I have searched the existing issues and discussions.
Problem Description
Do a Multi-Chat.
Click the merge icon that appears lower-right after all responses complete.
Merge response uses the 'Task Model'. Bad, very bad.
Merge responses instead, very different from other tasks, should be configured to use the best available chain-of-reasoning (aka. 'thinking') model (ideally OpRt_.anthropic/claude-3.7-sonnet:thinking , or at least OpRt_.deepseek/deepseek-r1-distill-llama-70b).
Web search query generation, as with all other tasks, by contrast suffers a sort of loss of creativity, a narrowing of the possible queries generated, if a reasoning/thinking model is used, resulting in both much more queries generated and much less relevance. Ideally, such tasks are best done with OpRt_.meta-llama/llama-3.1-405b-instruct , or at least Llama-3.1-8b , or some fine-tuned Llama-3.1-8b .
Directly results in complete reports which can be used as comprehensive end-user user manuals, or certification of whether code is ready for testing. Whereas all of the contributing models miss multiple crucial features outside their knowledge, logic, or search capabilities.
Merging with a non-reasoning model such as OpRt_.meta-llama/llama-3.1-405b-instruct , however, results in a report too similar to the contributions of other non-reasoning Llama models, still missing crucial features, and poorly formatted.
Separate setting for Merge Responses using Local Models and External Models.
Merge Response treated more as an ordinary response in the chat session.
Alternatives Considered
No response
Additional Context
No response
Originally created by @mirage335 on GitHub (Mar 25, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/12022
### Check Existing Issues
- [x] I have searched the existing issues and discussions.
### Problem Description
Do a Multi-Chat.
Click the merge icon that appears lower-right after all responses complete.
Merge response uses the 'Task Model'. Bad, very bad.
Merge responses instead, **very different from other tasks**, should be configured to use the best available chain-of-reasoning (aka. 'thinking') model (ideally OpRt_.anthropic/claude-3.7-sonnet:thinking , or at least OpRt_.deepseek/deepseek-r1-distill-llama-70b).
Web search query generation, as with all other tasks, by contrast suffers a sort of loss of creativity, a narrowing of the possible queries generated, if a reasoning/thinking model is used, resulting in both much more queries generated and much less relevance. Ideally, such tasks are best done with OpRt_.meta-llama/llama-3.1-405b-instruct , or at least Llama-3.1-8b , or some fine-tuned Llama-3.1-8b .
The differences are astonishing.
Merging...
Many-Chat [ OpRt_.perplexity/sonar-deep-research , ChatGPT_.gpt-4.5-preview , OpRt_.anthropic/claude-3.7-sonnet:thinking , ChatGPT_.o3-mini , OpRt_.deepseek/deepseek-r1-distill-llama-70b , OpRt_.meta-llama/llama-3.1-405b-instruct , ChatGPT_.o1 ] -> Merge OpRt_.anthropic/claude-3.7-sonnet:thinking
With prompts such as these:
https://github.com/mirage335-special/kit-researchEngine/blob/a66ed0c8eedef3aa813c52e216a917de69006862/_ref/DUBIOUS-speculativePrompts/special-damageControl-ExplainCommand.txt
https://github.com/mirage335-special/kit-researchEngine/blob/a66ed0c8eedef3aa813c52e216a917de69006862/_ref/DUBIOUS-speculativePrompts/review-code-condensed.txt
Directly results in complete reports which can be used as comprehensive end-user user manuals, or certification of whether code is ready for testing. Whereas all of the contributing models miss multiple crucial features outside their knowledge, logic, or search capabilities.
Merging with a non-reasoning model such as OpRt_.meta-llama/llama-3.1-405b-instruct , however, results in a report too similar to the contributions of other non-reasoning Llama models, still missing crucial features, and poorly formatted.
Web search...
Will correctly search down the syntax for something like the ${{ github.repository }} variable for GitHub Actions workflow files with a prompt like this, with all the more relevant search results explicitly mentioning ${{ github.repository }} and most being from official documentation, but only if a non-reasoning model (preferably OpRt_.meta-llama/llama-3.1-405b-instruct ) is used. Many alternative approaches do not provide useful results at all.
https://github.com/mirage335-special/kit-researchEngine/blob/6b0cd98f409672a01b83a98c61dc6a96929ca5fa/_ref/DUBIOUS-speculativePrompts-task/query-openwebui.txt#L71
### Desired Solution you'd like
Separate setting for Merge Responses using Local Models and External Models.
Merge Response treated more as an ordinary response in the chat session.
### Alternatives Considered
_No response_
### Additional Context
_No response_
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Originally created by @mirage335 on GitHub (Mar 25, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/12022
Check Existing Issues
Problem Description
Do a Multi-Chat.
Click the merge icon that appears lower-right after all responses complete.
Merge response uses the 'Task Model'. Bad, very bad.
Merge responses instead, very different from other tasks, should be configured to use the best available chain-of-reasoning (aka. 'thinking') model (ideally OpRt_.anthropic/claude-3.7-sonnet:thinking , or at least OpRt_.deepseek/deepseek-r1-distill-llama-70b).
Web search query generation, as with all other tasks, by contrast suffers a sort of loss of creativity, a narrowing of the possible queries generated, if a reasoning/thinking model is used, resulting in both much more queries generated and much less relevance. Ideally, such tasks are best done with OpRt_.meta-llama/llama-3.1-405b-instruct , or at least Llama-3.1-8b , or some fine-tuned Llama-3.1-8b .
The differences are astonishing.
Merging...
Many-Chat [ OpRt_.perplexity/sonar-deep-research , ChatGPT_.gpt-4.5-preview , OpRt_.anthropic/claude-3.7-sonnet:thinking , ChatGPT_.o3-mini , OpRt_.deepseek/deepseek-r1-distill-llama-70b , OpRt_.meta-llama/llama-3.1-405b-instruct , ChatGPT_.o1 ] -> Merge OpRt_.anthropic/claude-3.7-sonnet:thinking
With prompts such as these:
https://github.com/mirage335-special/kit-researchEngine/blob/a66ed0c8eedef3aa813c52e216a917de69006862/_ref/DUBIOUS-speculativePrompts/special-damageControl-ExplainCommand.txt
https://github.com/mirage335-special/kit-researchEngine/blob/a66ed0c8eedef3aa813c52e216a917de69006862/_ref/DUBIOUS-speculativePrompts/review-code-condensed.txt
Directly results in complete reports which can be used as comprehensive end-user user manuals, or certification of whether code is ready for testing. Whereas all of the contributing models miss multiple crucial features outside their knowledge, logic, or search capabilities.
Merging with a non-reasoning model such as OpRt_.meta-llama/llama-3.1-405b-instruct , however, results in a report too similar to the contributions of other non-reasoning Llama models, still missing crucial features, and poorly formatted.
Web search...
Will correctly search down the syntax for something like the ${{ github.repository }} variable for GitHub Actions workflow files with a prompt like this, with all the more relevant search results explicitly mentioning ${{ github.repository }} and most being from official documentation, but only if a non-reasoning model (preferably OpRt_.meta-llama/llama-3.1-405b-instruct ) is used. Many alternative approaches do not provide useful results at all.
https://github.com/mirage335-special/kit-researchEngine/blob/6b0cd98f409672a01b83a98c61dc6a96929ca5fa/_ref/DUBIOUS-speculativePrompts-task/query-openwebui.txt#L71
Desired Solution you'd like
Separate setting for Merge Responses using Local Models and External Models.
Merge Response treated more as an ordinary response in the chat session.
Alternatives Considered
No response
Additional Context
No response
@tjbck commented on GitHub (Apr 3, 2025):
959995c715