Building and Evaluating a RAG-based Menopause Information Chatbot

vira_health · Healthcare · 2023 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Clinical menopause information chatbot

Vira Health needed a chatbot that could answer menopause questions using clinically verified sources, with traceable grounding and safety checks strong enough for healthcare-adjacent use.

Constraints:

• General-purpose LLMs cannot guarantee clinically verified answers.
• Healthcare responses must minimize harmful advice and incorrect recall.
• The first evaluation used a narrow trusted-source set and required future expansion.
• Clinical correctness needed both AI and manual clinical review.

Architecture

Vira Health built a RAG chatbot for Stella using peer-reviewed Menopause Society position statements as trusted documents. The system retrieves the top five similar chunks, builds a context prompt, answers with GPT-4, and is evaluated on faithfulness, relevance, harmfulness, and clinical correctness across 40 open-ended questions.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  stella[Stella] --> rag-chatbot[RAG Chatbot]
  rag-chatbot --> retriever[Retriever]
  trusted-clinical-documents[Trusted Clinical Documents] --> retriever[Retriever]
  retriever --> context-prompt-builder[Context Prompt Builder]
  context-prompt-builder --> gpt-4-generator[GPT-4 Generator]
  gpt-4-generator --> evaluation-harness[Evaluation Harness]
  evaluation-harness --> clinical-reviewers[Clinical Reviewers]

Components

Data flow

A menopause question is embedded, matched against trusted clinical documents, the top five chunks are inserted into a context prompt, GPT-4 generates an answer, and AI plus clinician evaluators score the response for faithfulness, relevance, harm, and clinical correctness.

Why this architecture

The RAG design was chosen because clinical menopause support requires traceable, trusted sources rather than free-form model knowledge. The evaluation combined AI judging and clinical review because automated safety signals alone were insufficient for healthcare confidence.

Models used: GPT-4 · Claude 2.1 · ChatGPT

Scale

• The evaluation used 40 open-ended questions.
  “we designed 40 open-ended questions and we of used two methods to do this one we actually took the whole position statements”transcript, stated
• The retriever used top-k similarity with five trusted chunks.
  “top five um topk similarity of five so five CH selected trusted documents and used open AI gp4 model um at the time GPT for uh 0125 preview we also built a conversational interface using gradia around this so we”transcript, stated
• Automated faithfulness scored 95%.
  “this automated faithfulness score um scored very highly at 95% and so this gave us confidence alongside the high clinician reviews um the content here was being all the responses were being generated um almost exclusively using material from the”transcript, stated

Alternatives considered (roads not taken)

Rely on general-purpose model knowledge.

Rejected because: The product needed clinically verified and referenceable sources.

Use broad internet or unvetted medical content.

Rejected because: The first system constrained retrieval to peer-reviewed position statements.

Rely only on automated evaluation.

Rejected because: Clinical correctness showed stronger evidence in manual clinical evaluations.

Patterns

rag-retrieval (clinical QA)

The chatbot retrieves top-k trusted clinical document chunks and uses them to constrain GPT-4 responses.

safety-policy-enforcement (healthcare safety)

Responses are evaluated for harmfulness and clinical correctness before clinical deployment claims.

llm-as-judge (evaluation)

The evaluation used AI judges alongside manual clinical review.

human-in-the-loop-gate (clinical review)

Manual clinical evaluations checked clinical correctness and harmfulness.

Tradeoffs

Clinical grounding

Chose RAG over trusted clinical statements over Open-domain chatbot answers.

Rationale: The product needed verifiable clinical sources and referenceable answers.

Evaluation design

Chose AI judges plus manual clinical review over Single automated score.

Rationale: Clinical correctness and harm risk benefit from clinician judgment.

Corpus scope

Chose Small peer-reviewed statement set first over A broad medical corpus from day one.

Rationale: A narrow trusted set improves initial safety and faithfulness but needs expansion later.

Gotchas (what bit them)

Clinical chatbot answers require trusted provenance.

Root cause: Unconstrained model answers can be unverified or harmful in healthcare contexts.

Resolution: Restrict the corpus to peer-reviewed position statements and cite retrieved context.

Safety evaluation still requires further research.

Root cause: Biases and potential harms remain before integration into broader care workflows.

Resolution: Treat the prototype as a foundation and expand research before clinical deployment.

A narrow trusted corpus limits answer coverage.

Root cause: The prototype used only a small set of position statements.

Resolution: Integrate a wider range of trusted peer-reviewed sources.

Evaluation

Vira Health evaluated 40 open-ended questions with AI judges and manual clinical review across faithfulness, relevance, harmfulness, and clinical correctness.

Metrics: Faithfulness · Relevance · Harmfulness · Clinical correctness · Clinician review alignment

Outcomes

• The chatbot reached 95% automated faithfulness.
• Harmfulness was almost exclusively assessed as not harmful.
• Manual review found evidence of clinical correctness, while further research was still required.

The pilot used 40 open-ended questions, top-five retrieval, and a narrow trusted corpus of Menopause Society position statements.