PROJECT 03 / AI
RAG Database Research
A tested FastAPI, PostgreSQL and React learning playbook, plus a planned local RAG design for a healthcare referral system.
- FastAPI
- PostgreSQL
- pgvector
- React
- Local LLMs
Source private: not published
- Type
- AI
- Source
- Private
- Role
- Author of the playbook, working with Claude Code.
- Built with
- FastAPI, PostgreSQL, pgvector
01 / Overview
Overview
Built: a learning playbook
A 16-chapter FastAPI, PostgreSQL and React playbook, with chapters on LLM and RAG, auth, Redis caching and testing.
Planned: one referral database
Referrals from fax, phone and email would be centralized in one PostgreSQL database.
Planned: local RAG for reporting
A RAG pipeline with hybrid search and a reranker would automate the monthly reporting.
02 / Context and problem
Context and problem
This page keeps two things apart. Built is the learning playbook, my own project made with Claude Code. Planned is the referral system, which is a design only: there is no referral code, no reranker and no running system.
03 / My role
My role
Author of the playbook, working with Claude Code. The referral system is a plan, not shipped work.
04 / How it works
How it works
Built: re-run the tutorial code in three stages
The code was re-run with 9 tests (sync CRUD), then 11 (auth and relationships), then 14 (async, Redis and RAG).
Built: tests ran on SQLite and fakeredis
The Postgres-only SQL was compiled but not run on a live server, and no real model API was called.
Planned: one database
FastAPI and PostgreSQL with pgvector, holding the referrals that arrive by fax, phone and email.
Planned: hybrid search and reranker
Hybrid search is a planned playbook chapter. The reranker is part of the referral design only.
Planned: all local
A local embedding model and a local LLM, so the data never leaves internal servers.
05 / Results and numbers
Results and numbers
16
chapters in the Built playbook
9, 11, 14
tests per Built stage
06 / Limitations
Limitations
- Built: SQLite passing does not prove Postgres works. The Postgres-only SQL was never run on a live server, and the React snippets were never compiled.
- Planned: no referral code, no reranker and no results exist yet.
07 / What I learned
What I learned
- SQLite passing does not prove Postgres works, and library versions drift under tutorials, so verify by running.
08 / What's next
What's next
- Run the RAG code on real Postgres with pgvector.
- Add chapters on roles, multi-turn chat and hybrid search.
09 / Next project