could not parse response

[georgemac-labs]

I'm seeing the issue described here browser-use/browser-use#349. Seems to be occurring by Step 2-3 at the latest.

I've applied the fix here: browser-use/browser-use#417 but no improvement. Tried multiple models

Same setup works fine with ChatGPT as backend.

📍 Step 3
ERROR [agent] ❌ Result failed 1/3 times:
Could not parse response.
INFO [agent]
📍 Step 3
ERROR [agent] ❌ Result failed 2/3 times:
Could not parse response.
INFO [agent]
📍 Step 3
ERROR [agent] ❌ Result failed 3/3 times:
Could not parse response.
ERROR [agent] ❌ Stopping due to 3 consecutive failures

[galatian44to7]

@georgemac-labs Which model are you using?

[georgemac-labs]

@galatian44to7 I started with qwen2.5:32b-instruct-q4_K_M and also tried a couple of other models listed as having tools support, e.g. dorian2b/vera:latest, mistral-small:22b

I noticed that when I use local Ollama, the process runs within a separate Chromium instance, whereas with the OpenAI API it re-uses the Chrome Beta browser window that the extension is running in. Maybe that's as intended, but I mention it in case it's relevant.

[galatian44to7]

@georgemac-labs The issue has been fixed in the latest push. In mainOllama.py, browser was set incorrect when the agent was created. Please inform if this doesn't fix it on your end

[georgemac-labs]

@galatian44to7 yep, LGTM.

The other issue has also been answered in the browser-use discussion (browser-use/browser-use#349 (comment)) – allegedly local models simply do not adhere well to the schema. Will look into this when I have time, but doesn't seem to be in scope for this project, so this issue would be resolved for me.

[georgemac-labs]

... actually, one further question. I'm no longer getting Chromium, but with both cloud or local models, when I run my command from the extension, it immediately opens a new tab and then has to re-open the website I was already on. Is that intentional?

[georgemac-labs]

@galatian44to7 I have now resolved the original issue. IMO the explanation here may be wrong, and is certainly not the issue in this case.

The problem is simply that requirements.txt specifies browser-use==0.1.21, and that seems to have some inconsistent behaviour, such that the model is sometimes returning the response in the content field and sometimes in tool_calls. Thus it cannot be parsed.

After pip install --upgrade browser-use, I have v0.1.30 and no more parsing errors. So I would propose updating requirements.txt

[galatian44to7]

@georgemac-labs great call and will do. Suppose that should be a regular occurrence as browser-use constantly updates. Thanks for pointing this out!