feat: use new gpt-3.5-turbo model (#123)

Co-authored-by: Hiroki Osame <hiroki.osame@gmail.com>

src/commands/aicommits.ts

@@ -49,6 +49,10 @@ export default async (
 	);
 	s.stop('Changes analyzed');
 
+	if (messages.length === 0) {
+		throw new KnownError('No commit messages were generated. Try again.');
+	}
+
 	let message: string;
 	if (messages.length === 1) {
 		[message] = messages;

src/utils/openai.ts

@@ -1,6 +1,6 @@
 import https from 'https';
 import type { ClientRequest, IncomingMessage } from 'http';
-import type { CreateCompletionRequest, CreateCompletionResponse } from 'openai';
+import type { CreateChatCompletionRequest, CreateChatCompletionResponse } from 'openai';
 import { encoding_for_model as encodingForModel } from '@dqbd/tiktoken';
 import { KnownError } from './error.js';
 
@@ -50,13 +50,13 @@ const httpsPost = async (
 	request.end();
 });
 
-const createCompletion = async (
+const createChatCompletion = async (
 	apiKey: string,
-	json: CreateCompletionRequest,
+	json: CreateChatCompletionRequest,
 ) => {
 	const { response, data } = await httpsPost(
 		'api.openai.com',
-		'/v1/completions',
+		'/v1/chat/completions',
 		{
 			Authorization: `Bearer ${apiKey}`,
 		},
@@ -81,7 +81,7 @@ const createCompletion = async (
 		throw new KnownError(errorMessage);
 	}
 
-	return JSON.parse(data) as CreateCompletionResponse;
+	return JSON.parse(data) as CreateChatCompletionResponse;
 };
 
 const sanitizeMessage = (message: string) => message.trim().replace(/[\n\r]/g, '').replace(/(\w)\.$/, '$1');
@@ -90,8 +90,9 @@ const deduplicateMessages = (array: string[]) => Array.from(new Set(array));
 
 const getPrompt = (locale: string, diff: string) => `Write an insightful but concise Git commit message in a complete sentence in present tense for the following diff without prefacing it with anything, the response must be in the language ${locale}:\n${diff}`;
 
-const model = 'text-davinci-003';
+const model = 'gpt-3.5-turbo';
-const encoder = encodingForModel(model);
+// TODO: update for the new gpt-3.5 model
+const encoder = encodingForModel('text-davinci-003');
 
 export const generateCommitMessage = async (
 	apiKey: string,
@@ -110,9 +111,12 @@ export const generateCommitMessage = async (
 	}
 
 	try {
-		const completion = await createCompletion(apiKey, {
+		const completion = await createChatCompletion(apiKey, {
 			model,
-			prompt,
+			messages: [{
+				role: 'user',
+				content: prompt,
+			}],
 			temperature: 0.7,
 			top_p: 1,
 			frequency_penalty: 0,
@@ -124,7 +128,8 @@ export const generateCommitMessage = async (
 
 		return deduplicateMessages(
 			completion.choices
-				.map(choice => sanitizeMessage(choice.text!)),
+				.filter(choice => choice.message?.content)
+				.map(choice => sanitizeMessage(choice.message!.content)),
 		);
 	} catch (error) {
 		const errorAsAny = error as any;

tests/specs/cli.ts

@@ -3,9 +3,6 @@ import { createFixture } from 'fs-fixture';
 import { createAicommits, createGit } from '../utils.js';
 
 const { OPENAI_KEY } = process.env;
-if (!OPENAI_KEY) {
-	throw new Error('process.env.OPENAI_KEY is necessary to run these tests');
-}
 
 export default testSuite(({ describe }) => {
 	if (process.platform === 'win32') {
@@ -14,6 +11,11 @@ export default testSuite(({ describe }) => {
 		return;
 	}
 
+	if (!OPENAI_KEY) {
+		console.warn('⚠️  process.env.OPENAI_KEY is necessary to run these tests. Skipping...');
+		return;
+	}
+
 	describe('CLI', async ({ test }) => {
 		const data: Record<string, string> = {
 			firstName: 'Hiroki',
@@ -94,10 +96,8 @@ export default testSuite(({ describe }) => {
 				if (stdout.match('└')) {
 					const countChoices = stdout.match(/ {2}[●○]/g)?.length ?? 0;
 
-					// 2 choices or less should be generated
+					// 2 choices should be generated
-					// pretty common for it to return 2 results that are the same
+					expect(countChoices).toBe(2);
-					// which gets de-duplicated
-					expect(countChoices <= 2).toBe(true);
 
 					committing.stdin!.write('\r');
 					committing.stdin!.end();