Archon

A minimalistic AI-powered search assistant that combines SearXNG (local web search) with Ollama (local LLM) to answer your questions using real web results. It is privacy-focused with chat history saved locally in the browser. Searxng can be configured to use VPN and aggregate search results in various ways. And since everything is running locally on your privately hosted instance, there is no tracking or training of your data of any kind.

Prerequisites

• Node.js ≥ 18
• SearXNG running on http://localhost:8080
  • If searXNG cannot be reached, try add the following to nftables in archlinux:

    chain forward {
        type filter hook forward priority filter; policy drop;
        ip saddr 172.18.0.0/16 accept
        ip daddr 172.18.0.0/16 ct state established,related accept
    }
    

    If you are using a VPN (e.g. MullVad VPN or using MullVad Browser), you need to set the priority higher such as -1

    chain forward {
       type filter hook forward priority -1; policy drop;
       ip saddr 172.18.0.0/16 accept
       ip daddr 172.18.0.0/16 ct state established,related accept
     }
    

    Sometimes your firewall is blocking docker network access, so check ufw if applicable.
• Ollama running on http://localhost:11434 with model mistral:7b pulled

Setup

cd into searxng folder and then setup the .env file based on .env.example and then run docker compose up -d.

download ollama and pull your favorite model and ollama serve if ollama isn't already running as a service in your system, in which case you can check sudo systemctl status ollama.

• For a 16GB laptop with something like RTX 2060 for GPU, you can totally run mistral:7b, its very lightweight and efficient, anything above 7b will struggle on a 16GB laptop; if you have 96GB RAM with RTX 4090 GPU, you can run gpt-oss:20b, qwen3.5:35b or deepseek-r1:32b; consider running quantized versions in resource constrained environments. Also, llama3.2:1b can also run on CPU instead of GPU if you are running on a server without a GPU, it can be done.

npm install
npm run dev

Then open http://localhost:5173.

How it works

1. You type a question in the input box and hit Search.
2. The app queries your local SearXNG instance with search queries formulated by a lightweight AI model and then retrieves and merges the top 8 results per each queries (max 3 queries).
3. The results are bundled into a prompt and sent to Ollama (mistral:7b) via streaming.
4. The answer streams in token-by-token, with sources listed below.

Architecture

browser  →  Vite dev-server proxy  →  SearXNG  :8080
                                  →  Ollama   :11434

Vite's proxy avoids CORS issues — all requests go through /searxng/* and /ollama/*.

Build for production

npm run build   # outputs to dist/
npm run preview

For production you'll need a reverse proxy (nginx / Caddy) to forward /searxng and /ollama paths.

Screenshots

Main search input and chat interface while a reply streams in with model and timer for generating answers.

Answer with numbered web citations shown under the streamed response with link to source and which search engine they came from.