graph LR
    subgraph BUILD["COMPILE KG WORKFLOW (offline, Databricks)"]
        direction LR
        P[paper] --> EX[Semantic Parser\nLLM] --> F[findingsIR] --> R1[Resolve\nper entity type] --> CY1[compile to\nCypher query]
    end

    subgraph QUERY["COMPILE QUERY WORKFLOW (online, website)"]
        direction LR
        Q[question] --> DE[Semantic Parser\nLLM] --> M[queryIR\nlogical form for eval] --> R2[Resolve\nper entity type] --> CY2[compile to\nCypher query]
    end

    CY2 -->|"execute cypher query"| NEO["KG Instance\non Neo4j"]
    CY1 -->|"execute cypher merge"| NEO
    NEO --> SG["subgraph\nfor eval"] --> SYN["LLM synopsis\nanswer for eval"]

    style M fill:#f3e8ff,stroke:#7c3aed,stroke-width:2px
    style SG fill:#f3e8ff,stroke:#7c3aed,stroke-width:2px
    style SYN fill:#f3e8ff,stroke:#7c3aed,stroke-width:2px
            

graph LR
    U["User\nquestion"] --> W["Website\nFastAPI"]
    W --> LLM1["LLM Service\ndecompose\n(PICO extraction)"]
    LLM1 --> CR["Concept Resolver\nregex → UMLS → LLM → embed\n(4-step, cached)"]
    CR --> NEO["Neo4j\nCypher query\n(broad recall)"]
    NEO --> LLM2["LLM Service\npertinence filter\n(concurrent fan-out)"]
    LLM2 --> LLM3["LLM Service\nsynopsis\n(structured citations)"]
    LLM3 --> U2["User\nevidence + synopsis"]

    style CR fill:#e3f2fd,stroke:#1976d2
    style NEO fill:#e8f5e9,stroke:#388e3c
    style LLM1 fill:#fff3e0,stroke:#f57c00
    style LLM2 fill:#fff3e0,stroke:#f57c00
    style LLM3 fill:#fff3e0,stroke:#f57c00
            

Every LLM call in the entire system goes through the LLM Service (cached gateway). Zero direct Azure OpenAI calls. One gateway, one cache, one place to observe.

graph LR
    subgraph CI["GITHUB ACTIONS"]
        direction LR
        PUSH["push to prod"] --> BUILD["Build\nDocker image\n→ ACR"]
        BUILD --> STAGING["Deploy to\nSTAGING"]
        STAGING --> TEST{"E2E Smoke Test\nALL blue pills\nvia /smoke/query"}
        TEST -->|"all pass"| PROD["Deploy to\nPROD"]
        TEST -->|"any fail"| BLOCK["BLOCKED\nprod not updated"]
    end

    subgraph APPS["CONTAINER APPS"]
        S["nobsmed-staging\n0-1 replicas"]
        P["nobsmed.com\n1-10 replicas"]
        MCPAPP["nobsmcpserver\n1 replica (always on)\n/mcp + /health"]
    end

    STAGING --> S
    PROD --> P

    subgraph MCP_CI["MCP SERVER CI/CD"]
        direction LR
        MPUSH["push to prod\napps/mcp-server/**"] --> MBUILD["Build\nDocker image\n→ ACR"]
        MBUILD --> MDEPLOY["Deploy to\nnobsmcpserver"]
        MDEPLOY --> MHEALTH{"Health check\n+ MCP handshake"}
    end

    MDEPLOY --> MCPAPP

    style BLOCK fill:#f8d7da,stroke:#dc3545
    style TEST fill:#fff3cd,stroke:#ffc107
    style PROD fill:#d4edda,stroke:#198754
    style MCPAPP fill:#e3f2fd,stroke:#1976d2
    style MHEALTH fill:#e3f2fd,stroke:#1976d2
            

Single MCP server serves all AI clients. 2 read-only clinical evidence tools. No Custom GPT Actions — MCP-first architecture.

graph TD
    KG[":KG name"] --> META[":KGMeta\nrole, schema_version"]
    META -->|HAS_BUILD| BUILD[":BuildInfo\nstatus, resolver, git_commit"]
    META -->|HAS_AUDIT| AUDIT[":AuditReport\nresolved_ratio, smoke_test"]
    META -->|HAS_SHOWCASE| PILL[":ShowcaseQuestion\nquestion, score, blue"]

    META --> EV[":Evidence (unified label)"]

    EV --> FIND[":Finding:Evidence\nclinical study"]
    FIND -->|tested| INT[":Intervention"]
    FIND -->|for_condition| COND[":Condition"]
    FIND -->|on_outcome| OUT[":Outcome"]
    FIND -->|reported_in| STUDY[":Study pmc"]

    EV --> EXP[":Experience:Evidence\nReddit / anecdotal"]
    EXP -->|tested| INT2[":Intervention"]
    EXP -->|for_condition| COND2[":Condition"]
            

The drivers below are payoffs of three decisions: (1) canonical identity (resolver + SNOMED) → dedup, IS-A recall, pooling; (2) ontology-as-code (one schema, projected + constrained) → zero drift + no-hallucination; (3) typed artifacts, not prose → debuggable reasoning, re-runnable lineage, gradeable output.