {
chunks.push({
id: 'pm-context',
source: 'project-memory',
title: pmContext.label,
content: pmContext.content,
score: pmContext.relevance,
tokenEstimate: estimateTokens(pmContext.content),
metadata: { category: 'project', isProjectEvidence: true }
});
}
}
// Procedural Memory
const proc = memoryManager.getProceduralMemory();
const procContext = proc.buildContext(query);
if (procContext) {
chunks.push({
id: 'proc-context',
source: 'procedural-memory',
title: procContext.label,
content: procContext.content,
score: procContext.relevance,
tokenEstimate: estimateTokens(procContext.content),
metadata: { category: 'procedural' }
});
}
// Episodic Memory
const ep = memoryManager.getEpisodicMemory();
const epContext = ep.buildContext(query);
if (epContext) {
chunks.push({
id: 'ep-context',
source: 'episodic-memory',
title: epContext.label,
content: epContext.content,
score: epContext.relevance,
tokenEstimate: estimateTokens(epContext.content),
metadata: { category: 'episodic' }
});
}
return chunks;
}
// ─── Score Normalization ───
/**
* 서로 다른 스코어 스케일을 가진 소스들의 점수를 0~1로 정규화합니다.
*/
private normalizeScores(chunks: RetrievalChunk[]): void {
// Group by source
const groups = new Map<string, RetrievalChunk[]>();
for (const chunk of chunks) {
if (!groups.has(chunk.source)) groups.set(chunk.source, []);
groups.get(chunk.source)!.push(chunk);
}
// Normalize each group independently
for (const [, group] of groups) {
const maxScore = Math.max(...group.map((c) => c.score), 0.001);
for (const chunk of group) {
chunk.score = chunk.score / maxScore;
}
}
// Source priority boost (some sources are inherently more valuable for RAG)
const sourceBoost: Record<string, number> = {
'brain-trace': 1.0,
'brain-memory': 0.9,
'project-memory': 0.85,
'long-term-memory': 0.8,
'procedural-memory': 0.95, // Procedural is highly specific
'episodic-memory': 0.7,
'project-scan': 0.6,
'recent-knowledge': 0.75
};
for (const chunk of chunks) {
const boost = sourceBoost[chunk.source] || 0.5;
chunk.score *= boost;
}
}
// ─── Helpers ───
private isRawConversation(relativePath: string): boolean {
return /(^|[\\/])(00_Raw|raw-data|conversations?|transcripts?)([\\/]|$)/i.test(relativePath);
}
private inferCategory(relativePath: string): string {
const normalized = relativePath.toLowerCase();
if (/(decisions?|adr|planning)/i.test(normalized)) return 'decision';
if (/(records|development|bugs)/i.test(normalized)) return 'project-record';
if (/(architecture|design|pattern)/i.test(normalized)) return 'architecture';
if (/(knowledge|wiki|topics)/i.test(normalized)) return 'knowledge';
return 'general';
}
private isProjectEvidence(relativePath: string, content: string): boolean {
const normalized = relativePath.toLowerCase();
if (/(records|planning|development|bugs|retrospectives|projectchronicle)/i.test(normalized)) return true;
if (/adr-\d+|(^|[\\/])decisions?([\\/]|$)/i.test(normalized)) return true;
return false;
}
}
+241
View File
@@ -0,0 +1,241 @@
/**
* ============================================================
* Scoring Engine — TF-IDF + Bilingual Tokenizer
*
* 단순 includes() 키워드 매칭을 넘어서,
* TF-IDF 가중치 기반의 문서 스코어링을 제공합니다.
* 한국어/영어 양국어 토크나이저를 포함합니다.
* ============================================================
*/
// ─── Bilingual Tokenizer ───
const STOP_WORDS_EN = new Set([
'the', 'a', 'an', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for',
'of', 'with', 'by', 'from', 'is', 'are', 'was', 'were', 'be', 'been',
'have', 'has', 'had', 'do', 'does', 'did', 'will', 'would', 'could',
'should', 'may', 'might', 'can', 'this', 'that', 'these', 'those',
'it', 'its', 'not', 'no', 'what', 'how', 'when', 'where', 'which',
'who', 'whom', 'why', 'if', 'then', 'than', 'so', 'as', 'just',
'about', 'also', 'more', 'some', 'very', 'all', 'each', 'every',
'such', 'please', 'write', 'use', 'using', 'used'
]);
const STOP_WORDS_KO = new Set([
'그리고', '그런데', '그래서', '하지만', '또한', '또는', '해서', '하는',
'있어', '없어', '아래', '위에', '어떻게', '이것', '저것', '그것',
'이런', '저런', '그런', '여기', '거기', '필요', '사용', '관련',
'대한', '대해', '통해', '따라', '위해', '대로', '만큼'
]);
/**
* 한국어/영어 혼합 텍스트를 토큰으로 분리합니다.
*/
export function tokenize(text: string): string[] {
return text
.toLowerCase()
.split(/[^a-z0-9가-힣_.-]+/g)
.map((t) => t.trim())
.filter((t) => t.length >= 2)
.filter((t) => !STOP_WORDS_EN.has(t) && !STOP_WORDS_KO.has(t));
}
/**
* 동의어/관련어 확장을 수행합니다.
*/
export function expandQuery(tokens: string[]): string[] {
const synonymMap: Record<string, string[]> = {
'성능': ['performance', 'optimization', '최적화', 'speed'],
'performance': ['성능', '최적화', 'optimization', 'speed'],
'아키텍처': ['architecture', '구조', 'structure', 'design'],
'architecture': ['아키텍처', '구조', 'structure', 'design'],
'메모리': ['memory', '기억', 'cache', 'storage'],
'memory': ['메모리', '기억', 'cache', 'storage'],
'버그': ['bug', 'error', '오류', 'issue', 'defect'],
'bug': ['버그', 'error', '오류', 'issue'],
'설계': ['design', '아키텍처', 'architecture', 'pattern'],
'design': ['설계', '아키텍처', 'architecture', 'pattern'],
'배포': ['deploy', 'deployment', 'release', 'ci', 'cd'],
'deploy': ['배포', 'deployment', 'release'],
'테스트': ['test', 'testing', 'spec', 'jest', 'mocha'],
'test': ['테스트', 'testing', 'spec'],
'프로젝트': ['project', '프로그램', 'repo', 'repository'],
'project': ['프로젝트', '프로그램', 'repo'],
'방향': ['direction', '전략', 'strategy', '목표', 'goal'],
'direction': ['방향', '전략', 'strategy', '목표']
};
const expanded = new Set(tokens);
for (const token of tokens) {
const synonyms = synonymMap[token];
if (synonyms) {
for (const syn of synonyms) {
expanded.add(syn);
}
}
}
return Array.from(expanded);
}
// ─── TF-IDF Scoring ───
/**
* TF (Term Frequency): 문서 내 용어 빈도
*/
function termFrequency(term: string, documentTokens: string[]): number {
if (documentTokens.length === 0) return 0;
const count = documentTokens.filter((t) => t === term).length;
return count / documentTokens.length;
}
/**
* IDF (Inverse Document Frequency): 전체 문서 대비 희소도
*/
function inverseDocumentFrequency(
term: string,
allDocumentTokenSets: Array<Set<string>>
): number {
const containing = allDocumentTokenSets.filter((doc) => doc.has(term)).length;
return Math.log((allDocumentTokenSets.length + 1) / (containing + 1)) + 1;
}
export interface ScoredDocument {
index: number;
score: number;
titleBoost: number;
recencyBoost: number;
matchedTerms: string[];
}
/**
* TF-IDF 기반으로 문서 집합을 스코어링합니다.
*/
export function scoreTfIdf(
queryTokens: string[],
documents: Array<{
title: string;
content: string;
lastModified?: number;
}>
): ScoredDocument[] {
if (documents.length === 0 || queryTokens.length === 0) return [];
// Pre-tokenize all documents
const docTokenArrays = documents.map((doc) =>
tokenize(`${doc.title} ${doc.content}`)
);
const docTokenSets = docTokenArrays.map((tokens) => new Set(tokens));
// Expand query with synonyms
const expandedQuery = expandQuery(queryTokens);
// Compute IDF for each query term
const idfCache = new Map<string, number>();
for (const term of expandedQuery) {
if (!idfCache.has(term)) {
idfCache.set(term, inverseDocumentFrequency(term, docTokenSets));
}
}
const now = Date.now();
return documents.map((doc, index) => {
const docTokens = docTokenArrays[index];
const titleTokens = new Set(tokenize(doc.title));
let score = 0;
const matchedTerms: string[] = [];
for (const term of expandedQuery) {
const tf = termFrequency(term, docTokens);
const idf = idfCache.get(term) || 1;
const tfidf = tf * idf;
if (tfidf > 0) {
matchedTerms.push(term);
}
// Title match bonus (3x)
const titleMultiplier = titleTokens.has(term) ? 3.0 : 1.0;
score += tfidf * titleMultiplier;
}
// Recency boost: documents modified recently get a boost
let recencyBoost = 0;
if (doc.lastModified) {
const daysAgo = (now - doc.lastModified) / (1000 * 60 * 60 * 24);
if (daysAgo < 1) recencyBoost = 0.3;
else if (daysAgo < 7) recencyBoost = 0.2;
else if (daysAgo < 30) recencyBoost = 0.1;
}
// Title match bonus for exact query term presence
const titleBoost = queryTokens.some((t) => titleTokens.has(t)) ? 0.2 : 0;
return {
index,
score: score + recencyBoost + titleBoost,
titleBoost,
recencyBoost,
matchedTerms: [...new Set(matchedTerms)]
};
});
}
/**
* 텍스트에서 가장 관련성 높은 구간(excerpt)을 추출합니다.
* 단순 paragraph 단위가 아니라, 키워드 밀도가 높은 윈도우를 찾습니다.
*/
export function extractBestExcerpt(
content: string,
queryTokens: string[],
maxLength = 500
): string {
const expanded = expandQuery(queryTokens);
const expandedSet = new Set(expanded);
// Split into sentences (한국어 + 영어)
const sentences = content
.split(/(?<=[.!?。!?\n])\s*/)
.map((s) => s.trim())
.filter((s) => s.length > 10);
if (sentences.length === 0) return content.slice(0, maxLength);
// Score each sentence
const scored = sentences.map((sentence, idx) => {
const tokens = tokenize(sentence);
const matchCount = tokens.filter((t) => expandedSet.has(t)).length;
const density = tokens.length > 0 ? matchCount / tokens.length : 0;
return { sentence, idx, matchCount, density };
});
// Find the best window of consecutive sentences
let bestStart = 0;
let bestScore = -1;
let bestLen = 0;
for (let i = 0; i < scored.length; i++) {
let windowText = '';
let windowScore = 0;
let j = i;
while (j < scored.length && windowText.length < maxLength) {
windowText += scored[j].sentence + ' ';
windowScore += scored[j].matchCount + scored[j].density * 2;
j++;
}
if (windowScore > bestScore) {
bestScore = windowScore;
bestStart = i;
bestLen = j - i;
}
}
const excerptSentences = scored
.slice(bestStart, bestStart + bestLen)
.map((s) => s.sentence);
const result = excerptSentences.join(' ');
return result.length > maxLength ? result.slice(0, maxLength - 3) + '...' : result;
}
+50
View File
@@ -0,0 +1,50 @@
/**
* ============================================================
* Retrieval Types (검색 결과 통합 타입)
*
* 모든 검색 소스(Brain, Memory, Project, Episode)의 결과를
* 통합 인터페이스로 정의합니다.
* ============================================================
*/
export type RetrievalSource =
| 'brain-trace' // Second Brain Trace
| 'brain-memory' // findRelevantBrainMemory (legacy)
| 'long-term-memory' // Long-Term Memory
| 'project-memory' // Project Memory
| 'procedural-memory' // Procedural Memory
| 'episodic-memory' // Episodic Memory
| 'project-scan' // Local Project Path scan
| 'recent-knowledge'; // Recent Project Knowledge record
export interface RetrievalChunk {
id: string;
source: RetrievalSource;
title: string;
content: string;
score: number; // 0.0 ~ 1.0 normalized
tokenEstimate: number; // rough character / 4
metadata: {
filePath?: string;
category?: string;
isProjectEvidence?: boolean;
lastUpdated?: number;
};
}
export interface RetrievalResult {
query: string;
totalChunks: number;
selectedChunks: RetrievalChunk[];
droppedChunks: RetrievalChunk[];
totalTokensUsed: number;
contextBudget: number;
fusionLog: string[]; // 디버그용 융합 로그
}
export interface ContextBudgetConfig {
totalBudget: number; // 전체 토큰 예산
retrievalRatio: number; // 검색 결과 비율 (0.0~1.0)
minChunks: number; // 최소 포함 청크 수
maxChunks: number; // 최대 포함 청크 수
}