Skip to content

Commit a224543

Browse files
aigul-ermakaigul-ermak
committed
refactor: Refactor Gemini AI service with retry logic and JSDoc comments
1 parent f772e03 commit a224543

File tree

1 file changed

+103
-96
lines changed

1 file changed

+103
-96
lines changed

‎backend/src/services/geminiService.ts‎

Lines changed: 103 additions & 96 deletions
Original file line numberDiff line numberDiff line change
@@ -6,121 +6,128 @@ import {AudioRequest} from "../types/audioTypes.js";
66
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);
77

88
/**
9-
* Generates a response from the Gemini AI model based on the provided prompt text.
9+
* List of Gemini AI model versions to attempt in order of preference.
1010
*
11-
* @param {string} promptText - The text prompt to send to the Gemini AI model.
12-
* @returns {Promise<GeminiResponseType>} - A promise that resolves to an object containing the generated text and its summary.
11+
* The system will try each model sequentially until a successful response is received,
12+
* starting with the fastest and latest versions and falling back to older or alternative models if necessary.
13+
*
14+
* This ordering is designed to maximize reliability and performance.
1315
*/
14-
export const generateGeminiResponse = async (promptText: string): Promise<GeminiResponseType> => {
15-
// try {
16-
17-
const modelVersions = [
18-
"gemini-2.0-flash",
19-
"gemini-2.0-flash-001",
20-
"gemini-2.0-flash-lite",
21-
"gemini-1.5-flash",
22-
"gemini-1.5-flash-8b",
23-
"gemini-1.5-pro"
24-
];
25-
26-
for (const modelVersion of modelVersions) {
27-
try {
28-
const model = genAI.getGenerativeModel({model: modelVersion});
29-
30-
const result = await model.generateContent({
31-
contents: [
32-
{
33-
role: "user",
34-
parts: [{text: promptText}],
35-
},
36-
],
37-
});
38-
39-
const text = stripMarkdown(result.response.text());
40-
41-
const summaryResult = await model.generateContent({
42-
contents: [{
43-
role: "user",
44-
parts: [{
45-
text: `
46-
You are a smart assistant. Based on the interaction below, summarize it in no more than 35 words.
47-
Avoid using phrases like "the user" or "AI." Write a brief, neutral description of the request and response.
48-
49-
Prompt:
50-
${promptText}
51-
52-
Response:
53-
${text}
54-
`.trim()
55-
}],
56-
}],
57-
});
58-
59-
const summary = summaryResult.response.text().trim();
60-
61-
return {text, summary};
16+
const MODEL_VERSIONS = [
17+
"gemini-2.0-flash",
18+
"gemini-2.0-flash-001",
19+
"gemini-2.0-flash-lite",
20+
"gemini-1.5-flash",
21+
"gemini-1.5-flash-8b",
22+
"gemini-1.5-pro",
23+
];
6224

25+
/**
26+
* Attempts to generate content from Gemini AI by trying multiple model versions in order.
27+
*
28+
* Iterates through the predefined list of model versions, sending the provided prompt to each model.
29+
* If a model responds successfully, returns the stripped (plain text) result.
30+
* If a model fails with a 500 or 503 error, automatically retries with the next model.
31+
* Throws a fatal error immediately for other types of failures, or if all models fail.
32+
*
33+
* @param {any} prompt - The prompt payload to send to the AI model.
34+
* @returns {Promise<string>} - The generated plain text content.
35+
* @throws {Error} - Throws if all models fail or a non-retryable error occurs.
36+
*/
37+
async function tryGenerateContent(
38+
prompt: any,
39+
): Promise<string> {
40+
for (const modelVersion of MODEL_VERSIONS) {
41+
try {
42+
const model = genAI.getGenerativeModel({ model: modelVersion });
43+
const result = await model.generateContent({ contents: [prompt] });
44+
return stripMarkdown(result.response.text());
6345
} catch (error: any) {
6446
if (error?.status === 500 || error?.status === 503) {
6547
console.warn(`Model ${modelVersion} failed with status ${error.status}. Trying next...`);
66-
continue; // Try next model
48+
continue;
6749
}
68-
69-
console.error("Gemini Fatal Error:", error);
50+
console.error(`Gemini Fatal Error [${modelVersion}]:`, error);
7051
throw error;
7152
}
72-
7353
}
74-
7554
throw new Error("All Gemini models failed.");
55+
}
56+
57+
58+
/**
59+
* Generates a detailed AI response and a concise summary based on a given prompt text.
60+
*
61+
* Constructs two prompts:
62+
* - One to generate a full response from Gemini AI.
63+
* - Another to request a summarized version of the interaction in no more than 35 words,
64+
* avoiding references to "user" or "AI."
65+
*
66+
* Sends both prompts through the model retry mechanism to ensure robustness against failures.
67+
*
68+
* @param {string} promptText - The input text used to generate the AI response.
69+
* @returns {Promise<GeminiResponseType>} - An object containing both the generated text and its summary.
70+
* @throws {Error} - Throws if all model versions fail for either the full response or the summary.
71+
*/
72+
export const generateGeminiResponse = async (promptText: string): Promise<GeminiResponseType> => {
73+
const prompt = {
74+
role: "user",
75+
parts: [{ text: promptText }],
76+
};
7677

78+
const text = await tryGenerateContent(prompt);
79+
80+
const summaryPrompt = {
81+
role: "user",
82+
parts: [{
83+
text: `
84+
You are a smart assistant. Based on the interaction below, summarize it in no more than 35 words.
85+
Avoid using phrases like "the user" or "AI." Write a brief, neutral description of the request and response.
86+
87+
Prompt:
88+
${promptText}
89+
90+
Response:
91+
${text}
92+
`.trim(),
93+
}],
94+
};
95+
96+
const summary = await tryGenerateContent(summaryPrompt);
97+
98+
return { text, summary };
7799
};
78100

101+
/**
102+
* Generates a transcript from the provided audio input using the Gemini AI model.
103+
*
104+
* Converts the audio buffer to a base64-encoded string, builds a prompt requesting a transcription,
105+
* and attempts to generate a transcript by sending the prompt to Gemini AI.
106+
* Automatically retries across multiple model versions if initial attempts fail.
107+
*
108+
* @param {AudioRequest} param0 - An object containing the audio buffer and its MIME type.
109+
* @returns {Promise<string>} - The generated transcript as plain text.
110+
* @throws {Error} - Throws an error if all model attempts fail.
111+
*/
79112
export const generateGeminiAudioResponse = async (
80113
{audioBuffer, mimeType}: AudioRequest
81114
): Promise<string> => {
82-
const modelVersions = [
83-
"gemini-2.0-flash",
84-
"gemini-2.0-flash-001",
85-
"gemini-2.0-flash-lite",
86-
"gemini-1.5-flash",
87-
"gemini-1.5-flash-8b",
88-
"gemini-1.5-pro"
89-
];
90-
91115
const base64Audio = audioBuffer.toString('base64');
92-
for (const modelVersion of modelVersions) {
93-
try {
94116

95-
const model = genAI.getGenerativeModel({model: modelVersion});
96-
97-
const result = await model.generateContent({
98-
contents: [{
99-
role: "user",
100-
parts: [
101-
{text: "Generate a transcript of this audio:"},
102-
{
103-
inlineData: {
104-
data: base64Audio,
105-
mimeType
106-
}
107-
}
108-
]
109-
}]
110-
});
111-
112-
return stripMarkdown(result.response.text());
113-
114-
} catch (error: any) {
115-
if (error?.status === 500 || error?.status === 503) {
116-
console.warn(`Model ${modelVersion} failed with status ${error.status}. Trying next...`);
117-
continue; // Try next model
118-
}
119-
console.error("Gemini Audio Fatal Error:", error);
120-
throw error;
121-
}
122-
}
123-
throw new Error("All Gemini audio models failed.");
117+
const audioPrompt = {
118+
role: "user",
119+
parts: [
120+
{ text: "Generate a transcript of this audio:" },
121+
{
122+
inlineData: {
123+
data: base64Audio,
124+
mimeType,
125+
},
126+
},
127+
],
128+
};
129+
130+
return await tryGenerateContent(audioPrompt);
124131
};
125132

126133

0 commit comments

Comments
 (0)