🦉 Caddy

Caddy is a Retrieval-Augmented Generation (RAG) platform combining:

• model: A FastAPI-based LLM service backed by OpenSearch
• scraper: A document scraper supporting structured ingest from websites, files, and S3

📦 Components

Folder	Description
model	LLM + OpenSearch API and ingestion
scraper	Scraping, summarization, doc parsing

🚀 Getting Started

With Docker:

cp docker-compose-config/.env.example docker-compose-config/.env

And populate the docker-compose-config/.env file with the necessary vars. Then

docker compose up

If you need to work on GovAI client at the same time, you can use the local-client override. When the override is active, docker compose will draw the client from ../gov-ai-client. To activate the override, use the -f option for docker compose.

 docker compose -f docker-compose.yaml -f docker-compose-config/local-client.yaml up

Important note: you must allocate at least 4GB for Docker or performance will be horrible and opensearch will keep dying.

Working with dependencies

If you change a dependency in pyproject.toml you will need to rebuild the app.

Do this using

docker compose build model --no-cache

To drop data volumes, which you may want to do if you change the database init script in ./model/scripts/postgres-init.sql or some OpenSearch settings, you'll need to drop the volumes. docker volume rm will not work for volumes created by docker compose, so use the following command:

# drop all volumes including OpenSearch, postgres etc
docker compose down --volumes

📂 Features

• Semantic and lexical document search
• OpenSearch-backed vector store
• PDF/DOCX/web content ingestion
• Authenticated scraping support
• S3, local directory support
• FastAPI + Streamlit frontend
• Data Ingestion & Collection Management

🧪 Testing

make run_tests

🧠 Architecture

• FastAPI (REST API)
• OpenSearch (vector DB)
• PostgreSQL (metadata)
• Streamlit (query UI)

🧾 License

MIT