Use tools, structured output, and streaming

Test XiuRouter tool calls, JSON Schema output, and Chat Completions streaming separately with your target model.

Tool calling, structured output, and streaming are separate capabilities. First test ordinary text with the same model, then verify each capability you need.

Tool calling

This request lets the model select a tool; it does not execute the function:

curl https://router-api.xiu.ai/v1/chat/completions \
  -H "Authorization: Bearer $XIUROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "YOUR_MODEL_ID",
    "messages": [
      {
        "role": "user",
        "content": "What time is it in Taipei? Use the tool."
      }
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_time",
          "description": "Read the current time in the specified timezone",
          "parameters": {
            "type": "object",
            "properties": {
              "utc_offset": {
                "type": "string"
              }
            },
            "required": ["utc_offset"],
            "additionalProperties": false
          }
        }
      }
    ],
    "tool_choice": "auto"
  }'

A successful response includes the tool name and arguments in choices[0].message.tool_calls. Your application must then:

Validate the tool name and arguments.
Execute the tool in a controlled environment.
Return the result as a role: "tool" message.
Read the model’s final answer.

If text works but no tool_calls appear, check the selected model and service group’s tool support. Do not fabricate a tool result in the client.

Structured output

Chat Completions accepts response_format. This example requests a fixed JSON structure:

curl https://router-api.xiu.ai/v1/chat/completions \
  -H "Authorization: Bearer $XIUROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "YOUR_MODEL_ID",
    "messages": [
      {
        "role": "user",
        "content": "Classify Beijing as a city."
      }
    ],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "city",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "name": {"type": "string"},
            "country": {"type": "string"}
          },
          "required": ["name", "country"],
          "additionalProperties": false
        }
      }
    }
  }'

The gateway parses response_format, but the target model may support only json_object, reject strict schemas, or lack structured output entirely. Production code must still parse and validate the returned value.

Streaming

Use curl -N to avoid local buffering:

curl -N https://router-api.xiu.ai/v1/chat/completions \
  -H "Authorization: Bearer $XIUROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "YOUR_MODEL_ID",
    "messages": [
      {
        "role": "user",
        "content": "Explain an API gateway in three sentences."
      }
    ],
    "stream": true,
    "stream_options": {
      "include_usage": true
    }
  }'

Chat Completions streams normally contain SSE data: events and a completion event. Responses uses different event types. Parse the actual protocol rather than mixing the two stream formats.

Verify the result

Tools: receive correctly structured tool_calls and complete a read-only tool round trip.
Structured output: validate the response against your JSON Schema.
Streaming: receive multiple events and a complete termination signal, then confirm the streaming flag and final cost in usage logs.

Troubleshoot

Symptom	Next step
Plain text instead of a tool call	Try a model with verified tool support, or explicitly select a tool through `tool_choice`
`response_format` rejected	Use `json_object` or plain text and validate it yourself
Stream stops early	Confirm the direct `router-api.xiu.ai` endpoint; keep the request ID and check usage logs first
No usage event at the end	The upstream may not support `include_usage`; reconcile with usage logs
Tool arguments are not valid JSON	Reject execution, keep the original arguments, and ask the model to generate them again

Sources and review date

Reviewed on 2026-08-31. This translation preserves that date. Tool calls, structured output, and stream events vary by model and service tier. Verify each with the actual target model and reconcile the corresponding usage log before production use.