index-manager（指数管理器）

Index Manager Skill

Expert system for managing MPEP search index lifecycle: PDF downloads, index building, maintenance, updates, optimization.

FOR CLAUDE: All dependencies installed, system operational.

• Go directly to appropriate phase
• Scripts/tools in mcp_server/
• Use patent-creator CLI when available
• Only run diagnostics if operations fail

When to Use

Building/rebuilding MPEP index, corruption/missing files, optimization, adding content, troubleshooting.

Index Lifecycle

PDFs Not Present -> Download (2-5 min, 500MB)
  -> Extract & Parse (500MB data)
  -> Generate Embeddings (5-10 min GPU, 35-65 min CPU)
  -> Build FAISS + BM25 Indexes
  -> Index Ready (mcp_server/index/)
  -> Maintenance (Verify -> Optimize -> Update)

Phase 1: PDF Management

Check Status:

ls pdfs/  # Should show mpep-*.pdf, consolidated_laws.pdf, consolidated_rules.pdf

Download PDFs:

patent-creator download-mpep
# Or: python install.py (Select "Download MPEP PDFs")

Verify Integrity:

python -c "
import fitz
from pathlib import Path
for pdf in Path('pdfs').glob('*.pdf'):
    try:
        doc = fitz.open(pdf)
        print(f'[OK] {pdf.name}: {len(doc)} pages')
        doc.close()
    except Exception as e:
        print(f'[X] {pdf.name}: ERROR - {e}')
"

Phase 2: Index Building

patent-creator rebuild-index
# Or: python mcp_server/server.py --rebuild-index

Timeline:

• Load PDFs: 30s
• Extract text: 1-2 min
• Chunk text (500 tokens): 30s
• Generate embeddings: 5-10 min (GPU) or 35-65 min (CPU)
• Build FAISS/BM25: 1 min
• Save to disk: 10s

Total: 5-15 min (GPU) or 35-65 min (CPU)

Custom Build:

from mcp_server.mpep_search import MPEPIndex
index = MPEPIndex(use_hyde=False)
index.build_index(
    chunk_size=500,
    overlap=50,
    batch_size=32  # Reduce to 16/8 if OOM
)

Phase 3: Verification

# Check files
ls -lh mcp_server/index/
# Expected: mpep_index.faiss (~150MB), mpep_metadata.json (~80MB), mpep_bm25.pkl (~60MB)

# Verify health
patent-creator health
# Should show: [OK] MPEP Index: Ready (12,543 chunks)

# Manual test
python -c "
from mcp_server.mpep_search import MPEPIndex
index = MPEPIndex()
print(f'Chunks: {len(index.chunks)}')
results = index.search('claim definiteness', top_k=3)
print(f'Search results: {len(results)}')
"

Phase 4: Maintenance

When to Rebuild:

• MPEP updates (quarterly check uspto.gov)
• Index corruption
• After adding new PDFs
• Performance degradation
• Machine migration

Rebuild Process:

# Backup (optional)
cp -r mcp_server/index mcp_server/index_backup_$(date +%Y%m%d)

# Rebuild
patent-creator rebuild-index

# Verify
patent-creator health

# Remove backup if successful
rm -rf mcp_server/index_backup_*

Phase 5: Content Updates

# Download new PDF
wget https://www.uspto.gov/web/offices/pac/mpep/mpep-2900.pdf -O pdfs/mpep-2900.pdf

# Rebuild (includes new section)
patent-creator rebuild-index

Note: Incremental updates not supported. Full rebuild required.

Troubleshooting

• OOM errors during build
• Build taking too long
• Corrupted index files
• Search returning no results

Performance Tuning

• Embedding generation speed (GPU vs CPU)
• Search latency optimization
• Index size reduction
• Batch size tuning

Quick Reference

Best Practices:

1. Backup before rebuild
2. Verify PDFs before building
3. Use GPU for 10x faster builds
4. Test after rebuild
5. Keep PDFs until verified
6. Weekly health checks