RAG Medical Accuracy

Retrieval-augmented generation over clinical literature — passage retrieval, relevance ranking, top-K selection, and grounded answer generation

Current Stage

—

—

Passages retrieved

—

Avg relevance

—

Top-K used

—

Answer accuracy

—

Corpus Coverage of Topic55%

Retrieval Relevance Quality55%

Document corpus

Retrieved passage

Top-K selected

LLM generation

Grounded answer

Scroll for details

The Clinical Query Enters the Retrieval Pipeline

Placeholder lead: a natural-language clinical question kicks off retrieval.

Free text: Query type (Placeholder caption)
Dense vector: Encoding (Placeholder caption)
<300 ms: Latency budget (Placeholder caption)
Corpus search: Downstream step (Placeholder caption)

Placeholder section heading

Placeholder body text, short filler describing query embedding and intent.

Searching the Unstructured Document Corpus

Placeholder lead: the retriever scans literature for relevant passages.

~2.4M docs: Corpus size (Placeholder caption)
~300 tokens: Chunk size (Placeholder caption)
Hybrid dense+sparse: Search type (Placeholder caption)
~200: Candidates pulled (Placeholder caption)

Placeholder section heading

Placeholder body text about vector search over corpus chunks.

Relevance Scoring and Top-K Passage Selection

Placeholder lead: passages ranked, only the best K move forward.

Cross-encoder: Reranker (Placeholder caption)
5–10: Typical K (Placeholder caption)
0–1: Score range (Placeholder caption)
~0.6: Cutoff threshold (Placeholder caption)

Placeholder section heading

Placeholder body text about reranking and threshold cutoffs.

The LLM Generates an Answer Grounded in Retrieved Text

Placeholder lead: model composes answer citing retrieved passages only.

~4k tokens: Context window used (Placeholder caption)
Inline refs: Citation style (Placeholder caption)
Parametric memory: Fallback (Placeholder caption)
Post-hoc verify: Grounding check (Placeholder caption)

Placeholder section heading

Placeholder body text about prompt construction and grounding.

Answer Accuracy Tracks Retrieval Quality, Not Architecture Alone

Placeholder lead: good retrieval yields grounded answers, poor retrieval can still hallucinate.

~90%: High-quality retrieval accuracy (Placeholder caption)
~40%: Poor retrieval accuracy (Placeholder caption)
Grounded hallucination: Failure mode (Placeholder caption)
Corpus curation: Mitigation (Placeholder caption)

Placeholder section heading

Placeholder body text about retrieval quality driving final accuracy.

Placeholder highlight: RAG only grounds on what was actually retrieved.