Indexless hybrid search (BM25 + Vector) for fast, private document retrieval. Zero-setup context for agentic workflows - no database, no daemon, just a single binary.
Single Rust binary
no external database, daemon, or long-running service; transparent caching in standard user cache directories.
Layered retrieval pipeline
Query Expansion → Retrieval → Fusion → Reranking with page-index-hybrid as the current benchmark champion preset.
Hybrid search by default
combines BM25 and vector retrieval; supports bm25, vector, and hybrid strategies switchable per query.
Autonomous planner runtime
sift search --agent drives bounded linear or graph planning over a local corpus using heuristic or model-driven planner strategies.
Broad input support
indexes text, HTML, PDF, and OOXML files (.docx, .xlsx, .pptx) with no preprocessing required.
Embeddable library surface
search, assemble_context, search_controller, and search_autonomous APIs with view, protocol, or latent emission modes.