06.05
Evidence base
Executive summary
The evidence base retrieves literature from 9 open sources in parallel, ranks it deterministically, checks licences and retractions, and verifies whether each passage actually supports a claim. Every answer is bound to one immutable snapshot of exactly what was retrieved, so a citation can always be reopened to the passage behind it. All sources are free to access by policy, and queries leave the platform de-identified.
Overview
- ProductionReason · Evidence baseGround clinical answers and reports in retrievable, licensed, current sources — and say so plainly when they are not grounded.
Pipeline
- A dental query planner builds a structured query; a lite-tier model optionally expands it with ontology terms.
- The source router queries enabled sources in parallel, each failure mode returned as a distinct typed outcome, never as an empty success (EVID-05).
- A deterministic ranker scores results; the licence guard and a Crossref retraction gate remove what may not be used.
- Passages are extracted deterministically; dense retrieval embeds query and passages and fuses with lexical ranking by Reciprocal Rank Fusion (k = 60).
- An entailment verifier classifies each passage–claim pair as supports, contradicts or contextual; the result is sealed into an evidence snapshot with claim links.
Data owned
- Source documents and passages
- Holds
- Retrieved records and extracted passages
- Evidence snapshots
- Holds
- Immutable per-response record: query plan, provider traces, source versions, passages, engine versions
- Claims and claim–evidence links
- Holds
- Support relations per claim
- Guideline documents and recommendations
- Holds
- Curated guideline content (schema in place, content pending)
- Provider cache
- Holds
- De-identified provider responses
| Store | Holds |
|---|---|
| Source documents and passages | Retrieved records and extracted passages |
| Evidence snapshots | Immutable per-response record: query plan, provider traces, source versions, passages, engine versions |
| Claims and claim–evidence links | Support relations per claim |
| Guideline documents and recommendations | Curated guideline content (schema in place, content pending) |
| Provider cache | De-identified provider responses |
Models and engines
- Query expansion
- Model or engine
- Gemini 3 Flash-Lite (lite tier)
- Notes
- Ontology expansion of the structured query
- Dense retrieval
- Model or engine
- BGE-M3 embeddings, 1024-dimensional (self-hosted)
- Notes
- Embedded per query; a hashed bag-of-words fallback keeps retrieval deterministic when the worker is absent
- Entailment
- Model or engine
- DeBERTa-v3-large MedNLI cross-encoder (self-hosted)
- Notes
- Rule-based verifier when the neural one is unavailable
| Step | Model or engine | Notes |
|---|---|---|
| Query expansion | Gemini 3 Flash-Lite (lite tier) | Ontology expansion of the structured query |
| Dense retrieval | BGE-M3 embeddings, 1024-dimensional (self-hosted) | Embedded per query; a hashed bag-of-words fallback keeps retrieval deterministic when the worker is absent |
| Entailment | DeBERTa-v3-large MedNLI cross-encoder (self-hosted) | Rule-based verifier when the neural one is unavailable |
Guardrails
- A source may support a claim only from a licensed abstract, open full-text passage or curated recommendation that states that direction (EVID-01).
- Retracted, withdrawn or superseded material fails closed (EVID-07).
- Sources must be zero-cost and require no paid subscription, enforced by a registry validator (EVID-12).
- Queries are de-identified before they leave the platform (EVID-11).
Rollout and status
- ProductionMulti-source retrieval, ranking, snapshots, claim linksEnabled: PubMed, Cochrane-indexed via PubMed, PMC Open Access Subset, Europe PMC, OpenAlex, DailyMed (FDA SPL), MedlinePlus, ClinicalTrials.gov, openFDA
- Rolling outNeural embeddings and entailmentServed by the self-hosted clinical LLM worker
- PlannedCurated guideline repositoryDisabled sources: SNOMED CT, Clinically approved guideline repository, Clinic protocol
Interfaces
- Consumes
- What
- Claims from Copilot and Clinical Documents
- Emits
- What
- Evidence snapshots
- Emits
- What
- Support relations per claim
| Direction | What |
|---|---|
| Consumes | Claims from Copilot and Clinical Documents |
| Emits | Evidence snapshots |
| Emits | Support relations per claim |
Invariants
- EVID-01…12
- Evidence benchmark contract: top-5 precision ≥ 95 %, recall ≥ 95 %, support-direction precision ≥ 98 %
Known limits
- There is no persistent vector index: embeddings are computed per query. A vector store is planned, not built.
- No application path yet writes curated guideline recommendations, so the guideline lane is empty.
SOURCE OF TRUTH
server/services/evidence/docs/evidence-grounded-generation.mddocs/evidence-claim-traceability.mddocs/evidence-benchmark-contract.md