AI RAG Mongo + OpenAI Demo

A minimal RAG app using:

• Flask (single-file app with inline UI)
• MongoDB 4.4 (Docker, Apple Silicon friendly)
• OpenAI embeddings + chat
• LangChain text splitters (for chunking)

Upload a PDF/DOCX, generate embeddings, and ask questions over the uploaded knowledge.

Prerequisites

• macOS with Docker Desktop
• Python 3.11+
• OpenAI API key

Quick Start

1. Configure environment

Create .env in the project root. Example:

MONGO_INITDB_ROOT_USERNAME=mongo
MONGO_INITDB_ROOT_PASSWORD=mongo
MONGO_HOST=localhost
MONGO_PORT=27017
MONGO_DB_NAME=ragdb
VECTOR_COLLECTION_NAME=documents
OPENAI_API_KEY=sk-...your-key...
# Point to a custom OpenAI-compatible server (optional)
# Example for local server: http://192.168.1.238:8000/v1
OPENAI_BASE_URL=
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
OPENAI_CHAT_MODEL=gpt-4o-mini
FLASK_HOST=127.0.0.1
FLASK_PORT=8000
FLASK_DEBUG=0

2. Start MongoDB + seed sample docs

chmod +x setup.sh
./setup.sh

3. Create venv, install deps, run app

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt
export $(grep -v '^#' .env | xargs)
python app.py

Open http://127.0.0.1:8000

How To Demo (5–7 minutes)

• RAG = Retrieve relevant chunks + Generate answer with LLM.
• Show app UI:
  • Left: Upload a document (PDF/DOCX)
  • Right: Ask a question (search)
• Upload content:
  • Use the provided sample demo/sample_faq.md converted to DOCX/PDF (see below) or your own doc
  • After upload, app chunks text, calls OpenAI embeddings, stores vectors in MongoDB
• Ask a question:
  • e.g., "How do I start the database?" or "Which model generates embeddings?"
  • App embeds the query, finds similar chunks in MongoDB, sends context + question to Chat model, and returns the answer
• Mention fallback: If context doesn’t contain the answer, it falls back to general LLM answer

Preparing the demo document

• A sample is in demo/sample_faq.md. Optionally export it as DOCX or PDF:
  • Open it, copy into Word/Pages → Save as DOCX/PDF
  • Or use a markdown-to-pdf tool

Suggested Questions

• What database and vector approach does this app use?
• How do I start the database?
• Which model is used for embeddings?
• What file types can I upload?

Troubleshooting

• Apple Silicon + MongoDB: We use mongo:4.4 with platform: linux/amd64 in docker-compose.yml.
• If setup.sh times out, run docker logs rag-mongodb to ensure Mongod is listening.
• If OpenAI client errors about proxies: we pinned httpx/httpcore in requirements.txt.
• Reinstall deps after edits:

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

Tech Notes

• Chunking: LangChain RecursiveCharacterTextSplitter
• Embeddings: OpenAI text-embedding-3-small
• Chat: OpenAI gpt-4o-mini
• Storage: MongoDB collection documents with text + embedding per chunk
• Similarity: cosine similarity computed in app over a capped set of docs

Using a local OpenAI-compatible server

• Set OPENAI_BASE_URL to your server’s URL, including /v1, e.g. http://192.168.1.238:8000/v1.
• Keep OPENAI_API_KEY set (any non-empty value works if your server ignores it).
• The app will route both embeddings and chat completions to that base URL.

Security

• Secrets via .env only (do not commit real keys)
• Do not expose this demo publicly without appropriate security