swiss_army_llama

Why we included this project

Running semantic search locally means you can keep documents and embeddings on your own hardware, and this FastAPI service makes that straightforward. It wraps llama.cpp-backed embedding models behind REST endpoints, ingests PDFs (including scanned ones via OCR), Word files, and audio that gets transcribed with Whisper, then stores the resulting vectors in SQLite so repeated work is skipped. For retrieval, it uses FAISS, and when plain cosine similarity isn't discriminating enough, a second pass applies more sophisticated statistical measures like Kendall tau or Hoeffding's D. A Swagger UI and a simple Docker setup let a small team stand up an internal search tool without a big integration effort, and the codebase is a useful reference for how ingestion, embedding pooling, and vector retrieval fit together.