geminiClient.js

/**
 * src/geminiClient.js   —  RENDER-OPTIMISED (stateless + rate-limited)
 * ─────────────────────────────────────────────────────────────────────
 * Two critical changes from the original version:
 *
 *   1. STATELESS — no in-memory session Map.
 *      Render's free tier spins the instance down after 15 min idle;
 *      when it wakes again every variable in RAM is gone.
 *      Fix: the frontend owns conversation history (an array of
 *      {role, text} objects) and sends it with every request.  The server
 *      rebuilds the full prompt each time.  Cost impact is negligible —
 *      a 4-turn history is still only ~1 300 input tokens.
 *
 *   2. REQUEST QUEUE with exponential back-off.
 *      Gemini 1.5 Flash free tier allows 15 requests / minute.
 *      Under a burst the API returns HTTP 429.  The queue serialises
 *      outgoing calls and retries with back-off so no request is dropped.
 *
 * GITHUB PATH  →  src/geminiClient.js
 */

"use strict";

const { GoogleGenerativeAI }          = require("@google/generative-ai");
const { getFullKnowledgeBase, getRelevantContext } = require("./knowledgeBase");

// ── SDK init ─────────────────────────────────────────────────────────────────
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
const MODEL = "gemini-1.5-flash";

// ── System prompt (rebuilt on every request — stateless) ─────────────────────
function buildSystemPrompt() {
  return `You are the official Skedulelt Support Assistant.
Skedulelt is a mobile scheduling & payment app for service providers
and customers in Trinidad & Tobago.

Rules:
  • Answer ONLY based on the knowledge base below.
  • If the question falls outside the knowledge base, say:
    "I'm sorry, I don't have information on that. Please contact
     our support team for further assistance."
  • Be friendly, concise, and helpful.
  • Do NOT hallucinate features, policies, or prices.
  • Respond in English.

════════════════════════════════════════════════
 SKEDULELT KNOWLEDGE BASE
════════════════════════════════════════════════
${getFullKnowledgeBase()}
════════════════════════════════════════════════`;
}

// ── Rate-limit queue ─────────────────────────────────────────────────────────
// Simple FIFO.  Each item is { resolve, reject, fn }.  The runner pulls one
// at a time; on a 429 it waits and retries in-place before moving on.

const queue    = [];
let   running  = false;

const MAX_RETRIES   = 4;
const BASE_DELAY_MS = 2000;   // first retry after 2 s; then 4, 8, 16 s

async function enqueue(fn) {
  return new Promise((resolve, reject) => {
    queue.push({ resolve, reject, fn });
    if (!running) runQueue();
  });
}

async function runQueue() {
  running = true;
  while (queue.length > 0) {
    const { resolve, reject, fn } = queue.shift();
    try      { resolve(await executeWithRetry(fn)); }
    catch(e) { reject(e); }
  }
  running = false;
}

async function executeWithRetry(fn) {
  let lastErr;
  for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
    try { return await fn(); }
    catch (err) {
      lastErr = err;
      const status = err.status || (err.response && err.response.status);
      if (status === 429 && attempt < MAX_RETRIES) {
        const delay = BASE_DELAY_MS * Math.pow(2, attempt);
        console.warn(`[Queue] 429 — retry ${attempt + 1}/${MAX_RETRIES} in ${delay} ms`);
        await new Promise(r => setTimeout(r, delay));
      } else { throw err; }
    }
  }
  throw lastErr;
}

// ── Public API ───────────────────────────────────────────────────────────────

/**
 * Stateless chat.
 *
 * @param {Array<{role:"user"|"assistant", text:string}>} history
 *        Full conversation so far — owned & sent by the frontend.
 * @param {string} currentMessage   The user's latest message.
 * @returns {Promise<{ answer:string, matchedSections:string[] }>}
 */
async function chat(history, currentMessage) {
  const { matched } = getRelevantContext(currentMessage);

  // Build the Gemini history array:
  //   [0] user  → system prompt + KB
  //   [1] model → short grounding ack
  //   [2..n]    → prior conversation from frontend
  const geminiHistory = [
    { role: "user",  parts: [{ text: buildSystemPrompt() }] },
    { role: "model", parts: [{ text: "Got it. I'm the Skedulelt Support Assistant. I'll answer only based on the knowledge base provided. How can I help?" }] }
  ];

  for (const turn of history) {
    geminiHistory.push({
      role : turn.role === "assistant" ? "model" : "user",
      parts: [{ text: turn.text }]
    });
  }

  // ── Enqueue the Gemini call (respects 15 RPM free-tier limit) ──────────
  const answer = await enqueue(async () => {
    const model       = genAI.getGenerativeModel({ model: MODEL });
    const chatSession = model.startChat({ history: geminiHistory });
    const response    = await chatSession.sendMessage(currentMessage);
    return response.text;
  });

  return { answer, matchedSections: matched };
}

module.exports = { chat };