Assistant starts to ignore user prompts after a few Create Run API calls


I’ve encountered an issue where after creating several Runs from within my app via an API call, the Assistant will eventually start to ignore the user’s prompts and either regurgitate a previous response or continue the conversation in the direction it wants it to go.

For Further Context

I’m building a app that will enable non-technical founders to create project plans and scope statements for their new software ideas faster and easier. It works by providing users with a simple and intuitive template that contains a list of fields where they can add in relevant information about their idea. These fields are things like “Problem”, “Solution”, “Success Criteria”, etc.

To start using the template, the user must first create a project in the app. Doing so will automatically create an Assistant and Thread in OpenAI, and save the IDs of each to the project’s object. Each project has its own Assistant and Thread.

In the template, each field has a UI where users can generate content for that specific field. This is done by passing the field’s instructions through the Create Run API call.

The fields are Option Sets, and each Option Set has a text attribute that stores instructions for what the Assistant should generate for that field. So the “Problem” field’s instructions will be “Generate a short paragraph about the problem that this project’s product is looking to solve.”

This all works… at first.

The Issue

Sometimes, I can go down the entire list of fields and generate the appropriate content for each field. But sometimes the Assistant will repeat a previous response, or it will just ignore the field’s instructions and start doing its own thing.

For example, let’s say I generate content for the Problem field, then the Solutions field, then the Constraints and Features fields, but then I go back to regenerate content for the Problem field. What may happen is that the Assistant will keep generating a response as if it’s for the Constraint field or Features field.

OR, it will just take what it knows about the project so far to start generating an entire project scope, including an attempt at generating mockups, milestones, a development plan, etc.

I’m not entirely sure what’s going wrong, but I hope I articulated the problem clearly enough.

Has anyone encountered this issue? I’d be grateful for some tips.

Thank you

1 Like

Did you get this solved? I just started building mine and as soon as I started resuming threads, I’m getting this same issue, it will eventually just repeat its last response. Im sure its my code as the assistant plays nice in playground


Unfortunately, no. My current theory is the pre-determined text instructions that I assign to my fields are worded in a way that eventually initiates a sort of loop or something.

But seeing as how you’re also encountering the issue, and your setup is probably not like mine, maybe this isn’t the case.

I did find this: Assistant is repeating itself in a single run - API - OpenAI Developer Forum

It’s for an older model but it seems the solution back then was to change the model.

I’m using gpt-4o so maybe its not something we’re doing and is instead something about the model.