A medical retrieval augmented generation prototype for somali symptom guidance with cross lingual evidence retrieval and technical verification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Background Somali remains underrepresented in evaluated health-oriented natural-language systems. Direct use of a general-purpose large language model for symptom questions is difficult to audit because answers may not be traceable to a defined medical source and multilingual performance may differ substantially from English. Objective This work asks whether an implemented cross-lingual retrieval-augmented generation prototype can accept Somali symptom narratives, retrieve English medical-textbook evidence, and return a structured Somali response with explicit safety boundaries and page-level citations. Methods Five English medical textbooks were extracted and cleaned, including OCR correction for CURRENT Medical Diagnosis & Treatment. Text was divided into overlapping 275-word chunks with 40-word overlap. A total of 64,177 chunks were embedded with OpenAI text-embedding-3-small (1536 dimensions) and indexed using FAISS IndexFlatIP over L2-normalized vectors. At runtime, a Somali query is translated into clinical English, checked by deterministic bilingual emergency rules, embedded, and matched using cosine similarity (top-k = 5; minimum score = 0.25). Retrieved passages are supplied to OpenAI GPT-4o-mini (API identifier: gpt-4o-mini) under an evidence-only prompt. A citation guard rejects unavailable or missing chunk identifiers, and the structured answer is translated back to Somali without changing urgency. The prototype uses Flutter and FastAPI. Results Deterministic software verification covered data preprocessing, document cleaning, chunk generation, embedding consistency, vector-index construction, retrieval execution, metadata alignment, citation validation, deterministic safety routing, API validation, and structured response generation. The defined verification scenarios completed successfully, and the final index contains 64,177 vectors with book, section, and page metadata. An illustrative end-to-end transaction confirmed that the implemented components could execute as an integrated pipeline. These findings establish functional software implementation only; no clinical accuracy, linguistic quality, user acceptance, model superiority, or deployment effectiveness is claimed. Conclusions The contribution is a reproducible prototype architecture that separates cross-lingual translation, deterministic emergency escalation, semantic retrieval, evidence-constrained generation, citation validation, and Somali presentation. The system has not undergone clinician-labelled triage evaluation or prospective clinical testing and must not be used for diagnosis, prescribing, or autonomous emergency decisions.