Updates to functions might fail if done in same thread

I’ve encountered an issue with custom assistant functions that I’ve noticed others struggling with as well, so I wanted to share some insights.

Issue: If you modify a custom function by adding new parameters after it’s already been called in a thread, subsequent function calls within the same thread might not pass the newly required parameters.

Example Scenario: Consider a function defined as follows:

  "name": "get_weather",
  "description": "Determine weather in my location",
  "parameters": {
    "type": "object",
    "properties": {
      "location": {
        "type": "string",
        "description": "The city and state e.g., San Francisco, CA"
      "unit": {
        "type": "string",
        "enum": ["c", "f"]
    "required": ["location"]

Let’s say you run this assistant thread with the prompt “Get the weather in San Francisco”, and the function runs without passing the “unit” argument. Later, if you modify the function to make “unit” a required parameter, the function might not return that parameter when called again in the same thread.

Hypothesis: This could be because the Large Language Model (LLM) recognizes the previous version of the function where the “unit” parameter wasn’t required, and thus doesn’t include it in subsequent calls.

Solution: To ensure your modified functions work as intended, start a new thread after any modifications. This helps avoid issues where it seems like the LLM isn’t recognizing your updated parameter requirements.

I hope this tip helps others who might be facing similar challenges!

1 Like

If you look in the new thread option, you can see that the full function ‘set’ ie all functions json is part of the thread, together with the assistant prompt. So it makes sense that if you change function parameters on the fly that will not be reflected in an existing thread.
Ultimately I expect Assistants (and then threads) to become ‘versioned’, BUT that would of course still require us on the other side to support two different versions of function calls as well :slight_smile:

1 Like

Yes, that might be the case.

However, I have experienced this with the Chat API as well. When there is a message history where the function was called without a parameter, and then later the parameter is set to required, it doesn’t consider the new update.

Hence, the hypothesis.