Perplexica: self-hosted open-source alternative to Perplexity AI search

My notes

Summary

Perplexica is a free, open-source, self-hostable alternative to Perplexity AI search. It pairs with SearXNG (a meta-search engine) and Ollama (local LLM inference) to give you a fully private AI-powered search tool that runs on hardware as modest as a Raspberry Pi or an old Mac Mini with 8 GB RAM.

Key Insight

• The stack is three components: Perplexica (the AI search UI on port 3000), SearXNG (the search aggregator on port 8080), and Ollama (local LLM runtime on port 11434). All run in Docker containers.
• With a small model like Qwen 3.5 2B, 8 GB RAM is sufficient. Larger models (4B, 6B, 8B) need more RAM but produce deeper answers. Alternatively, you can point Perplexica at Claude or ChatGPT APIs instead of running local inference.
• SearXNG is highly configurable: you can enable/disable individual search engines per category (general, images, video, news, social media, IT/dev). Notably, enabling Reddit, npm, Hugging Face, and Ollama Hub searches makes this useful for dev work.
• Perplexica offers three search modes (speed, balanced, quality) and supports social media search if SearXNG has those engines enabled.
• The system prompt in Perplexica can be customized (e.g., “be to the point, factual, no fluff”) to shape response style.
• Running SearXNG inside a VPN container is mentioned as an option for additional privacy.