I’ve implemented a voice agent using the Realtime API with WebRTC in simple HTML/CSS/JS, and now want to enhance it with vector search capabilities.
We have vector stores on platform[dot]openai[dot]com. How can I use a vector store ID from there and pass it as a function tool call to this WebRTC Realtime API? Is it supported?
Realtime does not support “internal tools” like the Responses endpoint, to be able to directly employ “file_search”.
It does support function calling.
You can use the API endpoint for vector store search (not with any tool), writing your own RESTful call to it with a query.
You can then offer that as a developer function to the realtime AI model.
Then the idea falls apart when you are using client WebRTC, which doesn’t fit any pattern of security you’d actually want to implement and give an untrusted user.
At this point, 30.09.2025, the Open AI API Agents SDK DOES support tools, including file_search with your VS ID for Realtime. My testing shows that in conversation, it does have knowledge of the VS added to file_search tool.
More info - Tools | OpenAI Agents SDK
That said, the SDK is an abstraction layer on top of OpenAI APIs, haven’t dug in, how the file_search is actually implemented and how it can be replicated with pure API calls.