mirror of
https://github.com/open-webui/open-webui.git
synced 2026-05-07 03:18:23 -05:00
Web search prompt doesn't used, and noisy RAG #1246
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: github-starred/open-webui#1246
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @Fusseldieb on GitHub (Jun 13, 2024).
Bug Report
Description
Bug Summary:
When using the web search capability, it seemingly skips the web search prompt and simply uses the original question prompt.
Eg. if I ask "Make a Google search to tell me who's Fusseldieb", it just skips to
"Searching 'Make a Google search to tell me who's Fusseldieb'...When looking at the logs, it really does search for the whole original prompt. Seems like the prompt to transform the query into a Google search doesn't get executed, or ignored.
Also, when searching on the web, when it reads individual searches, it doesn't trim the output, which makes it very noisy, and uses A LOT of useless tokens.
See below.
Steps to Reproduce:
Expected Behavior:
Actual Behavior:
Environment
Open WebUI Version: 0.3.4
Operating System: Debian GNU/Linux 11 (bullseye)
Reproduction Details
Confirmation:
Logs and Screenshots
Docker Container Logs:
Screenshots (if applicable):
Installation Method
Docker
Additional Information
Suggestion: I think this would ideally be a pipeline, not a built-in function.
@justinh-rahb commented on GitHub (Jun 13, 2024):
@AeneasZhu that is unrelated to this issue.
@tjbck commented on GitHub (Jun 13, 2024):
You can change the prompt length threshold from admin settings!
@Fusseldieb commented on GitHub (Jun 13, 2024):
@tjbck Thanks for answering, however that doesn't quite address the issues.
The bug in which it skips through the query prompt is, well, a bug.
And while the prompt length threshold can indeed be adjusted, the issue is that it doesn't "clean up" endless ("\n", newlines) which may come from web searches. This is less important, but I still think it's worth mentioning, as both are related to the web search feature.
I think this issue should only be closed if at least the bug is solved.
@EFLKumo commented on GitHub (Jun 16, 2024):
The length threshold means that when your prompt word exceeds the threshold before a web search keyword is generated with the LLM instead of using your request directly, so if you want to generate web search keywords by LLM every time, you should set the threshold to
0.