Skip to main content

Hybrid Retrieval

Clinical Corvus uses hybrid retrieval to support evidence-backed answers: it combines keyword matching (fast, exact) with semantic retrieval (concept-aware) to improve recall and precision on real clinical queries.

The End-to-End Flow​

  1. Ingestion: PaddleOCR 3.x handles scanned documents and images, pdfplumber extracts structured content from born-digital PDFs, and pypdf provides a plain-text fallback.
  2. Chunking & metadata: content is split into coherent sections with basic tags (e.g., guideline vs narrative).
  3. Indexing: content is stored in:
    • a keyword index for exact matching,
    • a vector index for semantic similarity.
  4. Hybrid search: both retrieval modes are run and merged into a single ranked list.
  5. Context packaging: the top evidence is formatted into a traceable context pack.
  6. Cited synthesis: the answer is generated with citations tied to the retrieved snippets.

Structured facts extracted through OCR enter the clipboard path as drafts. They require explicit clinical confirmation before persistence as confirmed facts.

Why Hybrid Beats “Vector Only”​

  • Keyword search is often better for exact terms (drug names, lab thresholds, acronyms).
  • Semantic search is often better for conceptual questions (syndromes, management strategies).
  • Clinical questions frequently require both.

Conceptual Diagram​