Insurance Document Audit

This example shows a SQL-first pattern for insurance document review.

Vane Data does not ship a dedicated insurance workflow. Use SQL for deterministic checks, map_batches(...) for extraction or normalization, and prompt(...) when a document needs model-assisted structured review.

Goal

Audit insurance documents by extracting relevant text, applying deterministic SQL rules, and optionally using an LLM for structured review.

Input shape

Recommended input columns:

• document_id
• policy_id
• claim_id
• document_type
• text
• source_uri

OCR, document parsing, and policy-system extraction should happen upstream or in explicit UDF stages.

SQL rule pass

import vane


con = vane.connect()


docs = con.sql("""
    select *
    from read_parquet('data/insurance_documents/*.parquet')
    where text is not null
""")


docs.to_table("docs")


rule_hits = con.sql("""
    select
        document_id,
        claim_id,
        case
            when lower(text) like '%missing signature%' then 'missing_signature'
            when lower(text) like '%expired%' then 'expired_reference'
            else null
        end as rule_hit
    from docs
    where lower(text) like '%missing signature%'
       or lower(text) like '%expired%'
""")

Structured model pass

prompt(...) returns the configured output column. Explicitly combine that column with source metadata when the final output needs identifiers and document attributes, and validate row counts before writing the result.

from pydantic import BaseModel


class AuditResult(BaseModel):
    status: str
    reason: str


audit_only = docs.prompt(
    "text",
    provider="openai",
    system_message="Audit the insurance document for missing evidence. Return JSON.",
    return_format=AuditResult,
    output_column="audit_json",
    execution_backend="subprocess_actor",
)


docs_table = docs.to_arrow_table()
audit_table = audit_only.to_arrow_table()
audited = con.from_arrow(docs_table.append_column("audit_json", audit_table["audit_json"]))

Output shape

Recommended output columns:

• document_id
• claim_id
• document_type
• rule_hit
• audit_json
• source_uri

Validation

Before using this workflow in production:

• Keep deterministic rule hits auditable.
• Version prompts and model choices.
• Store source document references.
• Sample false positives and false negatives.
• Review model output before it drives operational decisions.

Scope notes

This page does not define legal advice, claims policy, OCR, insurance ontology, or policy-system integration. Add those capabilities upstream or as explicit UDF stages.