横展 · Yokoten — spread the lesson sideways
New engineers find the lesson before they repeat the failure.
An AI copilot over an automotive supplier's engineering record — 8D reports, lessons learned, FMEAs, test reports, design reviews, change notices, supplier quality data and scanned drawings. Natural-language questions, cited answers, measured accuracy.
The problem
Engineering knowledge lives in hundreds of reports, spreadsheets and scans. New engineers spend hours searching for past failures and the lessons behind them — and some lessons are only rediscovered after the same failure happens again.
The approach
A retrieval-augmented chatbot that searches the whole record in natural language, answers only from the documents, cites every claim, checks its own sentences for support and says so when the answer is not there.
Measured on a 129-question golden set
Loading evaluation results…
What it does
Cited answers
Every sentence links to the page and passage it came from; the viewer highlights it.
Knows when it doesn't know
A calibrated relevance gate abstains instead of guessing, and shows the closest documents.
Hybrid semantic search
Dense embeddings + BM25 fused with RRF, cross-encoder reranking, facets, score breakdown.
Scans and drawings
OpenCV denoise, deskew and threshold before OCR; title-block fields become metadata.
Revision aware
Prefers the latest revision of a document and tells you when an older one said otherwise.
Measured, not claimed
A golden set of 129 questions, ablations and hallucination checks - numbers on the Evaluation page.
Architecture: every step is visible in the trace
- 1Condense follow-up
- 2Expand acronyms
- 3Extract filters
- 4Route intent
- 5Dense + BM25 → RRF
- 6Cross-encoder rerank
- 7Relevance gate
- 8Small-to-big context
- 9Versioned prompt
- 10Streamed answer
- 11NLI check per sentence
- 12Confidence + trace
Ingestion parses PDF, Word, Excel, CSV, Markdown and images into layout elements with page coordinates, runs OCR on scans, chunks four ways and indexes into FAISS (flat and HNSW), ChromaDB and BM25. Answers come from Groq, Gemini, a local Ollama server or an in-process Hugging Face model.
Built with
Built by Ashwani Yadav as a portfolio project for an AI chatbot for engineering knowledge management internship. All documents describe a fictional supplier, Norvane Automotive Systems. Source code on GitHub.