RAG assistant for a retail chain

Where this started

Several shops and an online store, and the working knowledge split four ways: the stock system, Telegram, spreadsheets, and a few people heads. Anyone needing a price, a supplier term or a delivery date went asking.

Who we built it for

A board-game retail chain with several shops and an online store, whose working knowledge lived in a few people and a few dozen chats.

What fought back

1. 
   01

   The knowledge was in four places at once

   Part of it in the stock system, part in Telegram, part in spreadsheets, and part only in somebody head. A search that reads three of the four is not a search, it is a second thing to check.

2. 
   02

   Close enough is wrong here

   A price, a supplier term and a stock figure are either exact or useless, and a model that fills a gap with something plausible is worse than one that says it does not know. Hybrid search puts the exact fields beside the semantic ones, and the model answers from what it retrieved rather than from what it remembers.

3. 
   03

   An answer without a date is half an answer

   When the price was changed and when the delivery landed are the questions behind most of the other questions. Every answer carries the document it came from and the date it was true on, which is also what lets the manager see where the data still has holes.

What we actually built

All the sources collected and normalised, then RAG over a vector store with hybrid search, so the exact fields (stock, prices, article numbers) are matched exactly and the rest semantically. Telegram is the interface, because nobody was going to log into a seventh system. Access is by role, there is an admin panel for keeping the base current, and every query is logged. It does not just answer: it cites the document and the date behind the answer.

How the work went

1. Week 1

   Audit of the sources, the shape of the knowledge base, first test loads.

2. Week 2

   The RAG core, test answer scenarios, accuracy checked against known facts.

3. Week 3

   The Telegram interface, access rights by role, the admin panel, and the first round of feedback.

4. Week 4

   The real sources connected, the staff shown how it works, final corrections.

What came of it

Finding information got 8 to 10 times faster, and staff stopped chasing it through chats and spreadsheets. Built in three to four weeks.