RAG Index Rebuild

RAG Index Rebuild

Reindex the local RAG corpus to add new documents, update stale chunks, or recover from corruption. Choose between full rebuild (clean slate, slow) and incremental (fast, targets specific files).

When to rebuild

When NOT to rebuild

• Just queried and got bad results. Likely a curation/corpus gap, not an index bug. Try the adt-rag-curate skill first.
• Index is <1 week old. Drift detection runs automatically at session start. Manual rebuild only if you've made recent bulk edits.
• Testing retrieval logic. Use a static index; don't rebuild during active development.

Full rebuild (clean slate)

Drops the index and reconstructs all chunks from source files.

cd ~/.claude/rag-index
venv/bin/python build.py

What happens:

• Scans all source repos and doc dirs configured in build.py
• Creates new index.sqlite with all chunks from scratch
• Embeds each chunk with all-MiniLM-L6-v2
• Outputs final chunk count and time

Expected output:

Index built successfully.
Total chunks: 14,355
Time: 7m 42s

Incremental rebuild (targeted files)

Add new documents or refresh stale ones without rebuilding the entire index.

cd ~/.claude/rag-index
venv/bin/python build.py --incremental file1.md file2.md

Examples:

# Add a new skill
venv/bin/python build.py --incremental ~/.claude/skills/my-skill/SKILL.md

# Update multiple files
venv/bin/python build.py --incremental ~/.claude/plans/project.md ~/.claude/standards/code.md

# Update all Python files under a path
venv/bin/python build.py --incremental ~/Desenvolvimento/forgekit/src/*.py

What happens:

• Deletes any existing chunks from the named files
• Re-embeds and inserts new chunks for those files only
• Existing chunks from other files remain untouched
• Much faster than full rebuild

Finding what needs reindexing

Run the report script to see stale/missing chunks:

cd ~/.claude/rag-index
venv/bin/python report.py

Check the output file for stale sections:

cat ~/.claude/rag-index/weekly.md | grep -A 20 "Stale chunks"

Look for:

• Missing files: chunks indexed from a file that no longer exists
• Modified files: files edited after the last index time (sha mismatch)

Incremental reindex those files:

venv/bin/python build.py --incremental ~/.claude/standards/updated-file.md

Post-rebuild verification

Test a query against the fresh index:

cd ~/.claude/rag-index
venv/bin/python query.py "your test question" --top 5

Expected output shows rrf, cos, bm25 scores for each result. Cosine >0.55 indicates good retrieval quality.

Check chunk distribution:

sqlite3 ~/.claude/rag-index/index.sqlite \
  "SELECT source_type, COUNT(*) FROM chunks GROUP BY source_type ORDER BY COUNT(*) DESC;"

Compare to baseline:

• skills: ≥500 chunks
• handoffs: ≥200 chunks
• standards, codex: ≥50 chunks combined

If counts drop significantly, the rebuild may have skipped a source. Check build.py configuration.

Troubleshooting

"No chunks added" after incremental rebuild

• Verify the file path exists and is in a tracked source dir
• Check that the file extension matches the parser (.md, .py, .json, etc.)
• Run a full rebuild if unsure

Index size unchanged after full rebuild

• The corpus may not have changed. Check weekly.md for timestamps.
• Verify tracked repos/dirs in build.py exist and have recent commits

Embedding fails during rebuild

• Ensure venv is active and all-MiniLM-L6-v2 model is downloaded: venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"
• If model is missing, download it: the first run will auto-download (~100MB)

See also

• adt-rag-quality — evaluate retrieval quality and find zero-hit gaps
• adt-rag-drift — detect and fix stale chunks
• adt-rag-curate — improve corpus coverage by adding missing docs