feat(recall): progressive retrieval with G-document routing - #102
Conversation
- Remove docs/team-codebase/repos/ duplication; teamwiki evidence as sole knowledge carrier with AI narrative appended idempotently - Fix deep-enrich reading call-chains.md → dependency-paths.md (g2 bug) - Fix rebuild-wiki-index.ts still referencing old filename - Guard deepEnrich with skipEnrich flag to prevent unwanted AI calls - Fix router.md append link: [[code/...]] → [[evidence/code/...]] - Eliminate graph-index write race: extractCodebase writes per-repo only - Share Intl.Segmenter as module-level singleton (perf) - Fix --output CLI description (team-codebase → teamwiki)
Three-layer progressive retrieval system replacing flat BM25 search: - depth=route: returns routing suggestions only (no BM25) - depth=context (default): searches overview + modules + docs - depth=lookup: full search including raw symbol pages Production-side enhancements: - Rewrite routerTemplate with search-anchor + question routing table - Rewrite indexTemplate with module quick-reference - Upgrade G1 to N×N matrix with forward/reverse dependency indexes - Upgrade G2 with entry-grouped call chains and search-anchor - Add G5 (AI scenario sequence diagrams) and G6 (BFS multi-hop analysis) - Add graph/README.md dispatch routing table Consumption-side optimizations: - Pre-computed adjacency map for graph-boost (O(E²) → O(degree²)) - Exclude navigation files from BM25 corpus in context mode - Add MAX_RECURSION_DEPTH=10 to loadPagesRecursive - Warn on invalid --depth values with fallback to context - Update recall subagent to leverage G-documents and depth levels
jeff-r2026
left a comment
jeff-r2026
left a comment
There was a problem hiding this comment.
Code Review — medium effort
概览
PR 实现了三层渐进检索（route/context/lookup）、G5 场景序列图和 G6 多跳依赖分析、router.md/index.md 增强、graph-boost 性能优化。设计合理，变更范围清晰。
发现
🔴 src/deep-enrich.ts:570 — G6 BFS 方向反转
buildG6Content 的 "爆炸半径" BFS 遍历了正向边（X 依赖什么），而非反向边（谁依赖 X）。
adj 仅存储 from → Set<to>，BFS 从某节点出发跟随正向边，计算的是 "X 传递依赖了什么"，而输出标签声称是 "X 变更可能影响 N 个组件"。
复现： 边 [{from:'auth', to:'db'}, {from:'api', to:'auth'}]，auth 的爆炸半径应是 api（api 依赖 auth），但 BFS 跟随 adj.get('auth')={'db'}，错误报告 db 为受影响组件。
建议修复： 构建反向邻接表 revAdj: to → Set<from> 用于 BFS 遍历。
🟡 src/ci/extract-mr.ts:373 — 全局 graph-index.json 被单仓库覆盖
extractCodebase 写 per-repo graph 到 srcWiki/.indices/graph-index.json，然后 fse.copy(srcWiki, teamWikiRoot, { overwrite: true }) 将其覆盖到 teamWikiRoot/.indices/graph-index.json（原本是聚合后的全局 graph），导致多仓库 graph 数据丢失。
建议修复： copy 后调用 aggregateGlobalGraph()，或 copy 时排除 .indices/graph-index.json。
🟡 src/code-knowledge-recall.ts — NAVIGATION_FILES 可能误杀嵌套 index.md
NAVIGATION_FILES.has(entry.name) 在 CONTEXT_ALLOWED_PATTERNS 白名单检查之前执行。任何 docs/index.md（如 evidence/code/project/docs/index.md）会被 index.md 匹配跳过，即使它满足 /docs\/.+\.md$/ 模式。
建议修复： 将 NAVIGATION_FILES 检查限定为相对路径的顶层文件（如检查 relativePath 而非仅 entry.name），或将其移到 pattern 匹配之后作为 fallback。
🤖 Generated with Claude Code
- P1-1: overview.md handles missing frontmatter on empty base - P1-2: fs.copy srcGraph guarded with pathExists check - P1-3: depth=route only loads router.md, uses actual page title - P2-4: graphReadmeTemplate dynamically renders based on G5/G6 flags - P2-5: CONTEXT_ALLOWED_PATTERNS comment clarifies G1-G6 coverage
8da0259 to
c2b30bc
Compare
jeff-r2026
left a comment
jeff-r2026
left a comment
There was a problem hiding this comment.
All three findings from the previous review have been properly addressed:
- G6 BFS now uses revAdj (reverse adjacency map) — blast radius correctly computes "who depends on X"
- ci/extract-mr.ts no longer blanket-copies srcWiki; copies per-repo evidence only and calls aggregateGlobalGraph()
- NAVIGATION_FILES exclusion is now depth-gated (depthFromProject <= 1), protecting nested docs/index.md
LGTM
2 participants
Summary
Progressive retrieval layers
routecontext(default)lookupG-document routing
Test plan
recall-progressive.test.tsverifying route/context/lookup behaviorDepends on: #98 (merged)
🤖 Generated with Claude Code