Problem
7-Eleven technicians repair coffee machines, ovens and other revenue-generating equipment in stores, and every hour of downtime is lost sales. The manuals sit in a central repository of 600+ documents (20,000+ pages, some PDFs several hundred pages long). A technician has only a phone, so finding the right page is slow. New technicians forget their training and call the regional trainer, which adds delay.
Example from the article: a technician asks about part "56-2682/03". The internal catalogue does not have it; the web shows it is discontinued and replaced by a new number.
Solution in brief
A chatbot inside Microsoft Teams routes each question through guardrails and an intent-detecting handoff agent to catalogue, parts or troubleshooting agents. They retrieve chunks with hybrid search (BM25 plus embeddings, fused by reciprocal rank fusion) and fall back to the web when internal data is missing. MLflow Traces and a dashboard track agents and adoption.
Questions this case answers
- How to build a question-answering assistant over hundreds of long, image-heavy PDF manuals for mobile users who need one exact page quickly
- How to combine keyword and embedding retrieval rankings and fall back to web search when the internal reference data is outdated
- How to build a RAG assistant over hundreds of long technical PDF manuals for field technicians who only have a phone
- How to combine keyword (BM25) and embedding retrieval with reciprocal rank fusion to get more relevant chunks
Tools and methods
- Databricks Vector Search (Mosaic AI) — knowledge base with managed hybrid retrieval
- Reciprocal rank fusion (RRF) — merges BM25 and embedding rankings
- Multimodal LLM PDF-to-text conversion — custom prompt for messy layouts
- Guardrails and handoff agent — relevance filter, then routing by intent
- Web search agent — limited to specific websites
- MLflow Traces, AI/BI dashboard — agent traces, token cost, usage KPIs
Google or a one-time email code. No password.