LJTSG/gemma-webgpu-thinking-engine

1const http = require('http'), fs = require('fs'), path = require('path');2const fetch = globalThis.fetch || require('node-fetch'); // Node 18+ has built-in fetch3const PORT = 8150;4const ROOT = __dirname;5const SPLITS = path.join(ROOT, 'model_splits');6 7const MIME = {8  '.html': 'text/html', '.js': 'text/javascript', '.mjs': 'text/javascript',9  '.json': 'application/json', '.gguf': 'application/octet-stream',10  '.wasm': 'application/wasm', '.ts': 'text/javascript',11};12 13const ENTITY_DIR = path.join(ROOT, '..', 'entity');14 15async function extractKeywords(text) {16  // Try local xLAM at :8093 for intent-based keyword extraction17  try {18    const resp = await fetch('http://127.0.0.1:8093/v1/chat/completions', {19      method: 'POST',20      headers: { 'Content-Type': 'application/json' },21      body: JSON.stringify({22        model: 'xlam',23        messages: [{ role: 'user', content: `Extract 3-5 keyword phrases (1-3 words each) from this message that capture the user's intent. Return JSON: {"keywords":["k1","k2","k3"]}\n\nMessage: "${text}"` }],24        temperature: 0.1,25        max_tokens: 60,26      }),27    });28    if (resp.ok) {29      const data = await resp.json();30      const content = data.choices?.[0]?.message?.content || '';31      const match = content.match(/\{[\s\S]*"keywords"[\s\S]*\}/);32      if (match) {33        const parsed = JSON.parse(match[0]);34        if (Array.isArray(parsed.keywords)) return parsed.keywords.map(k => k.toLowerCase());35      }36    }37  } catch (e) { /* xLAM not available, fall back */ }38  // Fallback: extract names and nouns (4+ chars, skip common words)39  const stop = new Set(['what','that','this','with','from','have','your','been','does','were','they','their','about','would','could','should','there','where','which','these','those','before','after','other','being','still','never','always','remember']);40  return (text.toLowerCase().match(/[a-z]{4,}/g) || []).filter(w => !stop.has(w)).slice(0, 5);41}42 43function grepEntity(text, keywords) {44  const unique = [...new Set(keywords)];45  if (unique.length === 0) return [];46 47  const results = [];48  function scanDir(dir) {49    try {50      const entries = fs.readdirSync(dir, { withFileTypes: true });51      for (const e of entries) {52        const fp = path.join(dir, e.name);53        if (e.isDirectory()) { scanDir(fp); continue; }54        if (!/\.(md|json|txt)$/.test(e.name)) continue;55        try {56          const content = fs.readFileSync(fp, 'utf8');57          const lines = content.split('\n');58          for (let i = 0; i < lines.length; i++) {59            const line = lines[i].trim();60            if (line.length < 10 || line.length > 300) continue;61            const lower = line.toLowerCase();62            for (const kw of unique) {63              if (lower.includes(kw)) {64                results.push(line);65                break;66              }67            }68          }69        } catch (e) {}70      }71    } catch (e) {}72  }73  scanDir(ENTITY_DIR);74  return [...new Set(results)].slice(0, 6);75}76 77http.createServer((req, res) => {78  // Handle grep API79  if (req.method === 'POST' && req.url === '/api/grep') {80    let body = '';81    req.on('data', c => body += c);82    req.on('end', async () => {83      try {84        const { text } = JSON.parse(body);85        const keywords = await extractKeywords(text);86        console.log('[grep] keywords:', keywords);87        const results = grepEntity(text, keywords);88        console.log('[grep] found:', results.length, 'snippets');89        res.writeHead(200, { 'Content-Type': 'application/json', 'Access-Control-Allow-Origin': '*' });90        res.end(JSON.stringify({ results, keywords }));91      } catch (e) {92        res.writeHead(500, { 'Content-Type': 'application/json' });93        res.end(JSON.stringify({ error: e.message }));94      }95    });96    return;97  }98 99  if (req.method === 'OPTIONS') {100    res.writeHead(200, {101      'Access-Control-Allow-Origin': '*',102      'Access-Control-Allow-Methods': 'GET,HEAD',103      'Access-Control-Allow-Headers': 'Content-Type,Range',104      'Cross-Origin-Embedder-Policy': 'require-corp',105      'Cross-Origin-Opener-Policy': 'same-origin',106    });107    return res.end();108  }109 110  let p = decodeURIComponent(req.url.split('?')[0]);111  if (p === '/') p = '/index.html';112 113  let fp;114  if (p.startsWith('/model/')) {115    fp = path.join(SPLITS, p.slice('/model/'.length));116  } else {117    // Resolve relative to project root, allowing ../mamba_webgpu/ paths118    fp = path.resolve(ROOT, '.' + p);119  }120 121  fs.stat(fp, (e, st) => {122    if (e) { res.writeHead(404); return res.end('not found: ' + fp); }123    const total = st.size;124    const range = req.headers.range;125    const ct = MIME[path.extname(fp)] || 'application/octet-stream';126 127    const headers = {128      'Content-Type': ct,129      'Accept-Ranges': 'bytes',130      'Access-Control-Allow-Origin': '*',131      'Cross-Origin-Embedder-Policy': 'require-corp',132      'Cross-Origin-Opener-Policy': 'same-origin',133    };134 135    if (range) {136      const m = range.match(/bytes=(\d+)-(\d*)/);137      if (m) {138        const start = parseInt(m[1]);139        const end = m[2] ? parseInt(m[2]) : total - 1;140        headers['Content-Range'] = 'bytes ' + start + '-' + end + '/' + total;141        headers['Content-Length'] = end - start + 1;142        res.writeHead(206, headers);143        fs.createReadStream(fp, { start, end }).pipe(res);144        return;145      }146    }147 148    headers['Content-Length'] = total;149    res.writeHead(200, headers);150    fs.createReadStream(fp).pipe(res);151  });152}).listen(PORT, () => {153  console.log('Gemma WebGPU on :' + PORT);154  console.log('Model splits: ' + SPLITS);155  const files = fs.readdirSync(SPLITS).filter(f => f.endsWith('.gguf'));156  console.log('Split files: ' + files.length);157  console.log('CORS + COEP/COOP headers enabled for wllama multi-threading');158});159