AutoDefect Radar

AutoDefect Radar is a local retrieval-augmented generation (RAG) CLI for exploring vehicle recalls and consumer complaints published by the National Highway Traffic Safety Administration (NHTSA).

The application fetches public NHTSA records for a vehicle, redacts VINs from complaint text, builds a local Chroma vector index, and answers questions using a local Ollama model. Answers use retrieved records as evidence and include NHTSA campaign or ODI complaint identifiers.

AutoDefect Radar is an independent open-source project. It is not affiliated with NHTSA, vehicle manufacturers, or dealerships, and it does not provide safety, legal, repair, or diagnostic advice.

The CLI runs a local pipeline:

1. Fetch recall and complaint records from the public NHTSA API.
2. Normalize records into VIN-redacted Markdown documents.
3. Split documents and embed them with all-MiniLM-L6-v2.
4. Persist the embeddings in a local Chroma index.
5. Retrieve similar chunks and send them to a local Ollama model.

Raw responses, normalized documents, and the vector index are generated under data/ and deliberately excluded from Git.

• Python 3.12 or newer
• uv
• Ollama running locally

Install the default model before asking questions:

ollama pull gemma3:270m

uv sync --dev

The default corpus is the 2012 Honda Accord, including all recalls and the 100 most recently filed complaints.

uv run autodefect-radar ingest

Choose another vehicle or complaint limit with CLI options:

uv run autodefect-radar ingest \
  --make TOYOTA \
  --model CAMRY \
  --year 2020 \
  --complaints-limit 200

Use --refresh to bypass the local raw-response cache. Pass --complaints-limit -1 to index all returned complaints.

uv run autodefect-radar ask "What airbag issues appear?"
uv run autodefect-radar ask "Are there ABS module complaints?" --k 6

The ask command requires an index created by ingest and a running Ollama service.

By default, the answer prompt is relaxed so the model can synthesize from retrieved evidence without over-refusing. Set STRICT_ANSWER_PROMPT = True in src/config.py to restore strict context-only answering.

After building the index, launch the browser UI:

uv run autodefect-radar-ui

The UI calls the same answer path as the ask command and shows the request payload, answer, retrieved sources, and retrieved context.

NHTSA complaint records can contain vehicle identification numbers and free-form consumer narratives. AutoDefect Radar redacts VIN-shaped values when creating the normalized knowledge base. The original API cache is not modified, so the complete data/raw_cache/ directory remains local and ignored by Git.

The following generated directories are never committed:

• data/raw_cache/
• data/knowledge-base/
• data/vector_db/

Delete data/ and run ingest again to rebuild all local artifacts.

uv run pytest

The test suite uses temporary files, mocked HTTP responses, and deterministic fake embeddings. It does not require network access, Ollama, or downloaded embedding models.

• Retrieval uses vector similarity followed by a basic term-overlap guard. A weakly related question can still pass the guard when it shares one important term with retrieved text.
• Citation fallback reports retrieved evidence identifiers; it does not prove that every retrieved record contributed to the generated wording.
• NHTSA complaints are consumer reports and are not independently verified by this project.
• Generated indexes are local to the embedding model configured in the source.

Licensed under the MIT License.