Looking for tips for perfect forwarding of information

Prompting
,
1

Hello,

I have a few use cases where the conversation must be summarized, processed and the outcome becomes a synthetic result in a conversation. The best example is related to web searches. One of my use case requires that I ensure that the information provided to the search engines does not provide sensitive details so:

  • We summarize the inquiry;
  • Review the summary for sensitive details (actually request user approval);
  • Send it to search engines (ex. gpt-4o-search-preview or gpt-researcher)
  • Take the outcome and generate a synthetic assistant message where the assistant reflects on the next step - which is to use this information to answer the inquiry;

This flow works well most of the time, but regardless of my prompt, the response tends to summarize and paraphrase the results and removes information in the process.

My question is - has any of you succeeded in prompting such that all information from a source is perfectly replicated in the response? In most of my cases, the user inquiry can be fulfilled by repeating the result of the web search, and yet the model insists for paraphrasing.

I have tried several prompts, and here is one for reference:

After conducting a web search, I have found the following results.
I must now respond to the user inquiry, based on my web search results.
In my response, each statement must be supported by the source url the information is based on.
The user is interested in all the details of the search, I should not summarize anything unless it is clearly irrelevant.
Including the links that support the statements is important, I'll make sure I don't forget any.
At the end of the response, I will also summarize the links that were found in a list, 
even if they were not relevant to responding to the inquiry.
When providing the links, I use the hostname as a label for the link when I have nothing more precise to use.

If the response does not fully answer the inquiry, I will think of additional details that could be provided, or another way the user could formulate the inquiry.

### Web Search Results
{webContext}

Any tips appreciated, thank you!

1 Like