chore: Updated OpenAI instrumentation to skip creating a LlmChatCompletionMessage for an outgoing tool call response (#3655)

Author: amychisholm03

lib/llm-events/openai/chat-completion-message.js

@@ -26,7 +26,7 @@ module.exports = class LlmChatCompletionMessage extends LlmEvent {
     // Check if the given message is from the response.
     // The response object differs based on the API called.
     // If it's `responses.create`, we check against `response.output`.
-    // If it's `chat.completions.create` or langchain, we check against `response.choices`.
+    // If it's `chat.completions.create`, we check against `response.choices`.
     if (response?.object === 'response') {
       this.is_response = message.content === response?.output?.[0]?.content?.[0]?.text
     } else {

lib/subscribers/openai/utils.js

@@ -21,7 +21,8 @@ let TRACKING_METRIC = OPENAI.TRACKING_PREFIX
  * Parses the response from OpenAI and extracts the message content and role.
  *
  * @param {object} response The OpenAI SDK response object
- * @returns {{ content: string, role: string }} the message object with fields `content` and `role`
+ * @returns {object[]} an array with the message object with fields `content`
+ * and `role` or [] if response invalid
  */
 function getMessageFromResponse(response) {
   let content
@@ -30,11 +31,17 @@ function getMessageFromResponse(response) {
     content = response?.output?.[0]?.content?.[0]?.text
     role = response?.output?.[0]?.role
   } else {
-    content = response?.choices?.[0]?.message?.content
-    role = response?.choices?.[0]?.message?.role
+    const choice = response?.choices?.[0]
+    if (choice?.finish_reason === 'tool_calls') {
+      // A false response. Don't create a LlmChatCompletionMessage
+      // for this -- the full conversation isn't done yet.
+      return []
+    }
+    content = choice?.message?.content
+    role = choice?.message?.role
   }
 
-  return { content, role }
+  return [{ content, role }]
 }
 
 /**
@@ -165,7 +172,7 @@ function recordChatCompletionMessages({
   // the response message.
   const messages = [
     ...getMessagesFromRequest(request, logger),
-    getMessageFromResponse(response)
+    ...getMessageFromResponse(response)
   ]
 
   for (let i = 0; i < messages.length; i++) {

test/versioned/openai/chat-completions.test.js

@@ -155,6 +155,54 @@ test('chat.completions.create', async (t) => {
     })
   })
 
+  await t.test('should have not create LlmChatCompletionMessage for an outgoing tool call', (t, end) => {
+    const { client, agent } = t.nr
+    helper.runInTransaction(agent, async(tx) => {
+      const model = 'gpt-4'
+      const content = 'What is 2 + 2?'
+      await client.chat.completions.create({
+        max_tokens: 100,
+        temperature: 0.5,
+        model,
+        messages: [
+          { role: 'user', content },
+        ]
+      })
+
+      const chatCmplId = 'chatcmpl-87sb95K4EF2nuJRcTs43Tm9calc'
+
+      const events = agent.customEventAggregator.events.toArray()
+      assert.equal(events.length, 2, 'should create only 1 chat completion messages and 1 summary event')
+      const chatMsgs = events.filter(([{ type }]) => type === 'LlmChatCompletionMessage')
+      assert.equal(chatMsgs.length, 1, 'should only have the request message')
+
+      const [segment] = tx.trace.getChildren(tx.trace.root.id)
+
+      const requestMsg = chatMsgs[0]
+      if (requestMsg[1].sequence === 0) {
+        const expectedMsg = {
+          appName: 'New Relic for Node.js tests',
+          completion_id: /[a-f0-9]{36}/,
+          content,
+          id: `${chatCmplId}-0`,
+          ingest_source: 'Node',
+          is_response: false,
+          request_id: '49dbbffbd3c3f4612aa48def69059calc',
+          'response.model': model,
+          role: 'user',
+          sequence: 0,
+          span_id: segment.id,
+          trace_id: tx.traceId,
+          vendor: 'openai',
+        }
+        match(requestMsg[1], expectedMsg, { assert })
+      }
+
+      tx.end()
+      end()
+    })
+  })
+
   if (semver.gte(pkgVersion, '4.12.2')) {
     await t.test('should create span on successful chat completion stream create', (t, end) => {
       const { client, agent, host, port } = t.nr

test/versioned/openai/mock-chat-api-responses.js

@@ -299,3 +299,48 @@ responses.set('bad stream', {
   },
   streamData: 'do random'
 })
+
+responses.set('What is 2 + 2?', {
+  headers: {
+    'Content-Type': 'application/json',
+    'openai-model': 'gpt-4',
+    'openai-organization': 'new-relic-nkmd8b',
+    'openai-processing-ms': '1469',
+    'openai-version': '2020-10-01',
+    'x-ratelimit-limit-requests': '200',
+    'x-ratelimit-limit-tokens': '40000',
+    'x-ratelimit-remaining-requests': '199',
+    'x-ratelimit-remaining-tokens': '39940',
+    'x-ratelimit-reset-requests': '7m12s',
+    'x-ratelimit-reset-tokens': '90ms',
+    'x-request-id': '49dbbffbd3c3f4612aa48def69059calc'
+  },
+  code: 200,
+  body: {
+    choices: [
+      {
+        finish_reason: 'tool_calls',
+        index: 0,
+        message: {
+          role: 'assistant',
+          content: null,
+          tool_calls: [
+            {
+              id: 'call_calc123',
+              type: 'function',
+              function: {
+                name: 'calculator',
+                arguments: '{"a":2,"b":2,"operation":"add"}'
+              }
+            }
+          ]
+        }
+      }
+    ],
+    created: 1696888863,
+    id: 'chatcmpl-87sb95K4EF2nuJRcTs43Tm9calc',
+    model: 'gpt-4',
+    object: 'chat.completion',
+    usage: { completion_tokens: 15, prompt_tokens: 50, total_tokens: 65 }
+  }
+})