deijany/searxng-ai-summary repository overview

⁠searxng-ai-summary

AI summaries on top of your SearXNG results, from any OpenAI-compatible endpoint: LiteLLM, OpenAI, Anthropic, DeepSeek, Kimi, llama.cpp, LM Studio, vLLM, Ollama.

• Summary box above the results, streamed as it's written, with clickable [n] citations
• Dropdown to switch models (choice remembered per browser)
• No SearXNG changes: runs as a small proxy in front of it, survives SearXNG updates
• Uses the results already on the page (title, URL, snippet), so no extra searches
• Summaries cached per browser tab, so going back doesn't call the model again

⁠How it works

browser --> searxng-ai-summary --> SearXNG
                   |
                   └-------------> your LLM (summary)

searxng-ai-summary proxies every request to SearXNG and adds one script and one stylesheet to HTML pages. The script reads the top results from the page and asks /_ai/summarize for a summary.

⁠Quick start

1. Copy docker-compose.example.yml to docker-compose.yml and edit it.
2. Remove ports: from your SearXNG service; searxng-ai-summary takes over that port.
3. docker compose up -d

Check: curl http://localhost:8888/_ai/config lists your models.

⁠Configuration

You need four settings to get started:

• LLM_URL: the address of your model server, ending in /v1. For LiteLLM this is usually http://litellm:4000/v1.
• LLM_KEY: the API key for that server. Leave it out if your server doesn't need one.
• MODELS: the models to show in the dropdown, separated by commas. The first one is used by default. You can give a model a friendlier name with Name=model, for example Local=qwen3-8b.
• UPSTREAM: where to find SearXNG. Default http://searxng:8080. Change it only if your SearXNG container has a different name, e.g. http://my-searxng:8080. Keep port 8080 even if you mapped SearXNG to another port on your host.

Everything else is optional:

• AUTO_SUMMARIZE: set to false to only summarize when you click the button. Default true.
• MAX_RESULTS: how many results are sent to the model. Default 8.
• MAX_TOKENS: maximum length of the summary. Default 1200.
• RATE_LIMIT_PER_MIN: summaries allowed per user per minute. Default 20.
• LLM_TIMEOUT: seconds to wait for the model before giving up. Default 120.
• SYSTEM_PROMPT: replace the built-in instructions sent to the model.

⁠Notes

• Exposing it publicly. Anyone who can open the instance can spend your LLM credits. Keep it behind a VPN or auth, or point it only at a local model.
• SearXNG limiter. If enabled, requests now come from the searxng-ai-summary container. The real client IP is sent in X-Forwarded-For / X-Real-IP; add the Docker network to the limiter's trusted proxies.
• Unstyled page after upgrading. Your browser may still have a broken stylesheet cached from an older version: do a hard reload or clear the site's cache.

⁠Build locally

docker build -t searxng-ai-summary .