ItzCrazyKns/Vane

Vane is a self-hosted AI answering engine that searches the web and returns cited answers using local or cloud language models.

What it is

Vane is a web application that combines a metasearch backend with a large language model to answer questions and cite its sources. It is aimed at users who want a Perplexity-style answering experience but hosted on their own hardware, so that queries and search history stay private. It runs the whole stack locally, pairing a bundled SearxNG search engine with a model provider of the user's choice.

Key features

Works with local models through Ollama and with cloud providers including OpenAI, Anthropic Claude, Google Gemini, and Groq, and lets you mix models.
Three search modes named Speed, Balanced, and Quality to trade latency against answer depth.
Selectable source types (web, discussions, academic papers) and the option to restrict a search to specific domains.
Web search runs through SearxNG so queries are proxied across multiple engines; image and video search are also supported.
File uploads (PDFs, text files, images, and other documents) that can then be queried.
Inline widgets for things like weather, calculations, and stock prices, plus a Discover feed, typed suggestions, and locally stored search history.

Tech stack

TypeScript, built on Next.js ^16.0.7 with React ^18 and Tailwind CSS ^3.3.0.
SQLite storage via Drizzle ORM ^0.45.2 and better-sqlite3 ^11.9.1, with migrations managed by drizzle-kit ^0.18.1.
Model provider clients: openai ^6.9.0, @google/genai ^1.34.0, ollama ^0.6.3, and @huggingface/transformers ^3.8.1.
Document ingestion through pdf-parse ^2.4.5, mammoth ^1.9.1, officeparser ^6.0.7, @mozilla/readability, and jsdom.
Supporting libraries: zod ^4.1.12 for validation, js-tiktoken for token counting, mathjs and yahoo-finance2 for widgets, playwright ^1.59.1, and axios.
SearxNG as the search backend, deployed via Docker and docker-compose (a bundled image and a slim image that reuses an external SearxNG instance).

When to reach for it

You want a private, self-hosted alternative to hosted AI answer engines, with searches kept on your own machine.
You already run, or want to run, local models through Ollama and prefer to keep inference off third-party APIs.
You want answers that link back to cited web sources rather than ungrounded model output.
You want to query your own uploaded documents alongside web results.
You want to wire an AI search box into your browser or call it from your own application through its search API.

When not to reach for it

You do not want to operate infrastructure. Vane expects you to run Docker (or Node plus a SearxNG instance) and configure providers and API keys yourself.
You need multi-user access control today. Authentication is listed as an upcoming feature, not a shipped one.
You want a turnkey hosted product with no setup, in which case a managed service fits better.
You need Tavily or Exa as the search source now; the README lists these as coming soon, with SearxNG as the current backend.

Maturity signal

Vane is actively maintained. It carries roughly 35,800 stars and about 3,900 forks, was created in April 2024, and had commits pushed as recently as April 2026, with a current package version of 1.12.2. It is not archived and is MIT licensed. The 337 open issues are consistent with an active project that has a sizable user base rather than a sign of abandonment. The dependency set tracks recent releases (for example Next.js 16 and Zod 4), which points to ongoing upkeep.

Alternatives

Perplexity AI — use instead when you want a polished hosted product and do not care about self-hosting or keeping data local.
SearXNG — use instead when you only need private metasearch and do not want an LLM answering layer on top.
Morphic — use instead if you want another open-source AI answer engine and prefer its stack or hosting model.
Khoj — use instead when the priority is searching and chatting over your own documents and notes more than answering general web queries.

Notes

The repository's topics, one-click deploy templates, and the sponsor link all reference "perplexica" (for example the Sealos and Warp URLs use a perplexica template name), indicating Vane is a rebrand or successor of the author's earlier Perplexica project. The self-hosted setup has a hard dependency detail that is easy to miss: an external SearxNG instance must have JSON output and the Wolfram Alpha engine enabled, otherwise search and some widgets will not work.

Wiki: ItzCrazyKns/Vane