How to Set Up a Self-Hosted Perplexity Alternative with Perplexica

Yes. Perplexica is open-source under the MIT license and costs nothing beyond the hardware you run it on. No API keys are required when using Ollama for local inference.

If you connect a cloud LLM provider (OpenAI, Groq, or Anthropic) for faster or more capable responses, you pay those providers at their standard API rates. OpenAI GPT-4o costs $2.50 per million input tokens as of March 2026. For 50 to 100 search queries per day, that is under $1 per month. Groq offers a free tier with rate limits that covers moderate personal use.

For fully free operation: use Ollama with a local model. The electricity cost of running a 7B model on a desktop CPU is approximately $0.01 to $0.03 per hour of continuous use.

Yes. Perplexica supports three cloud LLM providers as alternatives to Ollama: OpenAI (GPT-4o, GPT-4o mini), Groq (Llama 3.3 70B, Mixtral 8x7B), and Anthropic (Claude 3.5 Sonnet, Claude 3 Haiku).

To enable a cloud provider, add the API key to the `[API_KEYS]` section of `config.toml`:

[API_KEYS]
OPENAI = "sk-proj-..."
GROQ = "gsk_..."
ANTHROPIC = "sk-ant-..."

Restart the backend after editing config.toml:

docker compose restart perplexica-backend

Then open Perplexica settings in the browser and change the LLM Provider from Ollama to your chosen provider. Groq is the fastest option for free use. Llama 3.3 70B on Groq responds at approximately 800 tokens per second, faster than any local GPU setup.

The core difference is architecture. Perplexica is built around search: every query triggers a SearXNG search, the top results are passed to the LLM as context, and the response always includes numbered citations. SearXNG queries multiple search engines simultaneously and deduplicates the results before passing them to Ollama.

Open-WebUI's web search uses DuckDuckGo or a configured API. The LLM decides whether to trigger a search based on the query content. Citations are less structured, and the feature is an add-on rather than the primary interface.

Use Perplexica when cited, sourced answers are the primary goal. Use Open-WebUI when you want a general-purpose chat interface that occasionally needs current information.

SearXNG is a meta-search engine that queries multiple backends simultaneously and aggregates the results. With default settings it queries Google, Bing, DuckDuckGo, and Brave Search in parallel.

Individual engines may fail or return blocked responses depending on your server IP. Google in particular rate-limits requests from cloud VPS IP addresses. SearXNG handles individual failures gracefully: if Google is blocked, results from Bing and DuckDuckGo still arrive.

You can configure which engines to include or exclude in `searxng/settings.yml` under the `engines` key. The SearXNG documentation at docs.searxng.org lists over 100 supported search backends including Qwant, Startpage, and Mojeek.

Yes. Perplexica binds to `0.0.0.0` by default, so any device on the same network can reach it at `http://[your-machine-ip]:3000`. Find your machine IP with `ip addr show` on Linux or `ipconfig getifaddr en0` on macOS.

For access over the internet, set up an Nginx reverse proxy with SSL as covered in the "Access Perplexica from Other Devices" section of this guide. Certbot (Let's Encrypt) provides free SSL certificates. Once configured, Perplexica is reachable at `https://search.yourdomain.com`.

For security on internet-facing instances, add HTTP basic authentication to the Nginx config to restrict access to your instance.

For most hardware (8 to 16 GB RAM), Mistral 7B or Llama 3.2 7B are the best starting points. Both run in Q4_K_M quantisation at 4.1 GB on disk and produce coherent cited answers in 8 to 15 seconds on CPU-only machines.

With 16 GB RAM available, Mistral-Nemo 12B produces noticeably better answer quality for research and academic queries, at 12 to 25 seconds per query on CPU.

On machines with an NVIDIA GPU with 12 GB VRAM or more (RTX 3060 12 GB or better), a 7B to 13B model responds in 2 to 4 seconds per query. Pull the model before starting the Perplexica stack: `ollama pull mistral:7b` or `ollama pull mistral-nemo:12b`.

Pull the latest source from GitHub, rebuild the Docker images, and restart the stack:

cd Perplexica
git pull origin main
docker compose down
docker compose up -d --build

The `--build` flag forces Docker to rebuild from the updated source. Your `config.toml` and `searxng/settings.yml` files are not modified by the update. Docker named volumes (chat history, uploaded files) also persist across rebuilds.

Check the Perplexica GitHub releases page before updating. Configuration file field names occasionally change between major releases, which requires manually updating `config.toml` after the git pull.

No. Perplexica depends on SearXNG making outbound HTTP requests to search engines (Google, Bing, DuckDuckGo). Without internet access, SearXNG returns no results and the LLM has no search context.

The Ollama component works offline, but without search results Perplexica falls back to the model's training knowledge and produces a plain response without citations. This is functionally equivalent to using Ollama directly.

For a fully offline AI assistant, use Open-WebUI connected to Ollama without the web search plugin. Perplexica is designed for the use case where privacy means "no data leaves your server" rather than "no internet connection at all".