ATR GPT is an enterprise RAG assistant helping 900+ employees access information distributed across 20+ departmental knowledge sources, through a policy-search mode and a general assistant mode.
Business Problem
Employees searched scattered documents and departmental repositories, and relied on direct communication with subject-matter experts to find information — a slow process that created knowledge silos across the organization.
Users
Employees across departments who need answers grounded in internal knowledge, without knowing which team or document holds them.
Constraints
Solution
I architected a FastAPI RAG platform backed by PostgreSQL with the pgvector extension for hybrid semantic-memory retrieval. A mode router directs each request to Oracle Cloud Infrastructure (OCI) GenAI for policy-document search or Anthropic Claude for general assistant conversation, with OpenAI used separately for image generation — chosen per request based on task and data sensitivity.
Architecture
Your Contribution
Technology Stack
Results
Challenges & Trade-Offs
Mode routing vs. a single fixed model
Routing policy-search requests to OCI GenAI and general assistant requests to Claude trades the simplicity of one fixed model for the flexibility of matching each request to the provider best suited to it — at the cost of maintaining a mode-routing layer instead of a single integration.
Confidentiality
Enterprise project. This page uses architecture diagrams and anonymized descriptions rather than internal interfaces or company data.