LLM Engineering — Prompt, RAG, Fine-tuning Cho AI App

const SYSTEM_PROMPT = `Bạn là senior developer hỗ trợ viết code.
Luôn tuân theo:
1. TypeScript, strict mode
2. Không dùng 'any'
3. Unit test kèm theo
4. Giải thích ngắn gọn quyết định`;

async function chat(userMessage: string) {
  const response = await openai.chat.completions.create({
    model: 'gpt-4o',
    messages: [
      { role: 'system', content: SYSTEM_PROMPT },
      { role: 'user', content: userMessage },
    ],
  });
  return response.choices[0].message.content;
}

import { z } from 'zod';

const ReviewSchema = z.object({
  summary: z.string(),
  bugs: z.array(z.object({
    file: z.string(),
    line: z.number(),
    severity: z.enum(['low', 'medium', 'high']),
    description: z.string(),
  })),
  score: z.number().min(1).max(10),
});

async function reviewCode(code: string) {
  const response = await openai.chat.completions.create({
    model: 'gpt-4o',
    messages: [
      { role: 'system', content: 'Review code và trả về JSON.' },
      { role: 'user', content: code },
    ],
    response_format: { type: 'json_object' },
  });

  return ReviewSchema.parse(JSON.parse(response.choices[0].message.content));
}

import { OpenAIEmbeddings } from '@langchain/openai';
import { PineconeStore } from '@langchain/pinecone';
import { Pinecone } from '@pinecone-database/pinecone';

// 1. Index documents
async function indexDocuments(docs: string[]) {
  const embeddings = new OpenAIEmbeddings();
  const pc = new Pinecone({ apiKey: process.env.PINECONE_API_KEY! });
  const index = pc.Index('my-docs');

  for (const doc of docs) {
    const vector = await embeddings.embedQuery(doc);
    await index.upsert([{ id: crypto.randomUUID(), values: vector, metadata: { text: doc } }]);
  }
}

// 2. Query với RAG
async function queryWithRAG(question: string) {
  const embeddings = new OpenAIEmbeddings();
  const pc = new Pinecone({ apiKey: process.env.PINECONE_API_KEY! });
  const index = pc.Index('my-docs');

  // Search
  const queryVector = await embeddings.embedQuery(question);
  const results = await index.query({ vector: queryVector, topK: 5 });
  const context = results.matches.map(m => m.metadata?.text).join('\n');

  // Generate
  const response = await openai.chat.completions.create({
    model: 'gpt-4o',
    messages: [
      { role: 'system', content: 'Answer based on context provided.' },
      { role: 'user', content: `Context:\n${context}\n\nQuestion: ${question}` },
    ],
  });

  return response.choices[0].message.content;
}

async function evaluateResponse(question: string, response: string, expected?: string) {
  const eval = await openai.chat.completions.create({
    model: 'gpt-4o',
    messages: [
      {
        role: 'system',
        content: `Evaluate the response on:
        1. Correctness (1-5)
        2. Relevance (1-5)
        3. Completeness (1-5)
        Return JSON with scores and explanation.`
      },
      {
        role: 'user',
        content: `Question: ${question}\nResponse: ${response}${expected ? `\nExpected: ${expected}` : ''}`
      },
    ],
    response_format: { type: 'json_object' },
  });

  return JSON.parse(eval.choices[0].message.content);
}