Confirm dialog for web search (for privacy reasons)

[slu-010101]

Hi,

I don't feel comfortable with the assistant being able to simply send search queries to the web.

Would it be possible to add a dialog box with an optional confirmation prompt?
For example: This search query "query" will be sent to the search engine "DuckDuckGo". [ok] [cancel]

Thanks.

Samuel

[KeithCu]

Thank you for your report and perspective!

This is not a true Agent like OpenClaw which is running 24x7 and will find a random credit card online and order food to your house, and figure out how to talk on the phone and start calling you! ;-)

This system will not send search queries to the web unless:

• you tell them to
• or they don't feel confident in their knowledge of a question you ask

It uses DDG which is considered a private search engine.

I was more focusing on figuring out a way to talk to Chrome or other browsers to use that as the browsing engine, with JS support for better results.

I love this search feature, but I could see adding a mode to prompt the user to ask on search, so I'll work on the checkbox.

The multi-threaded UI code is complicated already and this will add more to it, but I refactored the code so it should be easier to add features like this: https://github.com/KeithCu/writeragent/blob/master/docs/formal_verification.md#why-we-refactored-orchestration-into-pure-state-machines.

[KeithCu]

I'm going to add a feature for this. I'll change the buttons to "Accept" or "Reject" but it's tricky to know what to do if you click Reject, so I'll just stop the entire process for now. I could try to continue but if the agent thinks it needs information and you reject it, what should it do? I could have it try to just guess, but it would likely be wrong, the whole reason it made the query is because you told it or it was unsure, and I'm not sure making things up is good either.

Is just stopping the whole tool-call loop okay given no answer is better than a bad answer?

[slu-010101]

Is just stopping the whole tool-call loop okay given no answer is better than a bad answer?

Very good question, but yes I think stopping is better than a bad answer.

Or you have three options:
[ok] [reject] [AI only answer]

If you think this make no sense at all, please close this idea and reject it.
No issue for me :)

[KeithCu]

Also, another question is, do you want to see this prompt for web research? When you are in that mode, you explicitly are talking to a web research sub agent asking it to look up things for you, so it seems like verifying the request would be not necessary. They are separate codepaths. For now I'll make them both prompt you since I've already implemented it but I wanted to run it by you, whether 1 checkbox is enough or whether it should be a 3-state for the separate codepaths.

[slu-010101]

Also, another question is, do you want to see this prompt for web research?

Yes that was my idea, so I can check the content to accept, reject or modify it.

[KeithCu]

Okay I've got it implemented. It's a fairly large diff which I don't like for the general usefulness of the feature: 7df4f43

IMO, if you don't want an agent to search for the best pizzas in Madison Heights, MI, don't ask it the question.

Anyway I check it into master for now. It's more complicated because of the async nature of my code. Some of the human in the loop features I wanted to do one day anyway, but it was around document changes.

I could also have a checkbox to disable the tool-call, but the whole point is that because it uses DDG and all requests run from your computer it's already private.

[KeithCu]

Okay, try the new build of 0.71 I just uploaded a new version but didn't create a new release ;-)

It has the approval / change feature in so you can try it out and make sure the general language features are all working. Let me know if there are any bugs. Thank you for your helpful and thoughtful input.

[slu-010101]

Hi Keith,

I tested this new version (thank you very much for your work!) and like the request dialog before sending.

Two things I found:

1. Reject the request log a RuntimeError:

Exception ignored in: <generator object MultiStepAgent._run_stream at 0x7fc149ea5a80>
Traceback (most recent call last):
  File "/home/slu/.config/libreoffice/4/user/uno_packages/cache/uno_packages/lu8482ddsw.tmp_/WriterAgent.oxt/plugin/modules/chatbot/web_research.py", line 170, in execute
    return format_error_payload(
RuntimeError: generator ignored GeneratorExit

2. Modify request is a endless loop, the changed text is not shown in the chat window (saw it in on the terminal soffice --writer) and the tool ask the same original text again and again to approv.

[KeithCu]

Thank you!! I fixed both, tested, and am putting a new release up there for you to try.

I didn't update the version number since only you are using this feature.

Please let me know if you find any other issues and I'll work on them.

[slu-010101]

Working now, thank you very much!

[slu-010101]

Something I saw in the chat window:

[KeithCu]

Ah, newlines getting in there. Good catch! I uploaded a new release with some other cleanup and fixes.

[KeithCu]

This should be fixed. I'll close it for now. Let me know if any other bugs you find.