SearXNG Web Search Not Working in Open WebUI v0.10.x with Ollama — 3 Fixes (Solved)

[julibio]

Environment

• Open WebUI: v0.10.1 (Docker, network_mode: host)
• SearXNG: latest (Docker, bridge, port 8082:8080)
• Ollama: running on WSL2 Ubuntu 26.04 (CUDA GPU)
• Models tested: qwen3.5:9b, gpt-oss:20b
• Host: Windows 11 Pro + WSL2

Symptoms

• SearXNG works perfectly when tested with curl from inside the Open WebUI container
• Web Search toggle (globe icon) is enabled in chat
• Admin Panel settings are correct (engine: searxng, URL configured)
• But the model responds as if it has no web access ("I don't have access to real-time data...")
• No search-related entries appear in Open WebUI logs at all

Root Cause — Three Issues

1. Missing &format=json in SearXNG Query URL

Without &format=json, SearXNG returns HTML instead of JSON. Open WebUI silently fails to parse it.

Verify:

# Returns HTML (broken):
curl -s "http://localhost:8082/search?q=test" | head -3

# Returns JSON (working):
curl -s "http://localhost:8082/search?q=test&format=json" | head -3

Fix: In Admin Panel > Settings > Web Search, set the Query URL to:

http://localhost:8082/search?q=<query>&format=json

Note: Also ensure SearXNG's /etc/searxng/settings.yml has json in the formats list:

search:
  formats:
    - html
    - json

2. Concurrent Requests set to 0 in Admin Panel

Environment variables like RAG_WEB_SEARCH_CONCURRENT_REQUESTS=10 in docker-compose.yml are overridden by database-stored settings once you save anything in the Admin Panel. The Admin Panel value may default to 0.

Fix: Admin Panel > Settings > Web Search > Concurrent Requests: set to 10 (or any positive number).

3. num_ctx too small for Agentic Function Calling (the critical one!)

This is the main issue most people will miss.

Open WebUI v0.10.x changed web search to "agentic" mode — the model itself decides when to search by calling search_web and fetch_url tools via native function calling. The tool schemas are injected into the model's context.

With Ollama's default num_ctx of 2048 tokens, there isn't enough room for the tool schemas + conversation + the model's response. The model simply can't generate the function call, so it responds without searching.

Fix: In Admin Panel > Settings > Models > select your model > Advanced Params > num_ctx (Ollama): set to 16384 (minimum 8192).

This must be done per model for every model you want to use with web search.

Verification

After applying all three fixes:

1. Open a new chat (not an existing conversation)
2. Enable the Web Search toggle (globe icon near the input)
3. Ask something that requires current data (e.g., "What's the weather in São Paulo?")
4. The model should now call the search tool and respond with sourced information

Key Takeaways

• In Open WebUI v0.10.x, web search requires models with native function calling support — it's no longer automatic
• Legacy function calling mode does NOT support web search in v0.10.x
• The num_ctx parameter is critical — tool schemas consume significant context tokens
• Admin Panel (database) settings always override environment variables from docker-compose
• SearXNG needs both &format=json in the URL and json in settings.yml formats
• Qwen 3.5 models work well with function calling when given sufficient context

Credits

This issue was diagnosed and resolved by Claude (Anthropic) on behalf of @Julibio, through systematic research of the Open WebUI documentation, GitHub issues, and hands-on debugging.

Related Issues

• #9995 — No search query generated
• #25038 — Web Search retrieves results but fails during injection
• #25585 — Web search results not passed to model after v0.9.6
• #17820 — Web Search in Native function calling mode

[floatlesss]

Thank you so much! This fixed my exact problem and now my model queries searxng perfectly.
You were right in calling "num_ctx too small for Agentic Function Calling" critical. :)

[antigenx]

I was losing my mind trying to get this to work. Thank you. Ironically, AI was zero help here.

[Carlos2777777]

Thanks a ton, spent a few hours on this. Setting the num_ctx finally got it to work.