LLM Application Models from Ollama not auto-loaded

Help / Discuss
1

Hi everyone,

I just started experimenting with the LLM Application (v0.7.2). I’m not sure if what I’m experiencing is a misconfiguration on my end or a potential bug.

My Setup:
XWiki version: 17.7.0 (Debian)
LLM Application: 0.7.2
Only one provider configured: Ollama
Ollama server: Running on a separate machine in my local network

I added the provider via the AI LLM Application config in the administrator menu – just with a name and the correct URL, and streaming activated.

A query (curl http://*:11434/v1/models) to my Ollama looks like this:

{“object”:“list”,“data”:[{“id”:“llama3.1:latest”,“object”:“model”,“created”:1756733917,“owned_by”:“library”},{“id”:“qwen2:latest”,“object”:“model”,“created”:1756733814,“owned_by”:“library”},{“id”:“llama3:latest”,“object”:“model”,“created”:1756733053,“owned_by”:“library”},{“id”:“zephyr:latest”,“object”:“model”,“created”:1756732566,“owned_by”:“library”},{“id”:“gemma:2b”,“object”:“model”,“created”:1756732469,“owned_by”:“library”}]}

However, when I check the XWiki internal API:
GET /xwiki/rest/wikis/xwiki/aiLLM/v1/models
…I get an empty result:

{“data”:[],“object”:“list”,“first_id”:null,“last_id”:null,“has_more”:false}

Even after restarting XWiki, nothing changes.

What does work:
If I manually add a model in the configuration UI (e.g., “llama3.1:latest”), then it shows up in the list, and everything works from there — completions, chat, etc.

What I’ve Checked:

  • /models and /chat/completions endpoints are reachable from the XWiki server

  • CORS is working correctly (tested via curl and browser)

  • No relevant errors in the logs

  • Restarted XWiki

Is this a known bug, or am I missing a step?

I’d really appreciate any pointers. Thanks in advance.

2

It’s a known limitation. We should make it more comfortable to create models by allowing to select from the models that the server exposes, but at the moment, that’s how it works. The rationale for not automatically listing all models from Ollama is that a wiki admin might not want to expose all models, and at the same time, the wiki admin might want to expose the same model with different context configurations (using the index for the LLM application extension) under different names.

3

Hi Michael,
Thanks for the quick reply! What you’re saying definitely makes sense. I must have missed that part because in the demo video, it looked like everything worked automatically. :stuck_out_tongue: