Gitea
301 lines
9.6 KiB
JavaScript
301 lines
9.6 KiB
JavaScript
require("dotenv").config();
|
|
|
|
const express = require("express");
|
|
const cors = require("cors");
|
|
const OpenAI = require("openai");
|
|
const { pipeline } = require("@xenova/transformers");
|
|
const { QdrantClient } = require("@qdrant/js-client-rest");
|
|
|
|
const CONFIG = {
|
|
azure: {
|
|
endpoint: process.env.AZURE_OPENAI_ENDPOINT || "https://cpmindiayoda-resource.services.ai.azure.com",
|
|
deployment: process.env.AZURE_DEPLOYMENT || "gpt-4o-mini",
|
|
apiVersion: process.env.AZURE_API_VERSION || "2024-08-01-preview",
|
|
apiKey: process.env.AZURE_OPENAI_KEY,
|
|
},
|
|
qdrant: {
|
|
url: process.env.QDRANT_URL || "http://20.40.61.65:6333",
|
|
collection: process.env.QDRANT_COLLECTION || "pdf_rag",
|
|
},
|
|
search: {
|
|
topK: 20,
|
|
minScore: 0.10,
|
|
maxContextDocs: 10,
|
|
},
|
|
port: process.env.PORT || 5000,
|
|
};
|
|
|
|
// ─── Clients ──────────────────────────────────────────────────────────────────
|
|
const llm = new OpenAI({
|
|
baseURL: `${CONFIG.azure.endpoint}/openai/deployments/${CONFIG.azure.deployment}`,
|
|
apiKey: CONFIG.azure.apiKey,
|
|
defaultHeaders: { "api-key": CONFIG.azure.apiKey },
|
|
defaultQuery: { "api-version": CONFIG.azure.apiVersion },
|
|
});
|
|
|
|
const qdrant = new QdrantClient({
|
|
url: CONFIG.qdrant.url,
|
|
checkCompatibility: false,
|
|
timeout: 30000,
|
|
});
|
|
|
|
// ─── Embedding model (singleton, lazy-init) ───────────────────────────────────
|
|
let _embedder = null;
|
|
async function getEmbedder() {
|
|
if (!_embedder) {
|
|
console.log("Loading MiniLM model...");
|
|
_embedder = await pipeline("feature-extraction", "Xenova/all-MiniLM-L6-v2");
|
|
console.log("Embedding model ready");
|
|
}
|
|
return _embedder;
|
|
}
|
|
|
|
async function createEmbedding(text) {
|
|
const model = await getEmbedder();
|
|
const out = await model(text, { pooling: "mean", normalize: true });
|
|
return Array.from(out.data);
|
|
}
|
|
|
|
// ─── Qdrant search ────────────────────────────────────────────────────────────
|
|
async function searchQdrant(embedding, { topK, minScore, maxContextDocs } = CONFIG.search) {
|
|
const results = await qdrant.search(CONFIG.qdrant.collection, {
|
|
vector: embedding,
|
|
limit: topK,
|
|
with_payload: true,
|
|
score_threshold: minScore,
|
|
});
|
|
|
|
return results
|
|
.sort((a, b) => b.score - a.score)
|
|
.slice(0, maxContextDocs);
|
|
}
|
|
|
|
// ─── Build LLM context string ─────────────────────────────────────────────────
|
|
function buildContext(results) {
|
|
return results
|
|
.map((item, i) =>
|
|
`[${i + 1}] File: ${item.payload?.file ?? "unknown"} | Page: ${item.payload?.page ?? "?"}\n${item.payload?.text ?? ""}`
|
|
)
|
|
.join("\n\n---\n\n");
|
|
}
|
|
|
|
// ─── LLM call ─────────────────────────────────────────────────────────────────
|
|
const SYSTEM_PROMPT = `
|
|
You are CPM AI Assistant.
|
|
|
|
Rules:
|
|
- Answer only from the provided information.
|
|
- If the answer is not available, reply exactly:
|
|
"❌ I could not find this information in the uploaded documents."
|
|
- Do not make up information.
|
|
- Do not mention documents, context, or chunks.
|
|
|
|
Response Style:
|
|
- Use simple English.
|
|
- Keep answers short and clear.
|
|
- Use headings and bullet points.
|
|
- Highlight important words in **bold**.
|
|
- Use emojis in headings.
|
|
|
|
Format:
|
|
|
|
# 📋 Topic
|
|
|
|
## 🎯 Summary
|
|
Short answer in 1-2 sentences.
|
|
|
|
## ✅ Details
|
|
- Point 1
|
|
- Point 2
|
|
- Point 3
|
|
|
|
## ⚠️ Notes
|
|
- Extra information (if available).
|
|
`.trim();
|
|
|
|
async function askLLM(question, context) {
|
|
const completion = await llm.chat.completions.create({
|
|
model: CONFIG.azure.deployment,
|
|
temperature: 0,
|
|
max_tokens: 1500,
|
|
messages: [
|
|
{ role: "system", content: SYSTEM_PROMPT },
|
|
{ role: "user", content: `Context:\n${context}\n\nQuestion:\n${question}` },
|
|
],
|
|
});
|
|
|
|
return completion.choices[0].message.content;
|
|
}
|
|
|
|
// ─── Express app ──────────────────────────────────────────────────────────────
|
|
const app = express();
|
|
app.use(cors());
|
|
app.use(express.json({ limit: "1mb" }));
|
|
app.use(express.urlencoded({ extended: true }));
|
|
|
|
app.use((req, _res, next) => {
|
|
console.log(`→ ${req.method} ${req.path}`);
|
|
next();
|
|
});
|
|
|
|
// ─── Routes ───────────────────────────────────────────────────────────────────
|
|
app.get("/health", (_req, res) => {
|
|
res.json({ status: "ok", model: CONFIG.azure.deployment, collection: CONFIG.qdrant.collection });
|
|
});
|
|
|
|
app.post("/ask", async (req, res) => {
|
|
const { question } = req.body ?? {};
|
|
|
|
if (!question?.trim()) {
|
|
return res.status(400).json({ success: false, error: "question is required" });
|
|
}
|
|
|
|
const t0 = Date.now();
|
|
|
|
try {
|
|
const embedding = await createEmbedding(question.trim());
|
|
const results = await searchQdrant(embedding);
|
|
|
|
if (!results.length) {
|
|
return res.json({
|
|
success: true,
|
|
question,
|
|
answer: "❌ I could not find this information in the uploaded documents.",
|
|
sources: [],
|
|
ms: Date.now() - t0,
|
|
});
|
|
}
|
|
|
|
const context = buildContext(results);
|
|
const answer = await askLLM(question, context);
|
|
|
|
return res.json({
|
|
success: true,
|
|
question,
|
|
answer,
|
|
sources: results.map(r => ({
|
|
score: +r.score.toFixed(4),
|
|
file: r.payload?.file,
|
|
page: r.payload?.page,
|
|
chunk: r.payload?.chunk,
|
|
})),
|
|
ms: Date.now() - t0,
|
|
});
|
|
|
|
} catch (err) {
|
|
console.error("❌ /ask error:", err);
|
|
return res.status(500).json({ success: false, error: err.message });
|
|
}
|
|
});
|
|
|
|
app.post("/ask/stream", async (req, res) => {
|
|
const { question } = req.body ?? {};
|
|
|
|
if (!question?.trim()) {
|
|
return res.status(400).json({ success: false, error: "question is required" });
|
|
}
|
|
|
|
res.setHeader("Content-Type", "text/event-stream");
|
|
res.setHeader("Cache-Control", "no-cache");
|
|
res.setHeader("Connection", "keep-alive");
|
|
res.flushHeaders();
|
|
|
|
const send = (event, data) => res.write(`event: ${event}\ndata: ${JSON.stringify(data)}\n\n`);
|
|
|
|
try {
|
|
send("status", { message: "🔍 Searching documents..." });
|
|
const embedding = await createEmbedding(question.trim());
|
|
|
|
const results = await searchQdrant(embedding);
|
|
|
|
if (!results.length) {
|
|
send("token", { token: "❌", isWord: true });
|
|
send("token", { token: "I", isWord: true });
|
|
send("token", { token: "could", isWord: true });
|
|
send("token", { token: "not", isWord: true });
|
|
send("token", { token: "find", isWord: true });
|
|
send("token", { token: "this", isWord: true });
|
|
send("token", { token: "information", isWord: true });
|
|
send("token", { token: "in", isWord: true });
|
|
send("token", { token: "the", isWord: true });
|
|
send("token", { token: "uploaded", isWord: true });
|
|
send("token", { token: "documents.", isWord: true });
|
|
send("done", { sources: [] });
|
|
return res.end();
|
|
}
|
|
|
|
const sources = results.map(r => ({
|
|
score: +r.score.toFixed(4),
|
|
file: r.payload?.file,
|
|
page: r.payload?.page,
|
|
chunk: r.payload?.chunk,
|
|
}));
|
|
send("sources", { sources });
|
|
|
|
send("status", { message: "💬 Generating answer..." });
|
|
|
|
const context = buildContext(results);
|
|
const stream = await llm.chat.completions.create({
|
|
model: CONFIG.azure.deployment,
|
|
temperature: 0,
|
|
max_tokens: 1500,
|
|
stream: true,
|
|
messages: [
|
|
{ role: "system", content: SYSTEM_PROMPT },
|
|
{ role: "user", content: `Context:\n${context}\n\nQuestion:\n${question}` },
|
|
],
|
|
});
|
|
|
|
// Buffer to handle tokens that may be split mid-word
|
|
let wordBuffer = "";
|
|
|
|
for await (const chunk of stream) {
|
|
const rawToken = chunk.choices[0]?.delta?.content ?? "";
|
|
if (!rawToken) continue;
|
|
|
|
wordBuffer += rawToken;
|
|
|
|
// Split on whitespace — emit complete words, keep trailing partial
|
|
// We preserve newlines/markdown as separate tokens so markdown renders correctly
|
|
const parts = wordBuffer.split(/(\s+)/);
|
|
|
|
// Last element might be an incomplete word — buffer it
|
|
wordBuffer = parts.pop() ?? "";
|
|
|
|
for (const part of parts) {
|
|
if (part) {
|
|
send("token", { token: part, isWord: /\S/.test(part) });
|
|
}
|
|
}
|
|
}
|
|
|
|
// Flush any remaining buffered text
|
|
if (wordBuffer) {
|
|
send("token", { token: wordBuffer, isWord: true });
|
|
}
|
|
|
|
send("done", { sources });
|
|
|
|
} catch (err) {
|
|
console.error("❌ /ask/stream error:", err);
|
|
send("error", { error: err.message });
|
|
}
|
|
|
|
res.end();
|
|
});
|
|
|
|
app.use((_req, res) => res.status(404).json({ success: false, error: "Not found" }));
|
|
|
|
// ─── Start ────────────────────────────────────────────────────────────────────
|
|
async function start() {
|
|
await getEmbedder();
|
|
|
|
app.listen(CONFIG.port, () => {
|
|
console.log(`Server running on port ${CONFIG.port}`);
|
|
});
|
|
}
|
|
|
|
start().catch(err => {
|
|
console.error("Fatal startup error:", err);
|
|
process.exit(1);
|
|
}); |