index-manager

Building a FAISS search index for legal documents involves downloading PDFs, extracting text, generating embeddings with sentence-transformers, and constructing the FAISS index. This Skill automates the entire lifecycle from USPTO PDF acquisition through GPU-accelerated embedding generation and index creation, handling text chunking and integrity verification.

Yes, combining BM25 and FAISS enables hybrid search: BM25 provides keyword matching for patent terminology while FAISS delivers semantic similarity search. This Skill builds both indexes simultaneously, allowing you to leverage strengths of each approach for comprehensive MPEP search across regulatory updates.

When the MPEP updates, you need to rebuild your search index to include new regulations and ensure current search results. This Skill manages the full rebuild workflow: downloading updated PDFs from USPTO, re-extracting text, regenerating embeddings, and refreshing both FAISS and BM25 indexes with verification checks.

Index corruption can degrade search accuracy. This Skill performs index health checks, detects corruption events, and provides guided rebuilding to restore data integrity. It includes troubleshooting tools and maintenance workflows to ensure your FAISS and BM25 indexes remain reliable and performant.

GPU acceleration significantly speeds up embedding generation for large document collections. This Skill supports GPU-accelerated processing with sentence-transformers for embedding creation, reducing indexing time while maintaining the option to fall back to CPU-based processing when needed.

This Skill handles PDF downloads directly from USPTO sources (MPEP, 35 USC, 37 CFR) with PyMuPDF for text extraction and integrity validation. It manages the full acquisition-to-index pipeline, requiring only PDF availability and performing chunked text processing for reliable, searchable legal document indexes.