Challenge

A heavy equipment service network needed technicians to surface the right procedure from a corpus of 4M+ pages of OEM PDFs in under a second, without hallucination.

Solution

  • PDF ingestion pipeline with OCR fallback for scanned manuals.
  • Hybrid retrieval: BM25 (part numbers, DTC codes) + dense embeddings (semantic).
  • Claude claude-sonnet-4-20250514 for generation with strict citation prompting.
  • Eval harness with 500 golden Q&A pairs, run on every deploy.

Results

  • P95 latency: 380ms
  • Escalations to Tier-2 support: −32%
  • Technician satisfaction score: +27 NPS points