DB Intelligence Pipeline

SequentialAgent containing 6 specialist LLM agents translates natural language into safe SQL with validation and explicit approval preparation before execution, followed by chart-spec generation. Lanes: Agents, Safety Gate.

flowchart TD
    subgraph LANE_AG["Lane: Pipeline agents"]
        SD["schema_discovery_agent (flash-lite)
get_database_schema"]
        NL["nl_to_sql_agent (pro)
generate_sql"]
        EX["query_executor_agent (pro)"]
        IC["insight_chart_agent (pro)
generate_chart_spec, save_query_to_memory"]
    end

    subgraph LANE_GATE["Lane: Safety gate"]
        SV{sql_validator_agent (flash-lite)
validate_sql_safety}
        APPR{sql_preview_approval_agent (pro)
approve_sql_preview}
    end

    START([Natural-language question]) --> SD --> NL --> SV
    SV -->|unsafe| REJ([Rejected + logged])
    SV -->|safe| APPR
    APPR -->|direct| RUN["execute_clinical_query"]
    APPR -->|preview approved| RUN2["execute_approved_clinical_query"]
    RUN --> IC
    RUN2 --> IC
    IC --> DONE([Table / chart / CSV export + history + audit])

Key facts:

SQL never executes without passing validate_sql_safety; the product exposes a read-only SQL preview with an explicit execution boundary (Problem and Solution).
The executor is pro-customtools tier; generation is pro tier (Model Registry).
Results feed generate_chart_spec for visual output, and the query is persisted to memory (save_query_to_memory) for later recall.
Schema and query engine live in clinical_schemas.py and database.py (Module Reference).