RAG

Two LangChain retrieval demos: a FastAPI/LangServe service that answers questions about a company's rules and screens emails for phishing, and a notebook that answers questions about a SQL database by writing and running the query itself.

Overview

Both halves are built on the same idea — put real context in front of the model instead of trusting it to remember.

The service (app/) exposes two chains. /rules is a classic RAG pipeline: rules.txt is chunked, embedded with a local MiniLM model, stored in Chroma, and the chunks most relevant to your question are handed to Gemini 1.5 Flash along with the rlm/rag-prompt template from LangChain Hub. /emailchecker sends an email body through a local Ollama llama3.1 with a prompt that asks for nine phishing checks, each answered true/false with an English and an Arabic justification, returned as one JSON object.

The notebook (notebooks/sql_rag.ipynb) does retrieval over a database instead of documents: it feeds the live Chinook schema to Gemini, has it write a SQL query, runs that query, then feeds the schema, question, query and result back to Gemini to phrase an answer in plain English.

Features

• Rules Q&A (/rules) — RAG over app/rules.txt with MiniLM embeddings, a Chroma vector store, and Gemini 1.5 Flash.
• Phishing checker (/emailchecker) — nine checks (grammar, urgency, call-to-action, confidential-info requests, phishing phrases, fake alerts, fake offers, personalization, suspicious character encoding), each with a bilingual justification, run against a local Ollama model.
• Two ways to call each chain — LangServe's interactive playground (/rules/playground, /emailchecker/playground) and plain JSON endpoints (POST /api/rules, POST /api/emailchecker).
• Text-to-SQL (notebooks/sql_rag.ipynb) — schema → generated SQL → executed query → natural-language answer, over the Chinook sample database.
• Containerized — Dockerfile and Compose file for the service.

Tech stack

Python · FastAPI · LangServe · LangChain · Chroma · sentence-transformers (all-MiniLM-L6-v2) · Google Gemini 1.5 Flash · Ollama (llama3.1) · MySQL · Docker

Getting started

Prerequisites

• Python 3.11+
• A Google AI Studio API key (for the /rules chain and the notebook)
• Ollama with llama3.1 pulled (for /emailchecker): ollama pull llama3.1
• MySQL, only for the SQL notebook

Install

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure

Both GOOGLE_API_KEY and DATABASE_URL are read straight from the environment. Copy the template and fill it in:

cp .env.example .env
export $(grep -v '^#' .env | xargs)

Run

uvicorn app.server:app --reload --port 8000

Open http://localhost:8000/rules/playground to ask about the rules, or http://localhost:8000/emailchecker/playground to screen an email.

With Docker

docker compose up --build

The Compose file passes GOOGLE_API_KEY through from your shell. Note that /emailchecker calls Ollama on localhost, which inside a container means the container itself — so that endpoint needs Ollama reachable from the container to work.

Usage

curl -X POST http://localhost:8000/api/rules \
  -H 'Content-Type: application/json' \
  -d '{"text": "What is the latest I can arrive?"}'

{ "response": "You have to come in by 10:30 am." }

curl -X POST http://localhost:8000/api/emailchecker \
  -H 'Content-Type: application/json' \
  -d '{"text": "URGENT: verify your account within 24 hours or it will be suspended."}'

Returns one JSON object per check, e.g.:

{
  "urgencyCheck": {
    "result": true,
    "english_justification": ["The email demands action within 24 hours."],
    "arabic_justification": ["يطالب البريد باتخاذ إجراء خلال 24 ساعة."]
  }
}

Tests

The tests exercise the live API, so start the server (and Ollama) first. They skip themselves if nothing is listening on port 8000.

python -m unittest discover tests

Project structure

app/
  server.py     FastAPI app: routes, the phishing-check prompt, the Ollama model
  rag.py        the /rules RAG chain (chunk → embed → Chroma → Gemini)
  rules.txt     the company rules the /rules chain retrieves over
notebooks/
  rules_rag.ipynb   the RAG chain, built up step by step
  sql_rag.ipynb     text-to-SQL over Chinook
data/
  Chinook_MySql.sql sample database for the SQL notebook
tests/
  test_api.py   end-to-end tests against /api/emailchecker

SQL notebook

Load the sample database, then point DATABASE_URL at it:

mysql -u root -p < data/Chinook_MySql.sql

License

MIT