OpenAI Responses 完整交互示例

完整交互示例

下面用同一个跑步建议场景展示 Responses 的 item 化交互。

用户问：“北京今天适合跑步吗？如果空气质量不好，请查一下官方建议。”

Responses 同时支持用户端工具和 hosted/server 工具。这个例子里：

function get_weather 是用户端工具，模型返回 function_call，客户端执行后下一轮用 function_call_output 回填。
web_search 是 hosted/server 工具，上游服务自己执行，流里出现 web search 状态事件和对应 WebSearchToolCall item。

Client
  |
  | 1. input + function tool + web_search tool + stream=true
  v
proxai
  |
  | 2. 归一化 Responses 请求后转发
  v
OpenAI-compatible upstream
  |
  | 3. SSE: web_search 状态 + function_call arguments
  v
proxai
  |
  | 4. 保留 SSE bytes，Responses observer 监控 terminal/tool 参数超时
  v
Client
  |
  | 5. 本地执行 get_weather
  v
Local tool runtime
  |
  | 6. 下一轮 input 带 function_call_output
  v
proxai -> upstream -> proxai -> Client

第一轮请求

{
  "model": "gpt-5.4",
  "stream": true,
  "input": [
    {
      "type": "message",
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "北京今天适合跑步吗？如果空气质量不好，请查一下官方建议。"
        }
      ]
    }
  ],
  "tools": [
    {
      "type": "function",
      "name": "get_weather",
      "description": "查询指定城市的天气和空气质量摘要。",
      "parameters": {
        "type": "object",
        "properties": {
          "city": { "type": "string" },
          "date": { "type": "string" }
        },
        "required": ["city", "date"]
      },
      "strict": true
    },
    {
      "type": "web_search",
      "filters": {
        "allowed_domains": ["www.cma.gov.cn", "www.mee.gov.cn"]
      },
      "search_context_size": "low"
    }
  ],
  "tool_choice": "auto",
  "parallel_tool_calls": false
}

结构映射。这里是字段映射伪代码，json! 表示 serde_json::json!，... 表示其余字段省略：

RequestProjection {
    model: Some("gpt-5.4".to_string()),
    stream: Some(true),
    tools: Some(vec![
        Tool::Function(FunctionTool {
            name: "get_weather".to_string(),
            description: Some("查询指定城市的天气和空气质量摘要。".to_string()),
            parameters: Some(json!({
                "type": "object",
                "properties": {
                    "city": { "type": "string" },
                    "date": { "type": "string" }
                },
                "required": ["city", "date"]
            })),
            strict: Some(true),
            ...
        }),
        Tool::WebSearch(WebSearchTool {
            filters: Some(WebSearchToolFilters {
                allowed_domains: Some(vec![
                    "www.cma.gov.cn".to_string(),
                    "www.mee.gov.cn".to_string(),
                ]),
            }),
            search_context_size: Some(WebSearchToolSearchContextSize::Low),
            ...
        }),
    ]),
    parallel_tool_calls: Some(false),
    ...
}

RequestProjection 刻意不保留完整 input。如果只看 wire model，输入消息对应：

InputParam::Items(vec![
    InputItem::Item(Item::Message(MessageItem::Input(InputMessage {
        role: InputRole::User,
        content: vec![
            InputContent::InputText(InputTextContent {
                text: "北京今天适合跑步吗？如果空气质量不好，请查一下官方建议。".to_string(),
            }),
        ],
        ...
    }))),
])

第一轮 SSE

event: response.created
data: {
  "type": "response.created",
  "sequence_number": 0,
  "response": {
    "id": "resp_01",
    "object": "response",
    "status": "in_progress",
    "model": "gpt-5.4",
    "output": []
  }
}

event: response.web_search_call.in_progress
data: {
  "type": "response.web_search_call.in_progress",
  "sequence_number": 1,
  "output_index": 0,
  "item_id": "ws_01"
}

event: response.web_search_call.searching
data: {
  "type": "response.web_search_call.searching",
  "sequence_number": 2,
  "output_index": 0,
  "item_id": "ws_01"
}

event: response.web_search_call.completed
data: {
  "type": "response.web_search_call.completed",
  "sequence_number": 3,
  "output_index": 0,
  "item_id": "ws_01"
}

event: response.output_item.added
data: {
  "type": "response.output_item.added",
  "sequence_number": 4,
  "output_index": 1,
  "item": {
    "type": "function_call",
    "id": "fc_01",
    "call_id": "call_weather_01",
    "name": "get_weather",
    "arguments": "",
    "status": "in_progress"
  }
}

event: response.function_call_arguments.delta
data: {
  "type": "response.function_call_arguments.delta",
  "sequence_number": 5,
  "item_id": "fc_01",
  "output_index": 1,
  "delta": "{\"city\":\"北京\""
}

event: response.function_call_arguments.delta
data: {
  "type": "response.function_call_arguments.delta",
  "sequence_number": 6,
  "item_id": "fc_01",
  "output_index": 1,
  "delta": ",\"date\":\"today\"}"
}

event: response.function_call_arguments.done
data: {
  "type": "response.function_call_arguments.done",
  "sequence_number": 7,
  "item_id": "fc_01",
  "output_index": 1,
  "name": "get_weather",
  "arguments": "{\"city\":\"北京\",\"date\":\"today\"}"
}

event: response.output_item.done
data: {
  "type": "response.output_item.done",
  "sequence_number": 8,
  "output_index": 1,
  "item": {
    "type": "function_call",
    "id": "fc_01",
    "call_id": "call_weather_01",
    "name": "get_weather",
    "arguments": "{\"city\":\"北京\",\"date\":\"today\"}",
    "status": "completed"
  }
}

event: response.completed
data: {
  "type": "response.completed",
  "sequence_number": 9,
  "response": {
    "id": "resp_01",
    "object": "response",
    "status": "completed",
    "model": "gpt-5.4",
    "output": []
  }
}

结构映射：

ResponseStreamEvent::ResponseCreated(ResponseCreatedEvent {
    sequence_number: 0,
    response: Response {
        id: "resp_01".to_string(),
        object: "response".to_string(),
        status: Status::InProgress,
        model: "gpt-5.4".to_string(),
        output: vec![],
        ...
    },
})

ResponseStreamEvent::ResponseWebSearchCallInProgress(
    ResponseWebSearchCallInProgressEvent {
        sequence_number: 1,
        output_index: 0,
        item_id: "ws_01".to_string(),
    },
)

ResponseStreamEvent::ResponseWebSearchCallSearching(
    ResponseWebSearchCallSearchingEvent {
        sequence_number: 2,
        output_index: 0,
        item_id: "ws_01".to_string(),
    },
)

ResponseStreamEvent::ResponseWebSearchCallCompleted(
    ResponseWebSearchCallCompletedEvent {
        sequence_number: 3,
        output_index: 0,
        item_id: "ws_01".to_string(),
    },
)

ResponseStreamEvent::ResponseOutputItemAdded(ResponseOutputItemAddedEvent {
    sequence_number: 4,
    output_index: 1,
    item: OutputItem::FunctionCall(FunctionToolCall {
        id: Some("fc_01".to_string()),
        call_id: "call_weather_01".to_string(),
        name: "get_weather".to_string(),
        arguments: String::new(),
        status: Some(OutputStatus::InProgress),
        ...
    }),
})

ResponseStreamEvent::ResponseFunctionCallArgumentsDelta(
    ResponseFunctionCallArgumentsDeltaEvent {
        sequence_number: 5,
        item_id: "fc_01".to_string(),
        output_index: 1,
        delta: "{\"city\":\"北京\"".to_string(),
    },
)

ResponseStreamEvent::ResponseFunctionCallArgumentsDelta(
    ResponseFunctionCallArgumentsDeltaEvent {
        sequence_number: 6,
        item_id: "fc_01".to_string(),
        output_index: 1,
        delta: ",\"date\":\"today\"}".to_string(),
    },
)

ResponseStreamEvent::ResponseFunctionCallArgumentsDone(
    ResponseFunctionCallArgumentsDoneEvent {
        sequence_number: 7,
        item_id: "fc_01".to_string(),
        output_index: 1,
        name: Some("get_weather".to_string()),
        arguments: "{\"city\":\"北京\",\"date\":\"today\"}".to_string(),
    },
)

客户端聚合工具参数：

function_calls["fc_01"].call_id = "call_weather_01"
function_calls["fc_01"].name = "get_weather"
function_calls["fc_01"].arguments += "{\"city\":\"北京\""
function_calls["fc_01"].arguments += ",\"date\":\"today\"}"

delta 是新增参数片段，不是完整 JSON。arguments.done 到达后，客户端可以解析完整参数。proxai 的 Responses observer 会专门监控这种参数流，避免 delta 后长期没有 done 导致客户端卡住。

本地工具结果回填

客户端执行 get_weather 后，下一轮请求把工具结果作为 function_call_output input item 带回。

{
  "model": "gpt-5.4",
  "stream": true,
  "previous_response_id": "resp_01",
  "input": [
    {
      "type": "function_call_output",
      "call_id": "call_weather_01",
      "output": "北京今天气温 18-27C，轻度污染，PM2.5 约 85，傍晚有风。"
    }
  ]
}

结构映射：

RequestProjection {
    model: Some("gpt-5.4".to_string()),
    stream: Some(true),
    previous_response_id: Some("resp_01".to_string()),
    ...
}

InputParam::Items(vec![
    InputItem::Item(Item::FunctionCallOutput(FunctionCallOutputItemParam {
        call_id: "call_weather_01".to_string(),
        output: FunctionCallOutput::Text(
            "北京今天气温 18-27C，轻度污染，PM2.5 约 85，傍晚有风。".to_string(),
        ),
        ...
    })),
])

第二轮 SSE 最终回答

event: response.output_item.added
data: {
  "type": "response.output_item.added",
  "sequence_number": 0,
  "output_index": 0,
  "item": {
    "type": "message",
    "id": "msg_01",
    "role": "assistant",
    "content": [],
    "status": "in_progress"
  }
}

event: response.content_part.added
data: {
  "type": "response.content_part.added",
  "sequence_number": 1,
  "item_id": "msg_01",
  "output_index": 0,
  "content_index": 0,
  "part": {
    "type": "output_text",
    "text": "",
    "annotations": []
  }
}

event: response.output_text.delta
data: {
  "type": "response.output_text.delta",
  "sequence_number": 2,
  "item_id": "msg_01",
  "output_index": 0,
  "content_index": 0,
  "delta": "今天北京不太适合高强度户外跑步。"
}

event: response.output_text.delta
data: {
  "type": "response.output_text.delta",
  "sequence_number": 3,
  "item_id": "msg_01",
  "output_index": 0,
  "content_index": 0,
  "delta": "空气质量为轻度污染，建议改为低强度慢跑或室内训练。"
}

event: response.output_text.done
data: {
  "type": "response.output_text.done",
  "sequence_number": 4,
  "item_id": "msg_01",
  "output_index": 0,
  "content_index": 0,
  "text": "今天北京不太适合高强度户外跑步。空气质量为轻度污染，建议改为低强度慢跑或室内训练。"
}

event: response.completed
data: {
  "type": "response.completed",
  "sequence_number": 5,
  "response": {
    "id": "resp_02",
    "object": "response",
    "status": "completed",
    "model": "gpt-5.4",
    "output": []
  }
}

结构映射和聚合规则：

ResponseStreamEvent::ResponseOutputTextDelta(ResponseTextDeltaEvent {
    sequence_number: 2,
    item_id: "msg_01".to_string(),
    output_index: 0,
    content_index: 0,
    delta: "今天北京不太适合高强度户外跑步。".to_string(),
    ...
})

ResponseStreamEvent::ResponseOutputTextDelta(ResponseTextDeltaEvent {
    sequence_number: 3,
    item_id: "msg_01".to_string(),
    output_index: 0,
    content_index: 0,
    delta: "空气质量为轻度污染，建议改为低强度慢跑或室内训练。".to_string(),
    ...
})

ResponseStreamEvent::ResponseOutputTextDone(ResponseTextDoneEvent {
    sequence_number: 4,
    item_id: "msg_01".to_string(),
    output_index: 0,
    content_index: 0,
    text: "今天北京不太适合高强度户外跑步。空气质量为轻度污染，建议改为低强度慢跑或室内训练。".to_string(),
    ...
})

text[(output_index=0, content_index=0)] = ""
text[(0, 0)] += "今天北京不太适合高强度户外跑步。"
text[(0, 0)] += "空气质量为轻度污染，建议改为低强度慢跑或室内训练。"

output_text.delta 是新增文本片段；output_text.done.text 是该 content part 的完整文本。