colonelwatch/abstracts-search

Generates dense vector embeddings from 110M academic abstracts using the Stella 1.5B model, then builds a FAISS index for fast approximate nearest-neighbor retrieval—all components (embeddings, index, search interface) are published as separate Hugging Face datasets and spaces for modularity. Integrates with OpenAlex for publication metadata and supports incremental syncing against quarterly dataset snapshots to keep the index current. The modular architecture allows running just the search interface without rebuilding, or performing full reindexing on commodity hardware (RTX 3060, 32GB RAM+swap) in under a week.

102 stars.