RAGdio is an audio-focused RAG (Retrieval-Augmented Generation) framework combining transcription, semantic search, and multiple LLM backends — both local and cloud-based.

💡 Built with clean code, adapters, interfaces, and a minimal full-stack (FastAPI + Vite).

RAGdio is intended for local use, testing, and development purposes.
It's not production-ready and not designed for multi-user environments (yet).

• ❗ Great for experimenting with audio-to-RAG pipelines locally.
• ⚙️ Designed with modularity in mind — adapters, interfaces, and clean architecture.
• 🚫 Not optimized for production load, scaling, or secure multi-user handling.
• 📌 Requires manual setup of models (e.g., in Ollama) after first run.

Chat with memory/history feature (available in API and Local models)

git clone https://github.com/Kotmin/RAGdio
cd RAGdio

# 1. Configure environment variables
copy from examples or manually set up

# 2. Start services
docker compose up --build

Once everything is up and running, you can access:

• 🖥️ Frontend: http://localhost:5173
• 📡 Backend (API docs): http://localhost:8000/docs

The default docker-compose setup includes an Ollama server without any model pre-installed.
After running docker compose up, you must install a model manually.

You can install Zephyr or any other supported model.
👉 See Scenario 1: Ollama + Zephyr (lightweight local) for model installation instructions.

Then open:

• Frontend: http://localhost:5173
• Backend (API docs): http://localhost:8000/docs
• if using ollama: install model ()

ASR_MODEL=whisper                  # whisper, whisper-api
OPENAI_API_KEY=sk-...              # for OpenAI-based features

VECTOR_BACKEND=qdrant             # qdrant
QDRANT_HOST=qdrant                # if starded within docker 
EMBEDDING_BACKEND=huggingface     # huggingface, openai

LLM_PROVIDER_TYPE=ollama          # openai, ollama, deepseek, local
LLM_RAG_MODE=rag_fallback         # rag_fallback, rag_strict

LOCAL_OLLM_MODEL=zephyr
LOCAL_OLLM_API_URL=http://localhost:11434/api/generate

DEBUG=true

• "rag_strict" – LLM can only answer from documents

• "rag_fallback" – If docs are not useful or missing, LLM falls back to its own knowledge

VITE_API_BASE_URL=http://localhost:8000/api

Supported formats:

mp3, mp4, mpeg, mpga, m4a, wav, webm

We recommend whisper with "medium" or higher model for best local accuracy.

• Zephyr can be replaced with any llm

• ~3.5 GB disk (Zephyr model)
• 7 GB RAM
• ~1 GB VRAM (optional)

docker exec -it ollama ollama pull zephyr

.env:

LLM_PROVIDER_TYPE=ollama
LOCAL_OLLM_MODEL=zephyr

1. Upload 3 audio files (MP3/WAV/WEBM/etc.)
2. Transcribe + ingest to RAG
3. Ask a contextual question (e.g. "What did Krzysio say?")
4. Ask general question (e.g. "What's the capital of Poland?")
5. Ask "What was my previous question?" (test memory)

• ~13–15 GB disk
• 8–16 GB RAM
• GPU highly recommended

.env:

LLM_PROVIDER_TYPE=deepseek

Model downloads automatically via transformers.

Same test flow as Scenario 1.

.env:

ASR_MODEL=whisper-api
OPENAI_API_KEY=sk-...
VECTOR_BACKEND=qdrant
QDRANT_HOST=qdrant
EMBEDDING_BACKEND=openai
LLM_PROVIDER_TYPE=openai
LLM_RAG_MODE=rag_fallback

No local models needed. Ideal for low-resource devices.

• Stored in chat_id (browser localStorage)
• Up to 10-turn rolling memory (user + assistant)
• Injected as summary/context for compatible models
• Shared to LLM as part of prompt (unless using OpenAI’s own chain)

Sample structure:

{
  "chat_id": "uuid...",
  "turns": [ { "role": "user", "content": "..." }, ... ],
  "summary": "optional summary",
  "context": "optional injected context"
}

• 🧠 Langchain for RAG + chaining
• 🗃 Qdrant as vector DB (via Docker)
• 🧱 Clean adapter-based architecture (LLM, ASR, Embeddings)
• 🎙 Whisper for audio transcription
• 🌐 FastAPI backend + Vite frontend

# Run full app
docker compose up --build

# Access frontend
http://localhost:5173

# Access backend (API docs)
http://localhost:8000/docs

# Pull Zephyr model (inside container)
docker exec -it ollama ollama pull zephyr

RAGdio/
├── RAGdio-backend/
│   ├── app/
│   │   ├── adapters/         # LLM, audio, embedding adapters
│   │   ├── services/         # RAG pipeline, chat memory
│   │   ├── routers/          # FastAPI routes
│   │   └── core/             # Config, logging
├── rag-audio-frontend/       # React + Tailwind UI
├── docker-compose.yml

MIT License. Built to tinker, break, and rebuild — have fun with audio-first RAG!