OpenAI Functions Calling

OpenAI’s implementation of llm-native-function-calling on the Chat Completions API: you pass tools, the model replies with tool_calls, and you return results as role: "tool" messages.

Why it matters

It’s the most widely-copied dialect — most OpenAI-compatible servers (vLLM, Together, Groq, LiteLLM) speak it, so learning it transfers broadly. The 2023 functions/function_call fields are deprecated; the current shape is tools (a list of {type: "function", function: {...}}) plus tool_choice.

How it works

Each tool is a function with a JSON-Schema parameters block. The model returns one or more tool_calls, each with an id, a name, and arguments as a JSON string (you must json.loads it).

• tool_choice — "auto" (default), "none", "required" (must call something), or {"type":"function","function":{"name":"x"}} to force a specific one.
• The tool_call_id handshake — every tool result message must echo the tool_call_id from the call. Mismatch or a missing result for any call in the turn is a 400 error.
• strict: true — opt into Structured Outputs so arguments are guaranteed to satisfy the schema (requires additionalProperties: false and all fields required).
• Parallel calls — on by default; set parallel_tool_calls: false to force one at a time.

Field	Where	Note
tools	request	list of function schemas
tool_calls	response	arguments is a JSON string
role: "tool"	follow-up	must carry tool_call_id

Example

resp = client.chat.completions.create(
  model="gpt-4o", messages=msgs, tools=tools, tool_choice="auto")
call = resp.choices[0].message.tool_calls[0]
args = json.loads(call.arguments)          # arguments is a string!
out  = run_tool(call.function.name, args)
msgs += [resp.choices[0].message,
         {"role":"tool","tool_call_id":call.id,"content":str(out)}]
# call the API again with the appended tool result

The newer Responses API uses function_call / function_call_output items instead, but the loop is identical.

Pitfalls

• Forgetting arguments is a string — it’s not a dict; parse it (and guard against invalid JSON despite strict).
• Dropping the assistant tool_calls message — you must append the model’s call message before the tool result, or the next request 400s.
• Partial results — if the model made 3 calls you must return 3 results; one missing breaks the turn.
• strict schema rules — every property must be required and additionalProperties:false; optional fields are expressed via union with null.

See also

• anthropic-tool-use
• openai-assistant-api