Most "AI search" tools quietly route your queries through a vendor's servers and a vendor's index. This project inverts that: the entire retrieve-then-reason loop can run on a machine you own, with a self-hosted SearxNG doing the crawling and an Ollama model doing the reasoning. The insight is that a good answer engine is mostly plumbing — query rephrasing, source ranking, and citation — and none of that strictly requires a hosted black box (the project was recently renamed from Perplexica to Vane).
What Sets It Apart
- Bring-your-own-everything: a self-hosted SearxNG meta-search feeds any LLM, local (Ollama) or cloud (OpenAI, Claude, Groq, Gemini), so you control both the index and the model rather than a single vendor.
- Three explicit tiers — Speed, Balanced, Quality — instead of one opaque pipeline, letting you trade latency for how many sources get fetched and reranked per query.
- Switchable sources — web, academic papers, and discussions — plus image/video search and file upload, so one deployment covers more than plain text Q&A.
- Citations are first-class, which matters more here than in hosted tools: you can drop in a weaker local model and still verify every claim against its sources.
Who It's For
Great fit if you want a Perplexity-style experience without leaking queries to a third party, already run Ollama, or need a search layer you can embed and customize. Look elsewhere if you want a zero-setup hosted product — you will be running Docker, SearxNG, and a model yourself — or if you need the breadth of a frontier hosted index, since answer quality is bounded by your configured SearxNG engines and the model you pick.
