Clinical Graph

Hybrid GraphRAG & Evidence Grounding
Prompt Guard: Checking...
Hybrid Knowledge Graph + Typed Vector Retrieval

Strictly Grounded Clinical GraphRAG for Medical Guidelines & Drug Labels

Clinical Graph combines typed dense vector retrieval over guideline prose and tables with a directed clinical knowledge graph so multi-hop therapeutic questions and exact dosing lookups stay verifiably grounded. Every synthesized claim is anchored to verbatim source quotes and inspectable chunk IDs with zero reliance on unverified pretraining memory.

STRICT NO-PRETRAINING

Evidence-only synthesis with verbatim quotes anchored to chunk_id and <mark> highlights.

HYBRID GRAPH + VECTOR

Fuses is_chunk / is_table vector search with is_node entity seeds and k-hop DiGraph traversal via RRF.

VERIFIED WEB FALLBACK

1-shot allowlisted Tavily search via label_vectors.npz & gpt-6-luna Decisions with temporary vector-chunk RAG.

5-STAGE PIPELINE ARCHITECTURE
1. Guard & Rewrite
Prompt Guard + QueryRewriter v4
2. Hybrid Retrieval
Typed Vectors + k-Hop Graph RRF
3. Web Fallback
gpt-6-luna + Allowlisted Vector RAG
4. Grounded Synthesis
AnswerAgent v4 + Verbatim Citations

Unit & End-to-End Evaluation Telemetry

Latest isolated stage evaluations and E2E query routing metrics served from GET /v1/benchmarks.

Loading benchmark telemetry...