@arizeai/openinference-instrumentation-openai 2.2.0 → 2.3.1

package/src/instrumentation.ts

@@ -45,6 +45,18 @@ import {
   safelyJSONStringify,
   TraceConfigOptions,
 } from "@arizeai/openinference-core";
+import {
+  ResponseCreateParamsBase,
+  ResponseStreamEvent,
+  Response as ResponseType,
+} from "openai/resources/responses/responses";
+
+import {
+  consumeResponseStreamEvents,
+  getResponsesInputMessagesAttributes,
+  getResponsesOutputMessagesAttributes,
+  getResponsesUsageAttributes,
+} from "./responsesAttributes";
 
 const MODULE_NAME = "openai";
 
@@ -141,10 +153,10 @@ export class OpenAIInstrumentation extends InstrumentationBase<typeof openai> {
     // eslint-disable-next-line @typescript-eslint/no-this-alias
     const instrumentation: OpenAIInstrumentation = this;
 
+    // Patch create chat completions
     type ChatCompletionCreateType =
       typeof module.OpenAI.Chat.Completions.prototype.create;
 
-    // Patch create chat completions
     this._wrap(
       module.OpenAI.Chat.Completions.prototype,
       "create",
@@ -367,6 +379,102 @@ export class OpenAIInstrumentation extends InstrumentationBase<typeof openai> {
       },
     );
 
+    // Patch responses (if the patched module contains the Responses interface)
+    if (module.OpenAI.Responses) {
+      type ResponsesCreateType =
+        typeof module.OpenAI.Responses.prototype.create;
+
+      this._wrap(
+        module.OpenAI.Responses.prototype,
+        "create",
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        (original: ResponsesCreateType): any => {
+          return function patchedCreate(
+            this: unknown,
+            ...args: Parameters<ResponsesCreateType>
+          ) {
+            const body = args[0];
+            const { input: _messages, ...invocationParameters } = body;
+            const span = instrumentation.oiTracer.startSpan(
+              `OpenAI Responses`,
+              {
+                kind: SpanKind.INTERNAL,
+                attributes: {
+                  [SemanticConventions.OPENINFERENCE_SPAN_KIND]:
+                    OpenInferenceSpanKind.LLM,
+                  [SemanticConventions.LLM_MODEL_NAME]: body.model,
+                  [SemanticConventions.INPUT_VALUE]: JSON.stringify(body),
+                  [SemanticConventions.INPUT_MIME_TYPE]: MimeType.JSON,
+                  [SemanticConventions.LLM_INVOCATION_PARAMETERS]:
+                    JSON.stringify(invocationParameters),
+                  [SemanticConventions.LLM_SYSTEM]: LLMSystem.OPENAI,
+                  [SemanticConventions.LLM_PROVIDER]: LLMProvider.OPENAI,
+                  ...getResponsesInputMessagesAttributes(body),
+                  ...getLLMToolsJSONSchema(body),
+                },
+              },
+            );
+            const execContext = getExecContext(span);
+            const execPromise = safeExecuteInTheMiddle<
+              ReturnType<ResponsesCreateType>
+            >(
+              () => {
+                return context.with(trace.setSpan(execContext, span), () => {
+                  return original.apply(this, args);
+                });
+              },
+              (error) => {
+                // Push the error to the span
+                if (error) {
+                  span.recordException(error);
+                  span.setStatus({
+                    code: SpanStatusCode.ERROR,
+                    message: error.message,
+                  });
+                  span.end();
+                }
+              },
+            );
+            const wrappedPromise = execPromise.then((result) => {
+              const recordSpan = (result?: ResponseType) => {
+                if (!result) {
+                  span.setStatus({ code: SpanStatusCode.ERROR });
+                  span.end();
+                  return;
+                }
+                span.setAttributes({
+                  [SemanticConventions.OUTPUT_VALUE]: JSON.stringify(result),
+                  [SemanticConventions.OUTPUT_MIME_TYPE]: MimeType.JSON,
+                  // Override the model from the value sent by the server
+                  [SemanticConventions.LLM_MODEL_NAME]: result.model,
+                  ...getResponsesOutputMessagesAttributes(result),
+                  ...getResponsesUsageAttributes(result),
+                });
+                span.setStatus({ code: SpanStatusCode.OK });
+                span.end();
+              };
+              if (isResponseCreateResponse(result)) {
+                // Record the results, as we have the final result
+                recordSpan(result);
+              } else {
+                // This is a streaming response
+                // First split the stream via tee
+                const [leftStream, rightStream] = result.tee();
+                // take the right stream, consuming it and then recording the final chunk
+                // into the span
+                consumeResponseStreamEvents(rightStream).then(recordSpan);
+                // give the left stream back to the caller
+                result = leftStream;
+              }
+
+              return result;
+            });
+            return context.bind(execContext, wrappedPromise);
+          };
+        },
+      );
+    }
+
     _isOpenInferencePatched = true;
     try {
       // This can fail if the module is made immutable via the runtime or bundler
@@ -399,6 +507,12 @@ export class OpenAIInstrumentation extends InstrumentationBase<typeof openai> {
   }
 }
 
+function isResponseCreateResponse(
+  response: Stream<ResponseStreamEvent> | ResponseType,
+): response is ResponseType {
+  return "object" in response && response.object === "response";
+}
+
 /**
  * type-guard that checks if the response is a chat completion response
  */
@@ -449,7 +563,7 @@ function getLLMInputMessagesAttributes(
  * Converts each tool definition into a json schema
  */
 function getLLMToolsJSONSchema(
-  body: ChatCompletionCreateParamsBase,
+  body: ChatCompletionCreateParamsBase | ResponseCreateParamsBase,
 ): Attributes {
   if (!body.tools) {
     // If tools is undefined, return an empty object

package/src/responsesAttributes.ts

@@ -0,0 +1,286 @@
+import { SemanticConventions } from "@arizeai/openinference-semantic-conventions";
+import { Attributes } from "@opentelemetry/api";
+import {
+  ResponseCreateParamsBase,
+  ResponseInputItem,
+  ResponseOutputItem,
+  ResponseStreamEvent,
+  Response as ResponseType,
+} from "openai/resources/responses/responses";
+import { Stream } from "openai/streaming";
+
+/**
+ * Get attributes for responses api Items that are not typical messages with role
+ * @param item - The item to get attributes for
+ * @param prefix - The prefix to use for the attributes
+ * @returns The attributes for the item
+ */
+function getResponseItemAttributes(
+  item: Exclude<ResponseInputItem | ResponseOutputItem, { role: string }>,
+  prefix = "",
+): Attributes {
+  const attributes: Attributes = {};
+  // all items that are not typical messages with role
+  // things like images, files, etc.
+  const toolCallPrefix = `${prefix}${SemanticConventions.MESSAGE_TOOL_CALLS}.0.`;
+
+  switch (item.type) {
+    case "function_call": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "assistant";
+      // now using tool call prefix to simulate multiple tool calls per message
+      attributes[`${toolCallPrefix}${SemanticConventions.TOOL_CALL_ID}`] =
+        item.call_id;
+      attributes[
+        `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_NAME}`
+      ] = item.name;
+      attributes[
+        `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_ARGUMENTS_JSON}`
+      ] = item.arguments;
+      break;
+    }
+    case "function_call_output": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "tool";
+      attributes[`${prefix}${SemanticConventions.MESSAGE_TOOL_CALL_ID}`] =
+        item.call_id;
+      attributes[`${prefix}${SemanticConventions.MESSAGE_CONTENT}`] =
+        item.output;
+      break;
+    }
+    case "reasoning": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "assistant";
+      item.summary.forEach((summaryItem, index) => {
+        const summaryItemPrefix = `${prefix}${SemanticConventions.MESSAGE_CONTENTS}.${index}.`;
+        if (summaryItem.type === "summary_text") {
+          attributes[
+            `${summaryItemPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`
+          ] = "summary_text";
+          attributes[
+            `${summaryItemPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`
+          ] = summaryItem.text;
+        }
+      });
+      break;
+    }
+    case "item_reference": {
+      break;
+    }
+    case "file_search_call": {
+      if (!item.results) {
+        // its a tool call
+        attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] =
+          "assistant";
+        // now using tool call prefix to simulate multiple tool calls per message
+        attributes[`${toolCallPrefix}${SemanticConventions.TOOL_CALL_ID}`] =
+          item.id;
+        attributes[
+          `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_NAME}`
+        ] = item.type;
+        attributes[
+          `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_ARGUMENTS_JSON}`
+        ] = JSON.stringify(item.queries);
+      } else {
+        // its a tool call output
+        attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "tool";
+        attributes[`${prefix}${SemanticConventions.MESSAGE_TOOL_CALL_ID}`] =
+          item.id;
+        attributes[`${prefix}${SemanticConventions.MESSAGE_CONTENT}`] =
+          JSON.stringify(item.results);
+      }
+      break;
+    }
+    case "computer_call": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "assistant";
+      // now using tool call prefix to simulate multiple tool calls per message
+      attributes[`${toolCallPrefix}${SemanticConventions.TOOL_CALL_ID}`] =
+        item.id;
+      attributes[
+        `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_NAME}`
+      ] = item.type;
+      attributes[
+        `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_ARGUMENTS_JSON}`
+      ] = JSON.stringify(item.action);
+      break;
+    }
+    case "computer_call_output": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "tool";
+      attributes[`${prefix}${SemanticConventions.MESSAGE_TOOL_CALL_ID}`] =
+        item.call_id;
+      attributes[`${prefix}${SemanticConventions.MESSAGE_CONTENT}`] =
+        JSON.stringify(item.output);
+      break;
+    }
+    case "web_search_call": {
+      attributes[`${prefix}${SemanticConventions.MESSAGE_ROLE}`] = "assistant";
+      // now using tool call prefix to simulate multiple tool calls per message
+      attributes[`${toolCallPrefix}${SemanticConventions.TOOL_CALL_ID}`] =
+        item.id;
+      attributes[
+        `${toolCallPrefix}${SemanticConventions.TOOL_CALL_FUNCTION_NAME}`
+      ] = item.type;
+      // web search call does not share its arguments with the caller
+      // it will show "undefined" in the arguments when traced
+      break;
+    }
+  }
+  return attributes;
+}
+
+/**
+ * Get attributes for responses api Item, input or output
+ * Non message items are detected and handled by {@link getResponseItemAttributes}
+ * @param itemMessage - The message item to get attributes for
+ * @param prefix - The prefix to use for the attributes
+ * @returns The attributes for the message item
+ */
+function getResponseItemMessageAttributes(
+  itemMessage: ResponseInputItem | ResponseOutputItem,
+  prefix = "",
+): Attributes {
+  const message =
+    typeof itemMessage === "string"
+      ? ({ content: itemMessage, role: "user" } satisfies ResponseInputItem)
+      : itemMessage;
+  if (!("role" in message)) {
+    return getResponseItemAttributes(message, prefix);
+  }
+  const role = message.role;
+  const attributes: Attributes = {
+    [`${prefix}${SemanticConventions.MESSAGE_ROLE}`]: role,
+  };
+  // add contents from message
+  if (typeof message.content === "string") {
+    attributes[`${prefix}${SemanticConventions.MESSAGE_CONTENT}`] =
+      message.content;
+  } else if (Array.isArray(message.content)) {
+    message.content.forEach((part, index) => {
+      const contentsIndexPrefix = `${prefix}${SemanticConventions.MESSAGE_CONTENTS}.${index}.`;
+      if (part.type === "input_text") {
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`
+        ] = "input_text";
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`
+        ] = part.text;
+      } else if (part.type === "output_text") {
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`
+        ] = "output_text";
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`
+        ] = part.text;
+      } else if (part.type === "input_image") {
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`
+        ] = "input_image";
+        if (part.image_url) {
+          attributes[
+            `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_IMAGE}.${SemanticConventions.IMAGE_URL}`
+          ] = part.image_url;
+        }
+      } else if (part.type === "input_file") {
+        // TODO: Handle input file
+      } else if (part.type === "refusal") {
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TYPE}`
+        ] = "refusal";
+        attributes[
+          `${contentsIndexPrefix}${SemanticConventions.MESSAGE_CONTENT_TEXT}`
+        ] = part.refusal;
+      }
+    });
+  }
+  // add role based fields to attributes
+  // as of now, there are no role based fields as they are handled by alternatively typed
+  // items
+  switch (message.role) {
+    case "user": {
+      break;
+    }
+    case "assistant": {
+      break;
+    }
+    case "system": {
+      break;
+    }
+    case "developer": {
+      break;
+    }
+  }
+  return attributes;
+}
+
+export function getResponsesInputMessagesAttributes(
+  body: ResponseCreateParamsBase,
+): Attributes {
+  const attributes: Attributes = {};
+  const items: ResponseInputItem[] = [];
+  if (body.instructions) {
+    items.push({ content: body.instructions, role: "system" });
+  }
+  if (typeof body.input === "string") {
+    items.push({ content: body.input, role: "user" });
+  } else {
+    items.push(...body.input);
+  }
+  items.forEach((item, index) => {
+    const indexPrefix = `${SemanticConventions.LLM_INPUT_MESSAGES}.${index}.`;
+    const itemAttributes = getResponseItemMessageAttributes(item, indexPrefix);
+    Object.entries(itemAttributes).forEach(([key, value]) => {
+      attributes[key] = value;
+    });
+  });
+
+  return attributes;
+}
+
+export function getResponsesUsageAttributes(
+  response: ResponseType,
+): Attributes {
+  if (response.usage) {
+    return {
+      [SemanticConventions.LLM_TOKEN_COUNT_COMPLETION]:
+        response.usage.output_tokens,
+      [SemanticConventions.LLM_TOKEN_COUNT_PROMPT]: response.usage.input_tokens,
+      [SemanticConventions.LLM_TOKEN_COUNT_TOTAL]: response.usage.total_tokens,
+      [SemanticConventions.LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_READ]:
+        response.usage.input_tokens_details?.cached_tokens,
+      [SemanticConventions.LLM_TOKEN_COUNT_COMPLETION_DETAILS_REASONING]:
+        response.usage.output_tokens_details?.reasoning_tokens,
+      // no audio tokens for response inputs or outputs
+    };
+  }
+  return {};
+}
+
+export function getResponsesOutputMessagesAttributes(
+  response: ResponseType,
+): Attributes {
+  const attributes: Attributes = {};
+  const items = response.output;
+  items.forEach((item, index) => {
+    const indexPrefix = `${SemanticConventions.LLM_OUTPUT_MESSAGES}.${index}.`;
+    const itemAttributes = getResponseItemMessageAttributes(item, indexPrefix);
+    Object.entries(itemAttributes).forEach(([key, value]) => {
+      attributes[key] = value;
+    });
+  });
+
+  return attributes;
+}
+
+export async function consumeResponseStreamEvents(
+  stream: Stream<ResponseStreamEvent>,
+): Promise<ResponseType | undefined> {
+  let response: ResponseType | undefined;
+
+  for await (const event of stream) {
+    switch (event.type) {
+      case "response.completed": {
+        response = event.response;
+        break;
+      }
+    }
+  }
+
+  return response;
+}

package/src/version.ts

@@ -1,2 +1,2 @@
 // this is an auto-generated file. see js/scripts/version-update.js
-export const VERSION = "2.2.0";
+export const VERSION = "2.3.1";