Prism Case Study | Aryan Dani

A local enterprise AI agent that answers only from retrieved passages and refuses when support is weak.

Problem

Enterprise Q and A is often split across brittle domain bots or cloud RAG that leaks documents and ignores role boundaries at retrieval time. Teams need one conversational surface that stays honest about what it knows.

Approach

Prism routes one query across curated domains plus session uploads, filters Chroma chunks by role metadata before the model sees them, and refuses to invent when retrieval confidence is low. The same grounded answer can be exported as prose, JSON, XML, Excel, or email without asking the model again.

Architecture

React and Vite UI with a FastAPI backend, embedded Chroma, Ollama embeddings and generation, session memory, and demo RBAC logins. A CanonicalAnswer object fans out to multiple export formats from one retrieval pass.