RAVANA

Grounded Retail Operations Assistant is a FastAPI web app that answers retail operations questions from internal documents with cited, source-backed answers. A user asks, for example, how long customers have to return an item with a receipt, how often high velocity items are counted, or when price changes take effect, and the app returns only sentences taken from retrieved sample documents, each with a document and section citation. Retrieval is hybrid: BM25-style keyword scoring plus TF-IDF cosine similarity in pure Python, followed by a reranking step using query-term coverage, a verbatim phrase bonus, and a section-heading bonus. If no passage supports a question, the app refuses instead of inventing an answer. An evaluation script checks that every citation excerpt is verbatim in the sample documents and that unsupported questions are refused; it passes 4 of 4 checks on the sample set. The demo uses 3 clearly labelled sample documents (inventory policy, merchandising guide, returns policy; 12 passages). It is a working demo, not a production system and not a real retailer's data.