An Open WebUI plugin for hybrid vector+keyword document search via Qdrant, with recency-aware scoring that boosts more recent documents.
Forked from daradib/openwebui-plugins.
- Hybrid search — Dense vector + sparse BM25 retrieval via Qdrant
- Recency weighting — Gaussian decay boosts scores for recently modified documents (configurable half-life)
- Reranking — Optional semantic reranker to improve result quality
- Incremental updates — Only re-ingests changed files (rsync-like comparison using internal document dates)
source .venv/bin/activate
# Build document store
python3 utils/build_document_store.py <input_dir> [options]
# Test search
python3 test_search.py "your query" [--top-k N] [--debug]
# Export document summary
python3 utils/list_document_store.py [options]python3 utils/build_document_store.py Docs/Basic \
--qdrant-url http://localhost:6333 \
--qdrant-collection ssaskb \
--embed-rerank-url http://127.0.0.1:9997 \
--embedding-model mlx-community/Qwen3-Embedding-4B-4bit-DWQ \
--chunk-size 512For recency scoring, last_modified_date is extracted from:
| File Type | Source | Fallback |
|---|---|---|
.pdf |
Internal PDF /ModDate metadata |
Filesystem mtime |
.docx |
Internal core_properties.modified |
Filesystem mtime |
.md, .txt |
Filesystem mtime | N/A |
After semantic reranking, each result's score is adjusted:
weight = exp(-0.5 × (age_days / half_life_days)²)
final_score = rerank_score × (1 - α) + rerank_score × weight × α
Default: α = 0.3, half_life = 365 days. A brand-new document keeps its full score; a 1-year-old document loses ~15%; older documents are progressively discounted.
Use --debug with test_search.py to see per-result score breakdowns.
llama-index(core + embeddings/vector-stores integrations)qdrant-clientpymupdf(PyMuPDF) — PDF internal date extractionpython-docx— DOCX internal date extraction (via docling)docling— Document parsing
See README.md in the upstream repo for full installation instructions.