Teams sit on thousands of PDFs — contracts, specs, reports — and the only way to find an answer is to open them one by one. Generic "chat with your PDF" tools don't hold up when documents are company-scoped, permissions matter, and answers need a citation.
A microservice document platform where upload, storage, embedding and retrieval are separate concerns:
Auth.Service — registration, login, JWT issuing/refresh (ASP.NET Core)Document.Service — upload, metadata, S3-compatible storage via MinIORAG.service — chunking, embeddings, pgvector similarity search, Gemini-backed answer generation with source citationsdocumind-client — Next.js web appUploads are I/O-bound and bursty; embedding is CPU-bound and queue-friendly. Splitting them means the RAG service's vector store and model dependency never sit in the upload path, and each side scales independently.
Internal service traffic has a stable, typed contract — binary protocol and generated clients beat JSON-over-HTTP there. Public APIs stay REST behind Nginx where debuggability matters more than throughput.
Most retrieval queries join embeddings with document metadata and company scope. Keeping both in one Postgres instance removes a database to operate and makes permission filters a plain SQL predicate instead of a two-store sync problem.
A working end-to-end platform — sign up, upload, ask a question, get an answer with the exact chunk it came from. Runs from a single docker compose up.
I can design service boundaries, argue a storage decision, wire a model provider in production shape, and ship the frontend that makes it usable.