Skip to main content
Perceptron Mk1.5 can request functions that you define. Give the model descriptions and parameter schemas, then let it choose which functions to call while answering a question. Your application executes the calls and sends the results back so the model can continue. Use tools to connect visual understanding to information outside the supplied images or videos: look up a product identified in an image, retrieve an inspection record, or search your own document collection.

​
How tool calling works

  1. Send a question and function definitions to /v1/chat/completions.
  2. The model returns tool_calls with function names, argument strings, and call IDs.
  3. Your application validates and executes each call.
  4. Send the assistant message and a role: "tool" result for every call back to the model.
  5. Repeat until the model returns its answer or your application reaches its budget.
One response can request several functions. Each result is paired with its call through tool_call_id, so independent functions can run concurrently.
The public API returns function requests. Your application provides the implementations, credentials, and execution. Naming a function search does not enable a hosted search service.

​
Choose the right output

Function parameter schemas describe the expected arguments; they do not constrain decoding. Validate names and arguments before execution, even if a function declares strict: true. Tools cannot be combined with JSON-schema output or regex in the same request; the Perceptron SDK rejects this combination before sending it.

​
Build a tool workflow

Call a function

Run a complete Python example, handle streaming calls, and learn the supported controls.

Build a tool agent

Add execution budgets, return images, and combine retrieval with visual reasoning.
For the model ID and capabilities, see Perceptron Mk1.5. Applications selecting models programmatically can request GET /v1/models?extended=true and check for tool_calling in the model’s capabilities array.