Multi-agent EV ingestion Lambda (SQS -> Postgres)

This repo now ships a deployable Lambda that keeps the multi-agent guardrail approach and persists curated data into your PostgreSQL DDL (vehicles, specs).

What the Lambda does

For each SQS message:

1. PrimaryAgent: reads raw_extraction from message.
2. GapDetector: finds missing required fields.
3. SearchAgent: reads enrichment_candidates and filters by trusted domain allowlist.
4. VerifierAgent: fail-closed verification (contradictions and weak evidence are unresolved).
5. MergerAgent: merges verified values, nulls on conflicts.
6. CommitAgent: emits final record + explicit null_reasons.
7. Persists to Postgres with upserts into vehicles and specs.

Required env vars

Either set:

• DB_DSN

Or set:

• DB_HOST
• DB_PORT (optional, defaults to 5432)
• DB_NAME
• DB_USER
• DB_PASSWORD
• DB_SSLMODE (optional, defaults to require)

Optional:

• TRUSTED_DOMAIN_ALLOWLIST (comma-separated list)

SQS message format

{
  "request_id": "req-123",
  "market": "IN",
  "source": "firecrawl-gpt",
  "fetched_at": "2026-01-01T10:00:00Z",
  "source_url": "https://oem.example.com/model",
  "raw_extraction": {
    "make": "Tata",
    "model": "Nexon EV",
    "variant": "Creative+",
    "body_type": "SUV",
    "seating": 5,
    "status": "active",
    "notes": "variant dependent charging",
    "specs": {
      "price_ex_showroom_inr": 1999000,
      "range_km": 465,
      "battery_kwh": 45,
      "motor_power_kw": 106,
      "motor_torque_nm": 215,
      "motor_type": "PMSM",
      "battery_type": "LFP",
      "charge_ac_kw": 7.2,
      "charge_dc_kw": 60,
      "charger_options": "3.3kW/7.2kW",
      "accel_0_100_s": 8.9
    }
  },
  "enrichment_candidates": {
    "specs.range_km": [
      {
        "url": "https://oem.example.com/specs",
        "value": 465,
        "source_type": "oem_official"
      },
      {
        "url": "https://trusted-autospecs.example.net/specs",
        "value": 465,
        "source_type": "trusted_specs"
      }
    ]
  }
}

Deploy on AWS Lambda

1) Package code + dependency layer

Use either psycopg or psycopg2 in your deployment artifact.

Example (zip deployment, psycopg2):

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install psycopg2-binary -t package/
cp lambda_function.py package/
(cd package && zip -r ../lambda_bundle.zip .)

If you need src/ utilities in the future, copy src/ into package/ as well.

2) Create/update Lambda

• Runtime: Python 3.11+
• Handler: lambda_function.lambda_handler
• Upload lambda_bundle.zip
• Set env vars listed above
• Ensure Lambda network can reach your Postgres (VPC/subnets/security groups as needed)

3) Configure SQS trigger

• Add SQS event source mapping
• Enable Report batch item failures
• Choose batch size according to DB throughput (start with 5–10)
• Configure SQS DLQ redrive policy for poison messages

4) Smoke test

Use direct invoke payload (same shape as a message body). For SQS path, send the body JSON to queue and verify CloudWatch logs.

Local sanity check

python -m py_compile lambda_function.py