RAG Studio (Private-by-Default)

A local Retrieval-Augmented Generation (RAG) app with:

• CLI query (rag.py)
• Browser UI via FastAPI (web_app.py)
• Streamlit UI (streamlit_app.py)

It uses:

• ChromaDB for vector search
• SentenceTransformers embeddings
• Local Llama GGUF (llama-cpp-python) or OpenAI for generation

Project layout

• resources/documents/ - your source files (.txt, .md, .pdf)
• resources/public_demo/ - safe sample docs used for Streamlit public demo bootstrap
• resources/document_catalog.json - auto-generated catalog (ignored by git)
• resources/DOCUMENT_INDEX.md - auto-generated markdown index (ignored by git)
• chroma_db/ - vector index (ignored by git)

Quick start (local)

cd RAG
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements_local.txt

Add documents to resources/documents/, then index:

python ingest.py

Ask from CLI:

python rag.py "What is time series forecasting?"

Run UIs

FastAPI web UI:

python web_app.py

Open http://127.0.0.1:8765.

Streamlit UI:

streamlit run streamlit_app.py

Streamlit Cloud deployment

Use streamlit_app.py as the app entry point. This repo includes:

• requirements.txt (cloud-safe, faster deploy)
• requirements_local.txt (full local stack: FastAPI + local Llama)
• runtime.txt pinned to Python 3.11 for dependency compatibility.

Recommended public-demo environment/secrets:

• PUBLIC_DEMO_MODE=1
• AUTO_BOOTSTRAP_DEMO=1 (auto-ingest resources/public_demo/ when catalog is empty)
• SHOW_DOC_METADATA=0
• ALLOW_USER_UPLOAD=0
• ALLOW_USER_REINDEX=0

These defaults prevent exposing document names, chunks, and write operations in public demos.

If using OpenAI:

• set OPENAI_API_KEY in Streamlit secrets.

Privacy notes

• .gitignore excludes private docs and generated indexes:
  • resources/documents/* (except .gitkeep)
  • resources/document_catalog.json
  • resources/DOCUMENT_INDEX.md
  • .env
  • .streamlit/secrets.toml
• Keep private files only on your local machine or private storage.

Optional local Llama model

Download default GGUF:

python download_llm.py

Then query with local backend (default) or set LLM_BACKEND=openai for cloud model.