Openai Realtime Model Functional Calling Event not triggered

I’ve been working on a voice-to-voice agent using OpenAI’s realtime models and WebSockets. The pipeline was working fine before, but since yesterday I’ve started running into issues. Specifically, the function-calling event no longer gets triggered when I ask a query that should invoke the retriever function. Could you confirm if this is an issue on your side, or is it likely something on my end? Earlier it was working without problems.

Here is my code to declare functions

{

        "type": "session.update",

        "session": {

            "model": "gpt-4o-realtime-preview", # gpt-4o-realtime-preview

            'temperature': 0.6,

            "instructions" : System_prompt_updated,

            "input_audio_transcription": { "model": "gpt-4o-mini-transcribe" },

            "voice":"ash",

            "turn_detection": {

                    "type": "server_vad",

                    "threshold": 0.9,            #how confident the VAD must be before ending a turn

                    "prefix_padding_ms": 300,    #include 300 ms of audio before speech start

                    "silence_duration_ms": 200,  #require 200 ms of silence to mark end

                    "create_response": True      #auto-trigger a response.create for you

                    },

            "tools": \[

                {

                    "type": "function",

                    "name": "get_qdrantdb_result",

                    "description": "Fetch relevant info from documents using embeddings.",

                    "parameters": {

                        "type": "object",

                        "properties": {

                            "query": {

                                "type": "string",

                                "description": "The user's query for retrieval."

                            }

                        },

                        "required": \["query"\]

                    }

                }\],

            "tool_choice": "auto"

        }

    }
1 Like

I’ve got the same problem, func calls just stopped working at some point

2 Likes

Yes! Same here. Everything was working fine 1-2 days ago and then the functions just don’t get called anymore. I spent hours trying to figure it out and hoping it’s a system issue.

1 Like

Having the same issue! Had massive failures in production today due to this.

1 Like

I have the same issue. There is an earlier thread here Realtime API is currently unable to use function calling

1 Like

Same here! my Realtime assistants just are not working.
Function Tool Calls returns a generic “server error“

When will this be fixed?

1 Like

same issue I’m facing…
We are on production and the work arounds are not gonna do it.

I’ve tried passing it to completions endpoint via response.transcription.done but it’s not gonna pull it off

Are you still seeing the problem?

I’m not using mini-transcribe, but have a RT-STS assistant with over 200 tools that are working great. I have noticed that the tool chains are sensitive to the translation of the request to the tool name in the chain and description. The “request” from the user in verbal form needs to align to the tool chain. I’ve actually started naming my functions to match more based on vocabulary instead of thinking of them as a traditional function call. Not sure if this helps, but you can always experiment by adding the same function under different names/descriptions and see which is more likely to match. The other suggestion might be to eliminate the pass through the mini-transcribe so you aren’t getting a lossy translation as input (or add a debug output to show the transcription text that is being input to ensure it’s a good transcription).

2 Likes

The issue has been resolved — I tested the application and it’s working fine now.

1 Like

Everything Works now! Thanks for the heads up!