Image Extraction Pipeline

SequentialAgent containing 7 specialist LLM agents, with a nested LoopAgent, processes clinical images through quality assessment, OCR, AI vision analysis, structured field extraction, critic/refiner validation, and review-request preparation. Clinician review and persistence remain external deterministic boundaries. Lanes: Agents, Validation Loop, Clinician.

flowchart TD
    subgraph LANE_AG["Lane: Pipeline agents"]
        QA["quality_assessor_agent (flash-lite)
assess_image_quality"]
        OCR["ocr_processor_agent (flash-lite)
extract_clinical_text"]
        VA["vision_analyzer_agent (pro-customtools)
analyze_clinical_image"]
        CS["clinical_structuring_agent (pro)
structure_clinical_findings, store_to_gcs"]
        RR["clinical_review_request_agent (flash-lite)
prepare pending review packet"]
    end

    subgraph LANE_LOOP["Lane: validation_gate (LoopAgent, Day 1b)"]
        CRIT{extraction_critic_agent
confidence >= threshold?}
        REF["extraction_refiner_agent
flag_for_review (fields < 0.80)"]
        EXIT["exit_loop"]
    end

    subgraph LANE_HITL["Lane: Clinician (HITL, Day 2b)"]
        GATE{Clinician review
transition_extraction_review}
        PERSIST["persist_extraction_relational
persist_extraction_vector"]
        DISCARD["Discard + reason logged"]
    end

    START([Clinical image uploaded]) --> QA --> OCR --> VA --> CS --> CRIT
    CRIT -->|below threshold| REF --> CRIT
    CRIT -->|passes| EXIT --> RR --> GATE
    GATE -->|approve| PERSIST --> DONE([Stored + timeline updated + audit event])
    GATE -->|reject| DISCARD --> DONE2([Rejected, audit logged])

Key facts:

output_key plumbing: each agent writes its result to session.state; the next stage reads it — no direct coupling (Memory Layers Layer 2).
The LoopAgent repeats critic → refiner until field confidence passes the threshold; fields below 0.80 are flagged for review.
The seven LLM specialists end with clinical_review_request_agent; clinician approval, persistence, and audit are product/tool boundaries rather than extra agents.
State machine: transition_extraction_review only allows needs_review → approved or needs_review → rejected. Persistence tools require an approved review receipt — see Human-in-the-Loop Approval.