[8.x] [inference] add support for openAI native stream token count (#…

…200745) (#201007) # Backport This will backport the following commits from `main` to `8.x`: - [[inference] add support for openAI native stream token count (#200745)](#200745)  ### Questions ? Please refer to the [Backport tool documentation](https://github.com/sqren/backport)  Co-authored-by: Pierre Gayvallet <[email protected]>
elastic · Nov 20, 2024 · 63f1de7 · 63f1de7
1 parent 43d4730
commit 63f1de7
Show file tree

Hide file tree

Showing 9 changed files with 395 additions and 138 deletions.
diff --git a/x-pack/plugins/actions/server/lib/get_token_count_from_openai_stream.test.ts b/x-pack/plugins/actions/server/lib/get_token_count_from_openai_stream.test.ts
@@ -61,6 +61,16 @@ describe('getTokenCountFromOpenAIStream', () => {
     ],
   };
 
+  const usageChunk = {
+    object: 'chat.completion.chunk',
+    choices: [],
+    usage: {
+      prompt_tokens: 50,
+      completion_tokens: 100,
+      total_tokens: 150,
+    },
+  };
+
   const PROMPT_TOKEN_COUNT = 36;
   const COMPLETION_TOKEN_COUNT = 5;
 
@@ -70,55 +80,79 @@ describe('getTokenCountFromOpenAIStream', () => {
   });
 
   describe('when a stream completes', () => {
-    beforeEach(async () => {
-      stream.write('data: [DONE]');
-      stream.complete();
-    });
+    describe('with usage chunk', () => {
+      it('returns the counts from the usage chunk', async () => {
+        stream = createStreamMock();
+        stream.write(`data: ${JSON.stringify(chunk)}`);
+        stream.write(`data: ${JSON.stringify(usageChunk)}`);
+        stream.write('data: [DONE]');
+        stream.complete();
 
-    describe('without function tokens', () => {
-      beforeEach(async () => {
         tokens = await getTokenCountFromOpenAIStream({
           responseStream: stream.transform,
           logger,
           body: JSON.stringify(body),
         });
-      });
 
-      it('counts the prompt tokens', () => {
-        expect(tokens.prompt).toBe(PROMPT_TOKEN_COUNT);
-        expect(tokens.completion).toBe(COMPLETION_TOKEN_COUNT);
-        expect(tokens.total).toBe(PROMPT_TOKEN_COUNT + COMPLETION_TOKEN_COUNT);
+        expect(tokens).toEqual({
+          prompt: usageChunk.usage.prompt_tokens,
+          completion: usageChunk.usage.completion_tokens,
+          total: usageChunk.usage.total_tokens,
+        });
       });
     });
 
-    describe('with function tokens', () => {
+    describe('without usage chunk', () => {
       beforeEach(async () => {
-        tokens = await getTokenCountFromOpenAIStream({
-          responseStream: stream.transform,
-          logger,
-          body: JSON.stringify({
-            ...body,
-            functions: [
-              {
-                name: 'my_function',
-                description: 'My function description',
-                parameters: {
-                  type: 'object',
-                  properties: {
-                    my_property: {
-                      type: 'boolean',
-                      description: 'My function property',
+        stream.write('data: [DONE]');
+        stream.complete();
+      });
+
+      describe('without function tokens', () => {
+        beforeEach(async () => {
+          tokens = await getTokenCountFromOpenAIStream({
+            responseStream: stream.transform,
+            logger,
+            body: JSON.stringify(body),
+          });
+        });
+
+        it('counts the prompt tokens', () => {
+          expect(tokens.prompt).toBe(PROMPT_TOKEN_COUNT);
+          expect(tokens.completion).toBe(COMPLETION_TOKEN_COUNT);
+          expect(tokens.total).toBe(PROMPT_TOKEN_COUNT + COMPLETION_TOKEN_COUNT);
+        });
+      });
+
+      describe('with function tokens', () => {
+        beforeEach(async () => {
+          tokens = await getTokenCountFromOpenAIStream({
+            responseStream: stream.transform,
+            logger,
+            body: JSON.stringify({
+              ...body,
+              functions: [
+                {
+                  name: 'my_function',
+                  description: 'My function description',
+                  parameters: {
+                    type: 'object',
+                    properties: {
+                      my_property: {
+                        type: 'boolean',
+                        description: 'My function property',
+                      },
                     },
                   },
                 },
-              },
-            ],
-          }),
+              ],
+            }),
+          });
         });
-      });
 
-      it('counts the function tokens', () => {
-        expect(tokens.prompt).toBeGreaterThan(PROMPT_TOKEN_COUNT);
+        it('counts the function tokens', () => {
+          expect(tokens.prompt).toBeGreaterThan(PROMPT_TOKEN_COUNT);
+        });
       });
     });
   });

diff --git a/x-pack/plugins/actions/server/lib/get_token_count_from_openai_stream.ts b/x-pack/plugins/actions/server/lib/get_token_count_from_openai_stream.ts
@@ -25,44 +25,7 @@ export async function getTokenCountFromOpenAIStream({
   prompt: number;
   completion: number;
 }> {
-  const chatCompletionRequest = JSON.parse(
-    body
-  ) as OpenAI.ChatCompletionCreateParams.ChatCompletionCreateParamsStreaming;
-
-  // per https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
-  const tokensFromMessages = encode(
-    chatCompletionRequest.messages
-      .map(
-        (msg) =>
-          `<|start|>${msg.role}\n${msg.content}\n${
-            'name' in msg
-              ? msg.name
-              : 'function_call' in msg && msg.function_call
-              ? msg.function_call.name + '\n' + msg.function_call.arguments
-              : ''
-          }<|end|>`
-      )
-      .join('\n')
-  ).length;
-
-  // this is an approximation. OpenAI cuts off a function schema
-  // at a certain level of nesting, so their token count might
-  // be lower than what we are calculating here.
-
-  const tokensFromFunctions = chatCompletionRequest.functions
-    ? encode(
-        chatCompletionRequest.functions
-          ?.map(
-            (fn) =>
-              `<|start|>${fn.name}\n${fn.description}\n${JSON.stringify(fn.parameters)}<|end|>`
-          )
-          .join('\n')
-      ).length
-    : 0;
-
-  const promptTokens = tokensFromMessages + tokensFromFunctions;
-
-  let responseBody: string = '';
+  let responseBody = '';
 
   responseStream.on('data', (chunk: string) => {
     responseBody += chunk.toString();
@@ -74,39 +37,64 @@ export async function getTokenCountFromOpenAIStream({
     logger.error('An error occurred while calculating streaming response tokens');
   }
 
-  const response = responseBody
+  let completionUsage: OpenAI.CompletionUsage | undefined;
+
+  const response: ParsedResponse = responseBody
     .split('\n')
     .filter((line) => {
       return line.startsWith('data: ') && !line.endsWith('[DONE]');
     })
     .map((line) => {
       return JSON.parse(line.replace('data: ', ''));
     })
-    .filter(
-      (
-        line
-      ): line is {
-        choices: Array<{
-          delta: { content?: string; function_call?: { name?: string; arguments: string } };
-        }>;
-      } => {
-        return (
-          'object' in line && line.object === 'chat.completion.chunk' && line.choices.length > 0
-        );
-      }
-    )
+    .filter((line): line is OpenAI.ChatCompletionChunk => {
+      return 'object' in line && line.object === 'chat.completion.chunk';
+    })
     .reduce(
       (prev, line) => {
-        const msg = line.choices[0].delta!;
-        prev.content += msg.content || '';
-        prev.function_call.name += msg.function_call?.name || '';
-        prev.function_call.arguments += msg.function_call?.arguments || '';
+        if (line.usage) {
+          completionUsage = line.usage;
+        }
+        if (line.choices?.length) {
+          const msg = line.choices[0].delta!;
+          prev.content += msg.content || '';
+          prev.function_call.name += msg.function_call?.name || '';
+          prev.function_call.arguments += msg.function_call?.arguments || '';
+        }
         return prev;
       },
       { content: '', function_call: { name: '', arguments: '' } }
     );
 
-  const completionTokens = encode(
+  // not all openAI compatible providers emit completion chunk, so we still have to support
+  // manually counting the tokens
+  if (completionUsage) {
+    return {
+      prompt: completionUsage.prompt_tokens,
+      completion: completionUsage.completion_tokens,
+      total: completionUsage.total_tokens,
+    };
+  } else {
+    const promptTokens = manuallyCountPromptTokens(body);
+    const completionTokens = manuallyCountCompletionTokens(response);
+    return {
+      prompt: promptTokens,
+      completion: completionTokens,
+      total: promptTokens + completionTokens,
+    };
+  }
+}
+
+interface ParsedResponse {
+  content: string;
+  function_call: {
+    name: string;
+    arguments: string;
+  };
+}
+
+const manuallyCountCompletionTokens = (response: ParsedResponse) => {
+  return encode(
     JSON.stringify(
       omitBy(
         {
@@ -117,10 +105,42 @@ export async function getTokenCountFromOpenAIStream({
       )
     )
   ).length;
+};
 
-  return {
-    prompt: promptTokens,
-    completion: completionTokens,
-    total: promptTokens + completionTokens,
-  };
-}
+const manuallyCountPromptTokens = (requestBody: string) => {
+  const chatCompletionRequest: OpenAI.ChatCompletionCreateParams.ChatCompletionCreateParamsStreaming =
+    JSON.parse(requestBody);
+
+  // per https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
+  const tokensFromMessages = encode(
+    chatCompletionRequest.messages
+      .map(
+        (msg) =>
+          `<|start|>${msg.role}\n${msg.content}\n${
+            'name' in msg
+              ? msg.name
+              : 'function_call' in msg && msg.function_call
+              ? msg.function_call.name + '\n' + msg.function_call.arguments
+              : ''
+          }<|end|>`
+      )
+      .join('\n')
+  ).length;
+
+  // this is an approximation. OpenAI cuts off a function schema
+  // at a certain level of nesting, so their token count might
+  // be lower than what we are calculating here.
+
+  const tokensFromFunctions = chatCompletionRequest.functions
+    ? encode(
+        chatCompletionRequest.functions
+          ?.map(
+            (fn) =>
+              `<|start|>${fn.name}\n${fn.description}\n${JSON.stringify(fn.parameters)}<|end|>`
+          )
+          .join('\n')
+      ).length
+    : 0;
+
+  return tokensFromMessages + tokensFromFunctions;
+};