Skip to main content

Arbitrary context limit

BugsChat

The website artbitrarily limits the context window to 50 prompts (25 user prompts + 25 responses) which is ludicrusly low in most cases and completely negates models with large context windows. It leads to awkward fixes like downloading the conversation, manually removing all the <think> blocks (which is extremely time consuming) then feeding it back to a new conversation as a single prompt - and repeat, until one hits the actual context window.

You can easily test this if you tell it to create a prefix on each prompt, and go beyond 50, then tell it to give you the earliest prefix it recalls.

I doubt this is intentional, it might have been a remainder of early code? Hence why it’s flagged as a bug.

Status: Backlog2 comments

Log in to comment and vote

Comments2

  • Silver River

    ‱

    Jul 12, 2025

    As far as I know, the thinking portion of responses isn’t included in context. It’s likely the documents you’re uploading are being treated differently than normal context imo and that’s why you get a response that seems to “remember“.
    Could be wrong here.
    Oh but I agree about the 25 25 thing.

    • Magenta Volcano

      ‱

      Jul 19, 2025

      Exactly - The thinking isn’t included normally, but the only way to override the 25/25 limit is to download and repost it into a single prompt - which means you need to remove all the <think> tags otherwise they will be taken as normal input. It’s very annoying (and not really a good solution), but literally the only one available right now.Â