overview.
DocuSync is a local-first document intelligence system. It parses text-native and scanned PDFs (with OCR), DOCX files with embedded tables, and plain text; tags documents automatically with a zero-shot classifier; and answers queries with hybrid BM25 + vector search followed by cross-encoder reranking — all without any cloud API.
technical implementation.
Hybrid Retrieval
Keyword hits come from SQLite FTS5 with a Porter stemmer, semantic hits from GTE embeddings in a ChromaDB cosine HNSW index; the merged candidates are reranked with a GTE cross-encoder.
Evaluation
Retrieval quality is benchmarked on the SyllabusQA dataset (ACL 2024): 63 course syllabi and over 5,000 question-answer pairs.
key features.
- Upload PDF, DOCX, TXT, MD and JSON up to 120 MB
- OCR for scanned PDFs with a PaddleOCR + Tesseract ensemble
- Automatic tagging with a zero-shot classifier and keyword extraction
- Hybrid BM25 + semantic search with cross-encoder reranking
- Document summaries with BART
- Zero cloud — every model runs locally