OpenAI 代理 SDK (Node.js) 的 OpenInference 仪器

以 OpenTelemeth 为基础的仪器 OpenAI 代理服务器 SDK (单位:千美元)@openai/agents) (中文(简体) ). 将 SDK 的本地跟踪事件连接到 OpenTeleometry 之后的 开源语义公约,因此可以在任何 OpenTeleometry 兼容后端中观察到代理运行,例如 阿里兹·凤凰, (中文(简体) ), 调整 AXJaeger 或者你的收藏家选择的.

安装

npm install @arizeai/openinference-instrumentation-openai-agents @arizeai/openinference-semantic-conventions @openai/agents

快速启动

import * as agents from "@openai/agents";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { Resource } from "@opentelemetry/resources";
import { BatchSpanProcessor } from "@opentelemetry/sdk-trace-base";
import { NodeTracerProvider } from "@opentelemetry/sdk-trace-node";
import { ATTR_SERVICE_NAME } from "@opentelemetry/semantic-conventions";

import { OpenAIAgentsInstrumentation } from "@arizeai/openinference-instrumentation-openai-agents";

// 1. Configure OpenTelemetry.
const provider = new NodeTracerProvider({
  resource: new Resource({
    [ATTR_SERVICE_NAME]: "my-agent-app",
  }),
  spanProcessors: [
    new BatchSpanProcessor(new OTLPTraceExporter({ url: "http://localhost:6006/v1/traces" })),
  ],
});
provider.register();

// 2. Register the OpenInference processor with the agents SDK.
const instrumentation = new OpenAIAgentsInstrumentation({ tracerProvider: provider });
instrumentation.manuallyInstrument(agents);

// 3. Use the agents SDK as usual.
const agent = new agents.Agent({
  name: "Assistant",
  instructions: "You are a helpful assistant.",
});

const result = await agents.run(agent, "What is the capital of France?");
console.log(result.finalOutput);

如何运作

与大多数 OpenInference 仪器不同,此软件包确实 没有 给SDK打个针 特工SDK曝光头等舱 TracingProcessor 接口; 此软件包通过 SDK 执行并注册 setTraceProcessors 编号 : addTraceProcessor APIs (英语).

当模块自动被补丁时(通过) registerInstrumentations 或 NodeSDK),通过构造器选择模式取而代之: new OpenAIAgentsInstrumentation({ exclusiveProcessor: false })。 。 。 .

为了停止追踪,打电话 instrumentation.uninstrument()。在添加剂模式下,SDK剂不会暴露出单处理器去除API,所以 uninstrument() 不清理其他 SDK 处理器而禁用已设置的 OpenInference 处理器.

对于仅限无害环境管理的环境,通过进口的SDK命名空间:

import * as agents from "@openai/agents";

const instrumentation = new OpenAIAgentsInstrumentation({ tracerProvider: provider });
instrumentation.manuallyInstrument(agents, { exclusiveProcessor: false });

配置

new OpenAIAgentsInstrumentation({
  // Optional: an OTel TracerProvider. Defaults to the global provider.
  tracerProvider,

  // Optional: processor registration mode (see "How it works" above).
  // This is the only way to choose additive mode when the module is patched
  // automatically (registerInstrumentations / NodeSDK). Defaults to true.
  exclusiveProcessor: true,

  // Optional: OpenInference trace configuration for masking/redacting
  // sensitive data on emitted spans.
  // See https://github.com/Arize-ai/openinference/blob/main/js/packages/openinference-core
  traceConfig: {
    hideInputs: true,
    hideOutputs: true,
  },

  // Optional: maximum number of root trace spans kept in memory while traces
  // are in flight. If the limit is exceeded, the oldest root span is ended
  // and evicted. Defaults to 1000.
  maxRootSpansInFlight: 1000,
});

宽幅覆盖

每个代理SDK跨度类型被映射到 OpenInference 跨度类型:

两者 聊天 补全 财务报告和审定财务报表 答复 运输得到支持。 在聊天 补全模式中, SDK 将原始响应对象存储在 output[]; 此仪器从 choices[].message 并累积所有回复的符号使用.

实例

cd js/packages/openinference-instrumentation-openai-agents
pnpm install
pnpm -r build

OPENAI_API_KEY=sk-... npx tsx examples/chat.ts     # single agent + tool call
OPENAI_API_KEY=sk-... npx tsx examples/handoff.ts  # multi-agent handoff
OPENAI_API_KEY=sk-... npx tsx examples/streaming.ts # streamed agent run
OPENAI_API_KEY=sk-... npx tsx examples/guardrail.ts # input guardrail

共享的 OTel 设置在 examples/instrumentation.ts 修改它,以便在OTLP出口商或任何其他处理器中互换.

许可证

阿帕奇 - 2.0

快速缓存令牌计数

OpenAI 缓存读取( O)cached_tokens写作(cache_write_tokens)记录为 llm.token_count.prompt_details.cache_read 财务报告和审定财务报表 cache_write见 语义公约 说明这些分项与及时总数的关系,以及 缓存符号示例 凤凰城的往返旅行.