Gpt 4 refactor (#71)

* refactor: tokenCount counts array of strings instead of conversation * refactor: logic to decide when to use GPT4 --------- Co-authored-by: Lauris Skraucis <[email protected]>
getadaptly · Aug 15, 2023 · 841279c · 841279c
1 parent bebc683
commit 841279c
Show file tree

Hide file tree

Showing 4 changed files with 27 additions and 9 deletions.
diff --git a/.../issue_handler/actions/created/commands/go/source/breaking-changes/findBreakingChanges.ts b/.../issue_handler/actions/created/commands/go/source/breaking-changes/findBreakingChanges.ts
@@ -10,6 +10,7 @@ import { chatCompletion } from '@adaptly/services/openai/utils/chatCompletion';
 import { getMessageContent } from '@adaptly/services/openai/utils/getMessageContent';
 import { tokenCount } from '@adaptly/services/openai/utils/tokenCount';
 import { GPT35_MODEL, GPT4_MODEL, MAX_NUM_TOKENS_8K } from '@adaptly/services/openai/client';
+import { getConversationContents } from '@adaptly/services/openai/utils/getConversationContents';
 
 export type BreakingChanges = {
     cursorVersion: string;
@@ -121,7 +122,8 @@ async function extractBreakingChanges(packageName: string, cursorVersion: string
     let completionData: CreateChatCompletionResponse;
 
     try {
-        const firstCheckModel = tokenCount(breakingChangesConversation, GPT4_MODEL) < MAX_NUM_TOKENS_8K ? GPT4_MODEL : GPT35_MODEL;
+        const conversationContents = getConversationContents(breakingChangesConversation);
+        const firstCheckModel = tokenCount([...conversationContents, changelog], GPT4_MODEL) < MAX_NUM_TOKENS_8K ? GPT4_MODEL : GPT35_MODEL;
         // here we need to specify function to have good JSON reply structure
         const completion = await chatCompletion(breakingChangesConversation, firstCheckModel);
         Logger.info('ChatGPT: Breaking changes extracted', { packageName, cursorVersion, breakingChanges: completion.data.choices });
@@ -133,12 +135,20 @@ async function extractBreakingChanges(packageName: string, cursorVersion: string
     let breakingChanges = getBreakingChangesFromChatCompletion(completionData);
 
     if (breakingChanges.length > 0) {
+        const modelMessage = completionData.choices[0].message;
+
+        if (modelMessage) {
+            breakingChangesConversation.push(modelMessage);
+        }
+
         breakingChangesConversation.push({
             role: RoleSystem,
             content: `${breakingChangesDoubleCheckPrompt}`
         });
 
-        const secondCheckModel = tokenCount(breakingChangesConversation, GPT4_MODEL) < MAX_NUM_TOKENS_8K ? GPT4_MODEL : GPT35_MODEL;
+        const nextConversationContents = getConversationContents(breakingChangesConversation);
+        const secondCheckModel =
+            tokenCount([...nextConversationContents, JSON.stringify(completionData)], GPT4_MODEL) < MAX_NUM_TOKENS_8K ? GPT4_MODEL : GPT35_MODEL;
 
         const completion = await chatCompletion(breakingChangesConversation, secondCheckModel);
         Logger.info('Double check ChatGPT: Breaking changes extracted', { packageName, cursorVersion, breakingChanges: completion.data.choices });

diff --git a/source/services/openai/utils/getConversationContents.ts b/source/services/openai/utils/getConversationContents.ts
@@ -0,0 +1,7 @@
+import { ChatCompletionRequestMessage } from 'openai';
+
+export function getConversationContents(conversation: ChatCompletionRequestMessage[]): string[] {
+    const messages = conversation.map((message) => message.content);
+
+    return messages.filter((message) => message !== undefined) as string[];
+}
diff --git a/source/services/openai/utils/tokenCount.test.ts b/source/services/openai/utils/tokenCount.test.ts
@@ -1,4 +1,5 @@
 import { RoleSystem } from '../types';
+import { getConversationContents } from './getConversationContents';
 import { tokenCount } from './tokenCount';
 import { ChatCompletionRequestMessage } from 'openai';
 
@@ -19,7 +20,8 @@ describe('tokenCount', () => {
             }
         ];
 
-        const result = tokenCount(conversation, 'gpt-3.5-turbo-16k-0613');
+        const conversationContents = getConversationContents(conversation);
+        const result = tokenCount(conversationContents, 'gpt-3.5-turbo-16k-0613');
 
         // Replace with expected token count for the given conversation and model
         const expectedTokenCount = conversation.reduce((acc, message) => acc + message.tokenCount, 0);
@@ -43,7 +45,8 @@ describe('tokenCount', () => {
             }
         ];
 
-        const result = tokenCount(conversation, 'gpt-4-0613');
+        const conversationContents = getConversationContents(conversation);
+        const result = tokenCount(conversationContents, 'gpt-4-0613');
 
         // Replace with expected token count for the given conversation and model
         const expectedTokenCount = conversation.reduce((acc, message) => acc + message.tokenCount, 0);

diff --git a/source/services/openai/utils/tokenCount.ts b/source/services/openai/utils/tokenCount.ts
@@ -5,15 +5,13 @@ import { ChatCompletionRequestMessage } from 'openai';
 const enc4 = encoding_for_model('gpt-4');
 const enc35 = encoding_for_model('gpt-3.5-turbo');
 
-export function tokenCount(conversation: ChatCompletionRequestMessage[], model: 'gpt-3.5-turbo-16k-0613' | 'gpt-4-0613'): number {
+export function tokenCount(texts: string[], model: 'gpt-3.5-turbo-16k-0613' | 'gpt-4-0613'): number {
     let length = 0;
 
     const enc = model === 'gpt-3.5-turbo-16k-0613' ? enc35 : enc4;
 
-    for (let message of conversation) {
-        if (message.content) {
-            length += enc.encode(message.content).length;
-        }
+    for (let text of texts) {
+        length += enc.encode(text).length;
     }
 
     return length;