The challenge of getting and managing the discussion context

Request for Comments (RFC): Enhancing Discussion Context Retrieval and Maintenance

Background: In our recent series of discussions and system interactions, we’ve encountered challenges in maintaining and retrieving accurate and contextually relevant information from discussions spanning multiple requests. The objective is to achieve a cohesive, context-rich dialogue that can guide users effectively, avoiding side-tracks and ensuring purposeful conversation.

Original Request: We requested a list of all verbatim paragraph titles found within replies for a specified range of discussions. This included providing direct verbatim quotations of conclusions or the entire content of the last paragraph of each reply. The aim was to preserve the context and ensure accuracy in our documentation and discussion tracking. It was imperative that any interpretations or summarizations of excerpts would be incorrect and that titles and excerpts from conclusions or the last paragraphs must be exact verbatim copies from the replies, with no alterations.

Observed Incorrectness:

  1. The list of replies provided did not cover the complete specified range, omitting several key discussions.
  2. The extraction of paragraph titles and the conclusions was not complete, with many titles missing from the provided summaries.
  3. There was a tendency for the AI to either summarize or interpret the titles and conclusions, despite specific instructions for verbatim excerpts.
  4. There was a mismatch in information, where extracted paragraph titles and conclusion excerpts did not correspond to the correct replies, leading to context misalignment.

Request for Community Input: We are reaching out to the OpenAI community to partake in providing a more solid solution to these challenges. The goal is to refine the approach of extracting and presenting discussion contexts, particularly in a format that supports dynamic and multi-threaded dialogue scenarios.

Specific Areas for Feedback:

  • Methods or tools that could improve the accuracy of context extraction and retention over a range of discussions.
  • Strategies to enforce the adherence to verbatim extraction specifications without deviations or interpretations by the AI.
  • Suggestions on system adjustments or training methodologies that could help the AI better understand and fulfill the requirements of maintaining discussion integrity and context accuracy.

Conclusion: Maintaining the essence and continuity of discussions through accurate context handling is crucial for productive engagements on complex topics. Your insights and suggestions will be invaluable in enhancing the capabilities of AI-driven dialogue systems to manage discussions effectively, ensuring they remain on track and contextually coherent.

Please share your thoughts, experiences, or any relevant tools or strategies that could assist in addressing these challenges. Your contribution will play a significant role in shaping the next steps in our endeavor to improve discussion context management.

This RFC aims to open a productive dialogue within the community to find more robust solutions for the issues we’ve encountered. The feedback from this initiative will be crucial in optimizing how discussion contexts are handled and presented in AI-assisted platforms.


TLDR: you want endless context in chats, is that correct?

I am not sure what you mean by ‘endless’, but a periodic discussion context gathering would be inline with the objective.

You’re talking about the platform API and not ChatGPT, right? Langchain and other tools (as well as popular clients) have had features like this for quite a while I believe. :thinking:

1 Like

I guess I was referring to both. At any rates, thanks to the introduction to Langchain. I will look it up

1 Like