I’m considering DeepSeek for an app that asks the model to call tools, then uses the returned arguments to run a search or query a database. The API documentation looks compatible with familiar chat-completions patterns, but I’m not sure how much that guarantees about actual behavior.

For people using function calling in DeepSeek, do tool calls work similarly across DeepSeek’s own API and third-party hosts? I’d like to know whether tool selection and tool-call arguments tend to be predictable enough to build around.