Perplexica answers questions by searching the web, reading the results, and synthesising an answer with citations. It is an open-source take on the AI search category, and it runs entirely on hardware you control.

Why search plus a model

A language model knows what it was trained on and has no idea what happened last week. Search knows what happened last week and hands you ten links to read.

Combining them addresses both weaknesses: the search provides current material, the model reads it and answers the actual question, and the citations let you check whether it got it right.

That last part is what separates this from asking a model directly. An answer with sources you can open is verifiable; an answer without them is a claim.

What it needs

Two dependencies:

• SearXNG provides the search, with no API key and no per-query cost
• A model, either local through Ollama or a hosted API

Running both locally means nothing about your queries leaves your network, which is a meaningful difference from a hosted AI search product where every question is logged against you.

Focus modes

Different modes tune the search and prompting for different question types: general web search, academic search, YouTube, Reddit, and mathematical computation. The specialised modes produce noticeably better results than a general search for their domains.

Realistic expectations

Answer quality depends on your model. A small local model reading five search results will be less coherent than a large hosted one doing the same thing, and it will still be useful.

It also inherits search’s weaknesses. If the top results are wrong or SEO spam, the synthesis will faithfully reproduce that. Citations are what let you notice.

Running it

Docker Compose, with SearXNG alongside. Enable JSON output in your SearXNG configuration, since Perplexica needs it and it is off by default.

License

Perplexica is released under the MIT License.

Links

• Repository