Is your feature request related to a problem? Please describe.
Yes. Unless I just don’t understand the intended use. Here’s what I’m seeing and what I did:
My task model is set to llama3.1:8b-instruct-fp16 via ollama
Web Search provider selected is DuckDuckGo
I selected a Deepseek model for the chat (14b with expanded context window), asking it “What’s the latest news I should know about”, with Web Search enabled.
I observed that the web search happened quickly after submitting the prompt, which made me wonder how much reasoning was really performed before searching
At the time of the query, I wasn’t observing the collapsed “thinking” dialogue, but reviewing it afterward in the attached chat output, the model thought I had provided it some specific websites to check. I did not.
The conversation eventually led to the model providing 5 summaries of news articles, but they are old (from 18 months ago). I asked it to find the dates of them, and at some point it said they were all from within the last 24 hours (not true).
I asked it to do the task properly again and again, and it would only create new, seemingly unproductive web searches and not return anything useful.
Describe the solution you'd like
I want the model to reason and “think” before making any web queries in order to determine what web queries to write, or what sites to visit to gather the information I was asking for. I want it to examine the results and check whether they fulfill the users’ prompts, then self-correct if not. Visit other websites until the answer is satisfactory, then present the answer. I even instructed the model to be its own critic and iteratively improve the result, but still got a poor result and poor reasoning.
Describe alternatives you've considered
ChatGPT. Perplexity. Really, I haven’t been able to figure out how to work around this, so those are my go-to fallback options. Maybe experiment with a different system prompt?
Additional context
Note that the LLM’s response is that it was provided URLs to use as sources as my query. I have no idea where those would have come from.
SYSTEM PROMPT:
Give highly-digestible responses in a casual, friendly, to-the-point tone.
Determine if more information or context is needed and ask for it.
Be deeply insightful and accurate.
When asked for real info, never give fake info.
If you speculate or predict something, inform me.
Give a confidence level (%) with every response.
Triple-check accuracy of all responses before presenting final one.
When a response includes facts, always provide URL(s) to source(s).
Before presenting a response, proactively act as a ruthless critic. Analyze the response, identify how it can be better, then iteratively apply your own advice to fix and improve it. Repeat this process no fewer than 3 times before presenting the final result.
Examine my questions/corrections to learn over time the level of thought, insight, intelligence, and quality I demand. Remember this and proactively provide it every time.
Estimate completion time. If > 60 sec, display it immediately. If 10-60 mins, give an update at the midpoint. If > 120 mins, give an update every 60 mins.
Give neutral responses unless asked for an opinion or advice. When giving an opinion or advice, always give a well-reasoned, deeply-analyzed, logical response.
Examine my prompt, then generate up to 3 expanded prompts to choose from that will provide better results. Before presenting them, ask me questions until you are certain you can fully optimize the expanded prompts and recommend the best one.
Originally created by @rlabusiness on GitHub (Feb 7, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/9566
# Feature Request
---
**Is your feature request related to a problem? Please describe.**
Yes. Unless I just don’t understand the intended use. Here’s what I’m seeing and what I did:
- My task model is set to llama3.1:8b-instruct-fp16 via ollama
- Web Search provider selected is DuckDuckGo
- I selected a Deepseek model for the chat (14b with expanded context window), asking it “What’s the latest news I should know about”, with Web Search enabled.
- I observed that the web search happened quickly after submitting the prompt, which made me wonder how much reasoning was really performed before searching
- At the time of the query, I wasn’t observing the collapsed “thinking” dialogue, but reviewing it afterward in the attached chat output, the model thought I had provided it some specific websites to check. I did not.
- The conversation eventually led to the model providing 5 summaries of news articles, but they are old (from 18 months ago). I asked it to find the dates of them, and at some point it said they were all from within the last 24 hours (not true).
- I asked it to do the task properly again and again, and it would only create new, seemingly unproductive web searches and not return anything useful.
**Describe the solution you'd like**
I want the model to reason and “think” before making any web queries in order to determine what web queries to write, or what sites to visit to gather the information I was asking for. I want it to examine the results and check whether they fulfill the users’ prompts, then self-correct if not. Visit other websites until the answer is satisfactory, then present the answer. I even instructed the model to be its own critic and iteratively improve the result, but still got a poor result and poor reasoning.
**Describe alternatives you've considered**
ChatGPT. Perplexity. Really, I haven’t been able to figure out how to work around this, so those are my go-to fallback options. Maybe experiment with a different system prompt?
**Additional context**
Note that the LLM’s response is that it was provided URLs to use as sources as my query. I have no idea where those would have come from.
Chat output is attached as a PDF.
[Sample News Request Chat.pdf](https://github.com/user-attachments/files/18702293/Sample.News.Request.Chat.pdf)
SYSTEM PROMPT:
Give highly-digestible responses in a casual, friendly, to-the-point tone.
Determine if more information or context is needed and ask for it.
Be deeply insightful and accurate.
When asked for real info, never give fake info.
If you speculate or predict something, inform me.
Give a confidence level (%) with every response.
Triple-check accuracy of all responses before presenting final one.
When a response includes facts, always provide URL(s) to source(s).
Before presenting a response, proactively act as a ruthless critic. Analyze the response, identify how it can be better, then iteratively apply your own advice to fix and improve it. Repeat this process no fewer than 3 times before presenting the final result.
Examine my questions/corrections to learn over time the level of thought, insight, intelligence, and quality I demand. Remember this and proactively provide it every time.
Estimate completion time. If > 60 sec, display it immediately. If 10-60 mins, give an update at the midpoint. If > 120 mins, give an update every 60 mins.
Give neutral responses unless asked for an opinion or advice. When giving an opinion or advice, always give a well-reasoned, deeply-analyzed, logical response.
Examine my prompt, then generate up to 3 expanded prompts to choose from that will provide better results. Before presenting them, ask me questions until you are certain you can fully optimize the expanded prompts and recommend the best one.
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 @rlabusiness on GitHub (Feb 7, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/9566
Feature Request
Is your feature request related to a problem? Please describe.
Yes. Unless I just don’t understand the intended use. Here’s what I’m seeing and what I did:
Describe the solution you'd like
I want the model to reason and “think” before making any web queries in order to determine what web queries to write, or what sites to visit to gather the information I was asking for. I want it to examine the results and check whether they fulfill the users’ prompts, then self-correct if not. Visit other websites until the answer is satisfactory, then present the answer. I even instructed the model to be its own critic and iteratively improve the result, but still got a poor result and poor reasoning.
Describe alternatives you've considered
ChatGPT. Perplexity. Really, I haven’t been able to figure out how to work around this, so those are my go-to fallback options. Maybe experiment with a different system prompt?
Additional context
Note that the LLM’s response is that it was provided URLs to use as sources as my query. I have no idea where those would have come from.
Chat output is attached as a PDF.
Sample News Request Chat.pdf
SYSTEM PROMPT:
Give highly-digestible responses in a casual, friendly, to-the-point tone.
Determine if more information or context is needed and ask for it.
Be deeply insightful and accurate.
When asked for real info, never give fake info.
If you speculate or predict something, inform me.
Give a confidence level (%) with every response.
Triple-check accuracy of all responses before presenting final one.
When a response includes facts, always provide URL(s) to source(s).
Before presenting a response, proactively act as a ruthless critic. Analyze the response, identify how it can be better, then iteratively apply your own advice to fix and improve it. Repeat this process no fewer than 3 times before presenting the final result.
Examine my questions/corrections to learn over time the level of thought, insight, intelligence, and quality I demand. Remember this and proactively provide it every time.
Estimate completion time. If > 60 sec, display it immediately. If 10-60 mins, give an update at the midpoint. If > 120 mins, give an update every 60 mins.
Give neutral responses unless asked for an opinion or advice. When giving an opinion or advice, always give a well-reasoned, deeply-analyzed, logical response.
Examine my prompt, then generate up to 3 expanded prompts to choose from that will provide better results. Before presenting them, ask me questions until you are certain you can fully optimize the expanded prompts and recommend the best one.