AI2

Running

AI2 / server.js
Upload server.js
48ae3dd verified 4 months ago
24.6 kB
/**
 * GSS-TEC Sovereign Node — Hugging Face Space
 * Space: https://huggingface.co/spaces/Gaston895/AI
 *
 * Purpose: Receive requests from users, verify JWT, forward to Groq or Ollama.
 *          If the subscriber has a Business Profile (Turso creds in JWT),
 *          fetches their knowledge base and injects it into the system prompt.
 *
 * Routes:
 *   GET  /ping                   - Keep-alive (UptimeRobot)
 *   GET  /health                 - Health check
 *   POST /gss-internal/sync-keys - Receive AI keys from Cloudflare (bridge secret)
 *   POST /v1/chat/completions    - Main chat proxy (Groq or Ollama, JWT required)
 *
 * Three-Way Data Flow:
 *   1. SDK → Cloudflare /auth/lease  → JWT (with optional Turso creds embedded)
 *   2. SDK → HF /v1/chat/completions → JWT verified in RAM (zero DB call)
 *   3. HF  → Turso (if JWT has turso_url) → fetches 20MB business knowledge
 *   4. HF  → Groq/Ollama             → prompt + knowledge → answer → SDK
 *
 * HF Secrets to set:
 *   GSS_MASTER_SECRET  - JWT signing secret (same as Cloudflare Worker)
 *   GROQ_BASE_URL      - https://api.groq.com/openai/v1  (default)
 *   OLLAMA_BASE_URL    - e.g. http://your-ollama-server:11434/v1  (optional)
 *   SPACE_URL          - This space's public URL (for self keep-alive ping)
 */

const express = require('express');
const jwt     = require('jsonwebtoken');
const { createClient } = require('@libsql/client');

const CF_WORKER_URL = process.env.CF_WORKER_URL || 'https://qssn-d1-api.gastonsoftwaresolutions234.workers.dev';
console.log('[GSS] CF_WORKER_URL:', CF_WORKER_URL);

// Helper function to remove markdown formatting
function cleanMarkdown(text) {
  if (!text) return text;
  return text
    .replace(/\*\*(.+?)\*\*/g, '$1')  // Remove bold **text**
    .replace(/\*(.+?)\*/g, '$1')      // Remove italic *text*
    .replace(/\_\_(.+?)\_\_/g, '$1')  // Remove bold __text__
    .replace(/\_(.+?)\_/g, '$1')      // Remove italic _text_
    .replace(/\~\~(.+?)\~\~/g, '$1')  // Remove strikethrough ~~text~~
    .replace(/\`(.+?)\`/g, '$1');     // Remove inline code `text`
}

const app = express();
app.use(express.json({ limit: '100mb' }));
const SDK_JS = `
class GSSClient {
  constructor(opts) {
    if (!opts.apiKey || !opts.cfWorkerUrl || !opts.hfEngineUrl) {
      throw new Error('[GSSClient] apiKey, cfWorkerUrl, and hfEngineUrl are required');
    }
    this.apiKey      = opts.apiKey;
    this.cfWorkerUrl = opts.cfWorkerUrl.replace(/\\/$/,'');
    this.hfEngineUrl = opts.hfEngineUrl.replace(/\\/$/,'');
    this.model       = opts.model || 'llama-3.1-8b-instant';
    this.storageKey  = opts.storageKey || 'gss_jwt';
    this._token      = null;
    if (typeof window !== 'undefined') this._startKeepAlive();
  }
  _loadCachedToken() {
    try {
      const raw = localStorage.getItem(this.storageKey);
      if (!raw) return null;
      const { token, exp } = JSON.parse(raw);
      if (Date.now() / 1000 < exp - 300) return token;
      localStorage.removeItem(this.storageKey);
    } catch {}
    return null;
  }
  _decodeExp(token) {
    try {
      const p = JSON.parse(atob(token.split('.')[1].replace(/-/g,'+').replace(/_/g,'/')));
      return p.exp || 0;
    } catch { return 0; }
  }
  async _lease() {
    const res = await fetch(this.cfWorkerUrl + '/auth/lease', {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({ api_key: this.apiKey }),
    });
    if (res.status === 403) throw new Error('[GSSClient] Subscription expired or invalid API key');
    if (!res.ok) throw new Error('[GSSClient] Lease failed: ' + res.status);
    const { token } = await res.json();
    const exp = this._decodeExp(token);
    localStorage.setItem(this.storageKey, JSON.stringify({ token, exp }));
    this._token = token;
    return token;
  }
  async _getToken() {
    if (!this._token) this._token = this._loadCachedToken();
    if (!this._token) this._token = await this._lease();
    return this._token;
  }
  async chat(messages, opts = {}) {
    const token = await this._getToken();
    const body = {
      model:       opts.model || this.model,
      messages,
      temperature: opts.temperature ?? 0.7,
      max_tokens:  opts.max_tokens  ?? 1024,
    };
    if (opts.provider) body.provider = opts.provider;
    const res = await fetch(this.hfEngineUrl + '/v1/chat/completions', {
      method: 'POST',
      headers: { 'Authorization': 'Bearer ' + token, 'Content-Type': 'application/json' },
      body: JSON.stringify(body),
    });
    if (res.status === 401 || res.status === 403) {
      localStorage.removeItem(this.storageKey);
      this._token = null;
      return this.chat(messages, opts);
    }
    if (!res.ok) {
      const err = await res.json().catch(() => ({ error: res.statusText }));
      throw new Error('[GSSClient] Chat error ' + res.status + ': ' + (err.error || res.statusText));
    }
    const data = await res.json();
    return data.choices?.[0]?.message?.content ?? '';
  }
  async ping() { return (await fetch(this.hfEngineUrl + '/ping')).json(); }
  async renewToken() {
    localStorage.removeItem(this.storageKey);
    this._token = null;
    return this._lease();
  }
  async _syncKeys() {
    try {
      const token = await this._getToken();
      await fetch(this.cfWorkerUrl + '/auth/keep-alive', {
        method: 'POST',
        headers: { 'Authorization': 'Bearer ' + token, 'Content-Type': 'application/json' },
      });
    } catch (_) {}
  }
  _startKeepAlive() {
    setTimeout(() => this._syncKeys(), 5000);
    setInterval(() => this._syncKeys(), 30 * 60 * 1000);
  }
}
if (typeof module !== 'undefined' && module.exports) module.exports = GSSClient;
else if (typeof window !== 'undefined') window.GSSClient = GSSClient;
`;

// Serve SDK from both paths for compatibility
app.get('/sdk/gss-sdk.js', (_req, res) => {
  res.setHeader('Content-Type', 'application/javascript');
  res.setHeader('Cache-Control', 'public, max-age=3600');
  res.send(SDK_JS);
});
app.get('/gss-sdk.js', (_req, res) => {
  res.setHeader('Content-Type', 'application/javascript');
  res.setHeader('Cache-Control', 'public, max-age=3600');
  res.send(SDK_JS);
});

// ── CORS ──────────────────────────────────────────────────────────────────────
app.use((req, res, next) => {
  res.setHeader('Access-Control-Allow-Origin',  '*');
  res.setHeader('Access-Control-Allow-Methods', 'GET, POST, OPTIONS');
  res.setHeader('Access-Control-Allow-Headers', 'Content-Type, Authorization');
  if (req.method === 'OPTIONS') return res.sendStatus(204);
  next();
});

// ── Per-subscriber key pools ──────────────────────────────────────────────────
// Structure: { [subscriberId]: { groq: KeyEntry[], ollama: KeyEntry[], cursor, knowledge, profile } }
const subscriberPools = {};

function makeEntry(key, index) {
  return { index, key, failures: 0, coolUntil: 0 };
}

function loadSubscriberKeys(subscriberId, groqKeys = [], ollamaKeys = [], knowledge = [], profile = null) {
  subscriberPools[subscriberId] = {
    groq:      groqKeys.map((k, i) => makeEntry(k, i)),
    ollama:    ollamaKeys.map((k, i) => makeEntry(k, i)),
    cursor:    { groq: 0, ollama: 0 },
    knowledge: knowledge || [],
    profile:   profile   || null,
  };
  console.log(`[GSS] Subscriber ${subscriberId}: ${groqKeys.length} Groq + ${ollamaKeys.length} Ollama keys loaded, ${(knowledge||[]).length} knowledge rows`);
}

function pickKeyForSubscriber(subscriberId, provider = 'groq') {
  const pool = subscriberPools[subscriberId];
  if (!pool) return null;
  const keys = pool[provider] || [];
  const now = Date.now();
  const available = keys.filter(e => now >= e.coolUntil);
  if (!available.length) return null;
  const cursor = pool.cursor[provider] || 0;
  const entry = available[cursor % available.length];
  pool.cursor[provider] = (cursor + 1) % available.length;
  return entry;
}

function cooldownKey(entry) {
  entry.failures++;
  entry.coolUntil = Date.now() + 60_000;
  console.warn(`[GSS] Key #${entry.index} on cooldown (failures: ${entry.failures})`);
}

// ── Load keys from HF env var on startup ─────────────────────────────────────
// Fallback: GROQ_KEYS env var loads as a default subscriber (id=0)
function loadKeysFromEnv() {
  const raw = process.env.GROQ_KEYS || '';
  const keys = raw.split(',').map(k => k.trim()).filter(Boolean);
  if (keys.length) {
    const current = subscriberPools['env'] || { groq: [] };
    if (keys.length !== current.groq.length) {
      loadSubscriberKeys('env', keys, []);
    }
  }
}
loadKeysFromEnv();
setInterval(loadKeysFromEnv, 5000);

// Total key count for /ping
function totalKeys() {
  return Object.values(subscriberPools).reduce((n, p) => n + p.groq.length + p.ollama.length, 0);
}

// ── JWT middleware ────────────────────────────────────────────────────────────
const gssAuth = (req, res, next) => {
  const authHeader = req.headers['authorization'];
  const token = authHeader && authHeader.split(' ')[1];

  if (!token) {
    return res.status(401).json({ error: 'GSS-TEC: No Access Token Provided.' });
  }

  jwt.verify(token, process.env.GSS_MASTER_SECRET, (err, user) => {
    if (err) {
      return res.status(403).json({ error: 'GSS-TEC: Token Invalid or Expired.' });
    }
    req.user = user;
    next();
  });
};

// ── Keep-alive self-ping ──────────────────────────────────────────────────────
// Pings itself every 2 minutes to prevent HF Space from sleeping
const SPACE_URL = process.env.SPACE_URL || 'https://gaston895-ai.hf.space';
setInterval(async () => {
  try {
    await fetch(`${SPACE_URL}/ping`);
    console.log(`[GSS] Keep-alive ping OK — subscribers: ${Object.keys(subscriberPools).length}, keys: ${totalKeys()}`);
  } catch (err) {
    console.warn('[GSS] Keep-alive ping failed:', err.message);
  }
}, 2 * 60 * 1000); // every 2 minutes
app.get('/',       (_req, res) => res.redirect('/ping'));
app.get('/ping',   (_req, res) => res.json({ status: 'GSS Engine Online', subscribers: Object.keys(subscriberPools).length, keys: totalKeys(), ts: new Date().toISOString() }));
app.get('/health', (_req, res) => res.json({ status: 'healthy', ts: new Date().toISOString() }));

// ── Sync route (Cloudflare → HF RAM) ─────────────────────────────────────────
// Body: { subscribers: [{ id, groq_keys: [], ollama_keys: [] }] }
app.post('/gss-internal/sync-keys', (req, res) => {
  const secret = process.env.GSS_MASTER_SECRET;
  if (!secret || req.headers['x-gss-bridge-secret'] !== secret) {
    return res.status(403).json({ error: 'Forbidden' });
  }
  const { subscribers } = req.body;
  if (!Array.isArray(subscribers) || !subscribers.length) {
    return res.status(400).json({ error: 'subscribers[] required' });
  }
  for (const sub of subscribers) {
    if (sub.id && (sub.groq_keys?.length || sub.ollama_keys?.length)) {
      loadSubscriberKeys(String(sub.id), sub.groq_keys || [], sub.ollama_keys || [], sub.knowledge || [], sub.profile || null);
    }
  }
  res.json({ success: true, subscribers: Object.keys(subscriberPools).length, total_keys: totalKeys() });
});

// ── Main chat proxy ───────────────────────────────────────────────────────────
// Support both /chat and /v1/chat/completions for compatibility
app.post('/chat', gssAuth, async (req, res) => {
  return handleChatCompletion(req, res);
});

app.post('/v1/chat/completions', gssAuth, async (req, res) => {
  return handleChatCompletion(req, res);
});

async function handleChatCompletion(req, res) {
  const subscriberId = String(req.user.sub);
  const useOllama = req.body.provider === 'ollama';
  const provider  = useOllama ? 'ollama' : 'groq';
  delete req.body.provider;

  if (!req.body.model) {
    req.body.model = useOllama ? 'minimax-m2:cloud' : 'llama-3.1-8b-instant';
  }

  // ── Step 4: Inject business knowledge from RAM (pushed by Cloudflare cron) ──
  const tursoUrl   = req.user.turso_url;
  const tursoToken = req.user.turso_token;
  // Also check JWT-embedded knowledge (from /auth/lease)
  const jwtKnowledge = req.user.knowledge || [];
  const jwtProfile   = req.user.profile   || null;
  // RAM knowledge (pushed via keep-alive/cron sync)
  const ramPool      = subscriberPools[subscriberId];
  const ramKnowledge = ramPool?.knowledge || [];
  const ramProfile   = ramPool?.profile   || null;

  const messages    = req.body.messages || [];
  const lastUserMsg = [...messages].reverse().find(m => m.role === 'user')?.content || '';
  const words       = new Set(lastUserMsg.toLowerCase().replace(/[^a-z0-9\s]/g,' ').split(/\s+/).filter(w => w.length > 3));

  let snippets = [];
  let activeProfile = ramProfile || jwtProfile;

  if (tursoUrl && tursoToken) {
    // Turso path — large knowledge vaults (>5MB)
    try {
      const db = createClient({ url: tursoUrl, authToken: tursoToken });
      let rows = [];
      try {
        const r = await db.execute({ sql: `SELECT label, content FROM gss_business_knowledge WHERE content LIKE ?`, args: [`%${[...words].slice(0,3).join('%')}%`] });
        rows = r.rows;
      } catch(_) {}
      if (!rows.length) {
        const all = await db.execute('SELECT label, content FROM gss_business_knowledge ORDER BY label ASC');
        rows = all.rows.map(r => ({...r, score:[...words].filter(w=>(r.content||'').toLowerCase().includes(w)).length})).sort((a,b)=>b.score-a.score);
      }
      snippets = rows;
    } catch(err) { console.warn(`[GSS] Turso failed for ${subscriberId}:`, err.message); }
  } else {
    // RAM/JWT path — use knowledge pushed by Cloudflare (no outbound fetch needed)
    const allRows = ramKnowledge.length ? ramKnowledge : jwtKnowledge;
    if (allRows.length) {
      snippets = allRows
        .map(r => ({...r, score:[...words].filter(w=>(r.content||'').toLowerCase().includes(w)).length}))
        .sort((a,b) => b.score - a.score);
      if (!snippets.length) snippets = allRows;
    }
  }

  if (snippets.length > 0 || activeProfile) {
    let systemContent = '';
    if (activeProfile) {
      const p = activeProfile;
      const name = p.ai_name || 'Assistant';
      systemContent += `You are ${name}. Your name is ${name} and ONLY ${name}. Never use any other name. Never say your name is anything other than ${name}.\n`;
      if (p.biz_name)    systemContent += `You work for: ${p.biz_name}\n`;
      if (p.biz_type)    systemContent += `Industry: ${p.biz_type}\n`;
      if (p.biz_mission) systemContent += `Mission: ${p.biz_mission}\n`;
      if (p.biz_location)  systemContent += `Location: ${p.biz_location}\n`;
      if (p.biz_phone)     systemContent += `Phone/WhatsApp: ${p.biz_phone}\n`;
      if (p.biz_website)   systemContent += `Website: ${p.biz_website}\n`;
      if (p.biz_whatsapp)  systemContent += `WhatsApp: ${p.biz_whatsapp}\n`;
      if (p.biz_facebook)  systemContent += `Facebook: ${p.biz_facebook}\n`;
      if (p.biz_instagram) systemContent += `Instagram: ${p.biz_instagram}\n`;
      if (p.ai_role)     systemContent += `Your role: ${p.ai_role}.`;
      if (p.ai_tone)     systemContent += ` Tone: ${p.ai_tone}.`;
      if (p.ai_language) systemContent += ` Language: ${p.ai_language}.`;
      if (p.ai_goal)     systemContent += ` Goal: ${p.ai_goal}.`;
      if (p.ai_rules)    systemContent += `\nRules you MUST follow: ${p.ai_rules}`;
      systemContent += '\n\n';
    }
    if (snippets.length) {
      const knowledgeBlock = snippets.map(r => `[${r.label}]\n${r.content||''}`).join('\n\n---\n\n');
      // Check if this is auto-sync knowledge (contains "Auto-Sync: Enabled")
      const isAutoSync = snippets.some(r => (r.content || '').includes('Auto-Sync: Enabled'));
      
      if (isAutoSync) {
        // AUTO-SYNC MODE: AI can answer ANY question, using knowledge as context
        systemContent += `You are a helpful AI assistant. Use the following context about the user when relevant:\n\n${knowledgeBlock}\n\nYou can answer any question the user asks, not just questions about the context above. Be helpful, accurate, and conversational.`;
      } else {
        // NORMAL MODE: Restrict to business knowledge only
        systemContent += `Answer using ONLY this knowledge:\n\n${knowledgeBlock}\n\nIf unsure, say: "I don't have that information. Please contact us directly."`;
      }
    }
    const existingSystem = messages.findIndex(m => m.role === 'system');
    if (existingSystem >= 0) { messages[existingSystem].content = systemContent + '\n\n' + messages[existingSystem].content; }
    else { messages.unshift({ role: 'system', content: systemContent }); }
    req.body.messages = messages;
    console.log(`[GSS] Injected ${snippets.length} snippets + profile for sub ${subscriberId}`);
  }

  // Check subscriber has keys loaded
  const pool = subscriberPools[subscriberId];
  if (!pool || !(pool[provider]?.length)) {
    // Fallback to env pool if subscriber not found (e.g. env-loaded keys)
    const envPool = subscriberPools['env'];
    if (!envPool || !envPool.groq.length) {
      return res.status(503).json({ error: 'No AI keys loaded for your account. Please check your subscription.' });
    }
    // Use env pool only if subscriber matches env (single-tenant fallback)
  }

  const baseUrl = useOllama
    ? (process.env.OLLAMA_BASE_URL || 'https://ollama.com/api')
    : (process.env.GROQ_BASE_URL   || 'https://api.groq.com/openai/v1');

  // For Ollama cloud, use native /chat endpoint with different model format
  const isOllamaCloud = useOllama && baseUrl.includes('ollama.com');
  const chatEndpoint  = isOllamaCloud ? `${baseUrl}/chat` : `${baseUrl}/chat/completions`;
  if (isOllamaCloud && !req.body.model?.includes(':')) {
    req.body.model = 'gpt-oss:120b';
  }

  const poolKeys = (subscriberPools[subscriberId]?.[provider]) || (subscriberPools['env']?.groq) || [];
  const maxRetries = Math.min(poolKeys.length, 10);

  // Track if we've already fallen back to Ollama to avoid infinite loops
  let ollamaFallbackAttempted = false;

  for (let attempt = 0; attempt < maxRetries; attempt++) {
    const entry = pickKeyForSubscriber(subscriberId, provider)
                || pickKeyForSubscriber('env', 'groq'); // env fallback
    if (!entry) {
      // No Groq keys left — try Ollama fallback
      if (!ollamaFallbackAttempted && !useOllama) {
        const ollamaEntry = pickKeyForSubscriber(subscriberId, 'ollama');
        if (ollamaEntry) {
          ollamaFallbackAttempted = true;
          console.warn(`[GSS] Groq exhausted for sub ${subscriberId}, switching to Ollama cloud`);
          try {
            const ollamaUrl = process.env.OLLAMA_BASE_URL || 'https://ollama.com/api';
            const ollamaBody = {
              model: 'gpt-oss:120b',
              messages: req.body.messages || [],
              stream: false,
            };
            const upstream = await fetch(`${ollamaUrl}/chat`, {
              method: 'POST',
              headers: { 'Authorization': `Bearer ${ollamaEntry.key}`, 'Content-Type': 'application/json' },
              body: JSON.stringify(ollamaBody),
            });
            if (upstream.ok) {
              const ollamaData = await upstream.json();
              const content = ollamaData.message?.content || ollamaData.response || '';
              return res.status(200).json({
                choices: [{ message: { role: 'assistant', content }, finish_reason: 'stop' }],
                model: ollamaBody.model,
              });
            }
          } catch (_) { /* fall through to error */ }
        }
      }
      return res.status(429).json({ error: 'All your API keys are rate-limited. Try again in 60s.' });
    }

    try {
      const upstream = await fetch(chatEndpoint, {
        method:  'POST',
        headers: { 'Authorization': `Bearer ${entry.key}`, 'Content-Type': 'application/json' },
        body: JSON.stringify(req.body),
      });

      // For Ollama cloud responses, convert to OpenAI format
      const processResponse = async (r) => {
        const data = await r.json();
        if (isOllamaCloud && data.message) {
          return { choices: [{ message: { role: 'assistant', content: data.message.content || '' }, finish_reason: 'stop' }], model: req.body.model };
        }
        return data;
      };

      // 401/403 = invalid/expired key — cooldown and try next
      if (upstream.status === 401 || upstream.status === 403) {
        cooldownKey(entry);
        console.warn(`[GSS] Key #${entry.index} invalid/expired for sub ${subscriberId} (${upstream.status})`);
        // If all Groq keys are invalid and Ollama available, switch
        const remaining = (subscriberPools[subscriberId]?.[provider] || []).filter(e => Date.now() >= e.coolUntil);
        if (remaining.length === 0 && !ollamaFallbackAttempted && !useOllama) {
          const ollamaEntry = pickKeyForSubscriber(subscriberId, 'ollama');
          if (ollamaEntry) {
            ollamaFallbackAttempted = true;
            console.warn(`[GSS] All Groq keys invalid for sub ${subscriberId}, switching to Ollama cloud`);
            const ollamaUrl = process.env.OLLAMA_BASE_URL || 'https://ollama.com/api';
            // Ollama cloud native format
            const ollamaBody = {
              model: 'gpt-oss:120b',
              messages: req.body.messages || [],
              stream: false,
            };
            const ollamaUpstream = await fetch(`${ollamaUrl}/chat`, {
              method: 'POST',
              headers: { 'Authorization': `Bearer ${ollamaEntry.key}`, 'Content-Type': 'application/json' },
              body: JSON.stringify(ollamaBody),
            });
            if (ollamaUpstream.ok) {
              const ollamaData = await ollamaUpstream.json();
              const content = ollamaData.message?.content || ollamaData.response || '';
              return res.status(200).json({
                choices: [{ message: { role: 'assistant', content }, finish_reason: 'stop' }],
                model: ollamaBody.model,
              });
            }
            const errText = await ollamaUpstream.text().catch(() => '');
            console.warn(`[GSS] Ollama cloud fallback failed (${ollamaUpstream.status}): ${errText.slice(0,200)}`);
          }
        }
        continue;
      }

      if (upstream.status === 429 || upstream.status >= 500) {
        cooldownKey(entry);
        continue;
      }

      const data = await processResponse(upstream);
      // Clean markdown from response
      if (data.choices && data.choices[0] && data.choices[0].message && data.choices[0].message.content) {
        data.choices[0].message.content = cleanMarkdown(data.choices[0].message.content);
      }
      return res.status(upstream.status).json(data);

    } catch (err) {
      cooldownKey(entry);
      console.error(`[GSS] Network error on key #${entry.index} for sub ${subscriberId}:`, err.message);
    }
  }

  res.status(503).json({ error: 'All your API keys exhausted. Try again shortly.' });
}

// ── Start ─────────────────────────────────────────────────────────────────────
const PORT = process.env.PORT || 7860;
app.listen(PORT, () => console.log(`[GSS] Sovereign Node live on :${PORT}`));