Overview
The application organizes documents and operational knowledge, generates embeddings, retrieves relevant chunks from PostgreSQL/pgvector and streams answers with source citations.
Problem
Knowledge scattered across documents and tools is difficult to find and verify. A language model alone does not provide enough control over the provenance of an answer or the surrounding business process.
Solution
An ingestion pipeline chunks content, creates embeddings and retrieves context before Gemini answers. Citations, conversation history, website import and adapter-based mail, OCR, workflow and integration modules connect AI to conventional business software.
My role
I designed the full-stack architecture, data model, RAG ingestion and retrieval, streamed responses, provider boundaries and container deployment with private storage.
Architecture
Next.js combines UI and API, Prisma manages PostgreSQL 16, and pgvector stores 768-dimensional vectors. Provider interfaces isolate Gemini, OCR, mail and integrations; Docker Compose runs the app and database without exposing the DB publicly.
Engineering challenges
AI boundaries are explicit: answers depend on retrieved context, citations allow verification and document extractions require review. Rate limiting for selected endpoints remains identified hardening work.
Results and current status
Verified functionality includes upload, embeddings, retrieval and cited chat, plus demo OCR and business-module flows. Any future public demo will be sanitized; private data, prompts and organizational processes are not disclosed.