A multi-tenant enterprise assistant that answers from each tenant's own documents, spreadsheets and figures — with a citation on every answer.

The challenge

How do you let business users ask questions of their own knowledge base — including the spreadsheets and figures generic document chat can't see — and trust the answer enough to act on it?

What we built

A conversational agent where tenants upload their own knowledge across eight file formats, and an ingestion pipeline parses, chunks, embeds and indexes it into three specialised search indexes — narrative text, extracted table rows, and figures with vision-generated captions — plus a database row store for pure spreadsheets, so exact-value lookups are a query, not a semantic guess. At question time an intent analyser classifies the question, a router decides whether the answer lives in documents or tables, and the agent fetches the exact rows, passages and figures before it writes a word. Latency was engineered as a programme, not a hope: zero-LLM regex fast paths for follow-ups, conversation history compressed from ~10,800 tokens to ~500, and prompts right-sized per query type.