Automotive Standards Intelligence
Five specialised AI agents working together to answer questions from dense technical regulations, accurately and with sources.
Overview
Automotive standards documents are long, dense, technical PDFs. Finding a specific requirement means searching hundreds of pages, and a generic chatbot will confidently invent answers. This showcase build demonstrates how TAKE OVER makes AI reliable on exactly this kind of high-stakes document.
Instead of a single model answering a question, the system is a pipeline of five specialised AI agents, each with one job. Together they decide what to search, retrieve the right passages, rank them by relevance, write a grounded answer, and check that answer before it reaches the user. Retrieval is hybrid: it combines semantic vector search with keyword (BM25) search, so it catches both the meaning of a question and the exact technical terms and clause numbers engineers use.
The system was measured as well as built. It was evaluated with RAGAS, the standard framework for scoring RAG quality, and shipped with full documentation and a GitHub deployment. It grew out of an earlier version built on a Workflow-Agent-Tools (WAT) architecture, which shaped the final multi-agent design.
What we built
Five-agent pipeline (LangGraph)
- Orchestrator: interprets the question and directs the workflow.
- Retriever: pulls candidate passages from the standards library.
- Reranker: re-scores passages with a cross-encoder for precision.
- Generator: writes the answer grounded only in the retrieved evidence.
- Evaluator: checks the answer's quality and faithfulness before returning it.
Hybrid retrieval
ChromaDB vector search combined with BM25 keyword search.
Cross-encoder reranking
Puts the most relevant clauses first, reducing noise and hallucination.
Measured quality
Evaluated with the RAGAS framework for faithfulness and relevance.
Production-style delivery
A FastAPI backend, a Streamlit interface, full documentation and a GitHub deployment.
Iterative architecture
Evolved from an earlier WAT (Workflow-Agent-Tools) implementation into a multi-agent system.
Built with
- Python
- LangGraph
- LangChain
- ChromaDB
- BM25
- Cross-encoder reranking
- OpenAI GPT-4o-mini
- FastAPI
- Streamlit
- RAGAS evaluation