Skip to content

Healthtech

Clinical RAG

A retrieval-augmented answering system for clinical knowledge where every answer links to its source — built so a clinician can verify the citation, not just read a plausible summary.

  • pgvector
  • LlamaIndex
  • OpenAI
0

hallucinated citations in eval

source-linked answers

100%

In short

  • Every answer is composed only from retrieved passages, each linked to its approved source.
  • When the corpus doesn't support a claim, the system says so instead of guessing.
  • Citation support is scored on every change — an unbacked citation fails the gate before it ships.

The engagement

Sector
Healthcare / clinical tooling
Company shape
Growth-stage, regulated domain
Engagement
Build + eval harness

Anonymised by policy

Every engagement runs under NDA — we protect our clients' confidentiality as a matter of policy, and we'd extend the same discretion to your work. The constraints, the shape, and the outcome here are real; the identity is redacted.

The challenge

Clinicians won't act on a summary they can't verify. A generic chatbot that produces fluent but unsourced answers is worse than useless here — it's a liability. The system had to answer from an approved corpus and show its work for every claim.

The approach

  1. 1
    Source-linked retrieval

    Every answer is assembled from retrieved passages, each rendered with a link back to the approved source document.

  2. 2
    Citation-level evaluation

    The eval set scores not just answer quality but whether each cited passage actually supports the claim — hallucinated citations fail the gate.

  3. 3
    Refuse over guess

    When retrieval finds no supporting source, the system says so rather than filling the gap with a plausible guess.

The solution

Queries retrieve from an approved corpus embedded in pgvector; answers are composed only from retrieved passages and rendered with inline citations. The eval harness verifies citation support on every change, and low-support answers are suppressed in favour of an explicit 'no supported answer' — so trust is earned per response.

The system we built

  1. Queryclinician question
  2. Retrieveapproved corpus, pgvector
  3. Compose + citeanswer from sources only
  4. Citation evalsupport-checked
  5. Answer or refuseno unsupported claims
Abstract by design — the architecture we built, not a client screen. No confidential data is shown.
Illustrative
An answer with inline citations back to the approved source
Illustrative
Walkthrough: query → retrieve → cite → answer or refuse

The results

hallucinated citations in eval

0

on the citation eval set

source-linked answers

100%

every shipped answer

Data & safety

Evaluation
A citation-support eval checks that each cited passage actually backs the claim; failures block the release.
Guardrails
Answers draw only from the approved corpus and refuse when retrieval finds no support.
Data handling
Retrieval runs over an approved corpus; no patient data is sent to third-party models.

The stack

  • pgvector
  • LlamaIndex
  • OpenAI
Capability: RAG & Knowledge Systems

Questions we get

Every answer is composed only from retrieved passages and rendered with inline links back to the approved source document, so each claim can be checked against its citation.

The system returns an explicit 'no supported answer' rather than generating a plausible guess — refusing to answer is treated as the correct behaviour when the corpus doesn't support a claim.

The ingestion pipeline re-indexes on a schedule (or on change), so retrieval always reflects the current approved corpus rather than a stale snapshot.

An eval harness scores citation support on every change — an answer whose citations don't actually back the claim fails the gate before it ships.

More selected work

// Let's build

Want an outcome like this on your problem?

Tell us where you are. We reply within a day with a concrete next step.