Perplexity without the subscription or the telemetry

What it does

Farfalle is an open-source Perplexity clone that searches the web and synthesizes answers with an LLM. It pairs a Next.js frontend with a FastAPI backend, pulling search results from sources like SearXNG, Tavily, Bing, or Serper, then summarizing them via your chosen model. You can run the entire stack locally, in the cloud, or split between the two.

The interesting bit

The “Expert Search” feature treats search as an agentic task: it plans and executes multiple queries to refine results instead of doing a single-shot retrieval. The project also lets you drop commercial search APIs entirely by self-hosting SearXNG, making a fully offline pipeline possible.

Key highlights

• Local inference via Ollama (llama3, mistral, gemma, phi3) or cloud APIs (OpenAI GPT-4o, Groq/Llama3), plus custom models through LiteLLM.
• Choice of search providers, including self-hosted SearXNG to avoid external dependencies.
• Agentic “Expert Search” for multi-step research.
• Pre-built Docker image and split backend/frontend deployment.
• Can be registered as your browser’s default search engine.

Verdict

Worth a look if you want to own your search pipeline and toggle between offline and cloud models. Not for you if you need a managed service with chat-over-files, which is still on the roadmap.