Patient QA Pipeline

SequentialAgent containing 8 specialist LLM agents answers clinical questions with cited evidence from notes, images, and vector search — grounded in patient context.

flowchart TD
    START([Patient-scoped question]) --> VR
    VR["qa_request_validation_agent (flash-lite)
validate_qa_request"] --> CA
    CA["context_assembly_agent (flash-lite)
lookup_patient_record, load_memory,
search_past_conversations"] --> ER
    ER["evidence_retrieval_agent (pro-customtools)
search_clinical_notes, search_vector_store,
retrieve_imaging_evidence"] --> IE
    IE["image_evidence_agent (pro-customtools)
analyze_evidence_images, fetch_image_from_gcs"] --> CB
    CB["citation_builder_agent (flash-lite)
build_citations"] --> AS
    AS["answer_synthesis_agent (pro)
compose_clinical_answer"] --> AU
    AU["qa_audit_agent (flash-lite)
log_audit_event, save_qa_to_memory"] --> RS
    RS["qa_response_agent (flash-lite)
return cited answer"] --> DONE([Cited answer + audit trail])

Key facts:

Validation happens first: validate_qa_request rejects out-of-scope requests before any retrieval.
Evidence is multimodal — text notes, vector search hits, and imaging pulled from GCS and analyzed by Gemini Vision.
Every answer carries citations built before synthesis, so sources are traceable in the UI (Clinical App).
The audit stage persists the exchange to long-term memory (save_qa_to_memory) with PHI filtered first (Memory Layers Layer 3).