Evidence-grounded clinical retrieval
IrishClinicalRAG
A retrieval-first assistant for Irish clinical and biomedical guidance, built around traceable ingestion, hybrid search and exact chunk citations.
Clinical answers should be inspectable, not merely fluent. IrishClinicalRAG retrieves from a small, controlled corpus of Irish guidance and ties generated statements back to exact source passages.
How can an early-stage RAG system make source provenance, corpus integrity and insufficient evidence visible by design?
My contribution
- Built the ingestion, chunking, retrieval, reranking and source-grounded answer pipeline.
- Implemented corpus manifests, checksums, exact chunk citations and citation validation.
- Developed a FastAPI service, research interface, offline extractive baseline and configurable generation provider.
Signature view
Trace the evidence.
Move across the research chain. The case study below remains complete and readable without this interaction.
Question
Make evidence inspection part of the answer
Clinical retrieval systems need to show what they found, where it came from and when the evidence is insufficient.
Method, as a sequence of decisions
- 01
Immutable raw documents and manifests make corpus changes auditable.
- 02
Section-aware chunking retains metadata and produces 692 searchable passages.
- 03
BM25 and BGE-small dense retrieval are fused with reciprocal-rank fusion.
- 04
A cross-encoder reranks candidates before grounded answer generation and citation validation.
What the record actually supports
Controlled HSE/NCEC corpus
An implementation count, not a coverage claim for Irish medicine.
Indexed chunks
Section-aware passages with metadata and checksum-backed source integrity.
Repository-reported passing suite
Reported at 75.36% statement coverage; not independently rerun for this site.
Evidence: public repository README/status snapshot recorded 7 October 2026. Counts and timings are repository-reported, not independent clinical validation.
Choices and trade-offs
Why hybrid retrieval?
Lexical search is strong for exact clinical terminology; dense search can recover semantically related wording. Fusion reduces dependence on either signal alone.
Why an offline baseline?
An extractive path separates retrieval quality from generation-provider behaviour and keeps basic inspection possible without an API key.
Why exact chunk citations?
A citation is only useful if a reader can find the supporting text and determine whether the answer overstates it.
What has not been established
- The eight-document corpus is intentionally small and does not represent comprehensive clinical coverage.
- Formal retrieval and end-to-end RAG evaluation on the planned 100–150-case set is not complete.
- Repository smoke tests and latency measurements are engineering evidence, not clinical validation.
What I would investigate next
- Q1Which query strata expose the largest gaps in retrieval recall and citation entailment?
- Q2How should conflicting or superseded Irish guidance be represented?
- Q3Can evaluation distinguish a retrieval miss from an answer-generation failure?
Peer notes / ICRAG
See a missed assumption?
Research suggestions, methodological critique and relevant collaboration ideas are particularly useful.