Index management

What coverage means, when to rebuild, Embedding, and what to check first

Purpose

After ingest, RAG depends on a built index. Low coverage means many ingested repos are not searchable yet. Vectors stay empty if Chat is mapped but Embedding is not.

RAG workspace Index: coverage, vector progress, failed cards

Screenshot TODO · images/screenshots/knowledge-base/index-management.webp

Where to look

Workspace header or the Index tab: library size, coverage (keyword / vector as the UI shows), Embedding hints.

Vector progress is ready coverage for the current Embedding model. The toolbar status panel can inspect and pause; an unfinished build is not marked complete.

Cards distinguish: still in the library, vectorized for this model, updated time, approximate tokens. Opening pending / failed / stale cards jumps to that chunk.

On rebuild, the Meilisearch keyword index syncs text first, so keyword search can work before every vector is ready. Local FTS supports Chinese infix matching; naming a repo is more likely to hit README.

When to rebuild

Situation	Do this
Changed Embedding model	Rebuild or spaces will not match
Just ingested many repos	Wait for auto-index or fill gaps
Search went empty after clearing cache	Rebuild
“Nothing found” but the library is not empty	Check coverage and whether @ is too narrow

Time and API cost scale with library size. Watch progress and cancel on a large rebuild.

Checklist

1. Pro still valid → Restore
2. Embedding test passed → AI configuration
3. Target repo actually ingested → Adding repos
4. @ did not exclude the real source → Workspace
5. Chat mapped, Embedding empty → map Embedding

Related

• Citations
• FAQ · AI