Farfalle
An open-source AI search engine you can self-host with Docker, using Ollama for local models or OpenAI, Groq and LiteLLM for other backends.
Farfalle is a self-hosted AI search engine for people who want a Perplexity-style search app with a choice of local or cloud models. It combines web search with model-generated answers and includes an agent that plans and carries out searches. The code is open source under Apache 2.0.
Local model support comes through Ollama, with llama3, gemma, mistral and phi3 named as supported choices. You can keep answer generation on your own hardware, or use cloud models through OpenAI and Groq. LiteLLM provides another route for connecting custom LLMs, so the app doesn't tie you to a single model provider.
Search is a separate choice. Farfalle works with SearXNG, Tavily, Serper and Bing, letting you pair your preferred search backend with your model. Running a local LLM keeps model inference local, but web search still needs an internet connection. Choosing a cloud model also sends the answer-generation work to that provider.
The app runs with Docker and presents a browser interface. Its Next.js frontend and FastAPI backend can run together on your own machine or be deployed separately. Ollama use doesn't require model-provider API keys; cloud model services and some search providers use their own keys.
For people comparing AI search apps, its main distinction is the independent choice of search provider and LLM backend. Farfalle also supports use as your browser's default search engine, so queries from the address bar can open in your self-hosted instance.
