@ruvector/ruvllm 4.0.0 → 4.1.0

package/bin/cli.js

@@ -0,0 +1,1066 @@
+#!/usr/bin/env node
+/**
+ * RuvLLM CLI - Self-learning LLM orchestration
+ *
+ * Usage:
+ *   ruvllm query "What is machine learning?" --model ./model.gguf
+ *   ruvllm generate "Write a haiku about AI" --model ./model.gguf
+ *   ruvllm memory add "Important context"
+ *   ruvllm memory search "context"
+ *   ruvllm models list
+ *   ruvllm models download claude-code
+ *   ruvllm stats
+ *   ruvllm benchmark
+ */
+
+// `ruvllm gpu ...` delegates to @ruvector/ruvllm-gpu-loop before dist/ (and the
+// native binding) is loaded, so it works even where the runtime cannot load.
+if (process.argv[2] === 'gpu') {
+  require('./gpu-delegate.js').runGpuCommand(process.argv.slice(3));
+  return;
+}
+
+const { RuvLLM, SimdOps, version, hasSimdSupport, ModelDownloader, listModels, getModelInfo, RUVLTRA_MODELS, getDefaultModelsDir, runRoutingBenchmark, formatRoutingResults, baselineKeywordRouter, runEmbeddingBenchmark, formatEmbeddingResults, runFullBenchmark, formatFullResults, ROUTING_TEST_CASES, runFullComparison, formatComparisonResults, ContrastiveTrainer, AGENT_TRAINING_DATA, generateTrainingDataset, generateContrastivePairs, getDatasetStats, tripletLoss, cosineSimilarity } = require('../dist/cjs/index.js');
+
+const args = process.argv.slice(2);
+const command = args[0];
+
+// Parse CLI arguments
+function parseArgs(args) {
+  const result = { flags: {}, positional: [] };
+  for (let i = 0; i < args.length; i++) {
+    const arg = args[i];
+    if (arg.startsWith('--')) {
+      const key = arg.slice(2);
+      const nextArg = args[i + 1];
+      if (nextArg && !nextArg.startsWith('--')) {
+        result.flags[key] = nextArg;
+        i++;
+      } else {
+        result.flags[key] = true;
+      }
+    } else if (!result.command) {
+      result.command = arg;
+    } else {
+      result.positional.push(arg);
+    }
+  }
+  return result;
+}
+
+// Format output
+function formatJson(obj) {
+  return JSON.stringify(obj, null, 2);
+}
+
+function formatTable(data) {
+  const maxKeyLen = Math.max(...Object.keys(data).map(k => k.length));
+  return Object.entries(data)
+    .map(([k, v]) => `  ${k.padEnd(maxKeyLen)} : ${v}`)
+    .join('\n');
+}
+
+// Commands
+function generationConfig(flags) {
+  const config = {};
+  if (flags.temperature) config.temperature = parseFloat(flags.temperature);
+  if (flags['max-tokens']) config.maxTokens = parseInt(flags['max-tokens']);
+  if (flags['top-p']) config.topP = parseFloat(flags['top-p']);
+  if (flags['top-k']) config.topK = parseInt(flags['top-k']);
+  return config;
+}
+
+// query/generate need a model: without --model they fail with
+// RUVLLM_NO_LANGUAGE_MODEL (and the CLI exits 1).
+function loadModelFromFlags(llm, flags) {
+  if (!flags.model || flags.model === true) return;
+  const options = {};
+  if (flags['max-context']) options.maxContext = parseInt(flags['max-context']);
+  const info = llm.loadModel(flags.model, options);
+  if (!flags.json) console.error(`Loaded ${info.name} (${info.architecture}, ${info.quantization ?? 'quantization unknown'})`);
+}
+
+async function runQuery(llm, text, flags) {
+  loadModelFromFlags(llm, flags);
+  const response = llm.query(text, generationConfig(flags));
+
+  if (flags.json) {
+    console.log(formatJson(response));
+  } else {
+    console.log('\n' + response.text);
+    console.log(`\n--- Route: ${response.model} | Confidence: ${(response.confidence * 100).toFixed(1)}% | Tokens: ${response.promptTokens}+${response.completionTokens} (${response.finishReason}) | Latency: ${response.latencyMs.toFixed(2)}ms ---`);
+  }
+}
+
+async function runGenerate(llm, prompt, flags) {
+  loadModelFromFlags(llm, flags);
+  const config = generationConfig(flags);
+
+  if (flags.json) {
+    console.log(formatJson(llm.generateDetailed(prompt, config)));
+  } else {
+    console.log(llm.generate(prompt, config));
+  }
+}
+
+async function runMemoryAdd(llm, content, flags) {
+  const metadata = flags.metadata ? JSON.parse(flags.metadata) : undefined;
+  const id = llm.addMemory(content, metadata);
+  console.log(`Added memory with ID: ${id}`);
+}
+
+async function runMemorySearch(llm, query, flags) {
+  const k = flags.k ? parseInt(flags.k) : 10;
+  const results = llm.searchMemory(query, k);
+
+  if (flags.json) {
+    console.log(formatJson(results));
+  } else {
+    if (results.length === 0) {
+      console.log('No results found.');
+      return;
+    }
+    results.forEach((r, i) => {
+      console.log(`\n[${i + 1}] Score: ${r.score.toFixed(4)} | ID: ${r.id}`);
+      console.log(`    ${r.content.slice(0, 100)}${r.content.length > 100 ? '...' : ''}`);
+    });
+  }
+}
+
+async function runStats(llm, flags) {
+  let stats;
+  try {
+    stats = llm.stats();
+  } catch (e) {
+    stats = null;
+  }
+
+  if (!stats) {
+    if (flags.json) {
+      console.log(formatJson({ error: 'No inference data available' }));
+    } else {
+      console.log('\nNo inference data available. Run some queries first.');
+    }
+    return;
+  }
+
+  if (flags.json) {
+    console.log(formatJson(stats));
+  } else {
+    console.log('\nRuvLLM Statistics:');
+    console.log(formatTable({
+      'Total Queries': stats.totalQueries ?? 0,
+      'Memory Nodes': stats.memoryNodes ?? 0,
+      'Patterns Learned': stats.patternsLearned ?? 0,
+      'Avg Latency': `${(stats.avgLatencyMs ?? 0).toFixed(2)}ms`,
+      'Cache Hit Rate': `${((stats.cacheHitRate ?? 0) * 100).toFixed(1)}%`,
+      'Router Accuracy': `${((stats.routerAccuracy ?? 0) * 100).toFixed(1)}%`,
+    }));
+  }
+}
+
+async function runRoute(llm, text, flags) {
+  const decision = llm.route(text);
+
+  if (flags.json) {
+    console.log(formatJson(decision));
+  } else {
+    console.log('\nRouting Decision:');
+    console.log(formatTable({
+      'Model': decision.model,
+      'Context Size': decision.contextSize,
+      'Temperature': decision.temperature.toFixed(2),
+      'Top-P': decision.topP.toFixed(2),
+      'Confidence': `${(decision.confidence * 100).toFixed(1)}%`,
+    }));
+  }
+}
+
+async function runEmbed(llm, text, flags) {
+  const embedding = llm.embed(text);
+
+  if (flags.json) {
+    console.log(formatJson({ embedding, dimensions: embedding.length }));
+  } else {
+    console.log(`Embedding (${embedding.length} dimensions):`);
+    console.log(`  First 10: [${embedding.slice(0, 10).map(x => x.toFixed(4)).join(', ')}...]`);
+    console.log(`  Norm: ${Math.sqrt(embedding.reduce((s, x) => s + x * x, 0)).toFixed(4)}`);
+  }
+}
+
+async function runSimilarity(llm, text1, text2, flags) {
+  const score = llm.similarity(text1, text2);
+
+  if (flags.json) {
+    console.log(formatJson({ text1, text2, similarity: score }));
+  } else {
+    console.log(`Similarity: ${(score * 100).toFixed(2)}%`);
+  }
+}
+
+async function runBenchmark(flags) {
+  const simd = new SimdOps();
+  const dims = flags.dims ? parseInt(flags.dims) : 768;
+  const iterations = flags.iterations ? parseInt(flags.iterations) : 1000;
+
+  // Generate test vectors
+  const a = Array.from({ length: dims }, () => Math.random());
+  const b = Array.from({ length: dims }, () => Math.random());
+
+  console.log(`\nBenchmark: ${dims} dimensions, ${iterations} iterations`);
+  console.log(`SIMD: ${simd.isNative() ? 'Native' : 'JavaScript fallback'}`);
+  console.log(`Capabilities: ${simd.capabilities().join(', ')}`);
+  console.log('');
+
+  // Dot product benchmark
+  let start = Date.now();
+  for (let i = 0; i < iterations; i++) {
+    simd.dotProduct(a, b);
+  }
+  let elapsed = Date.now() - start;
+  console.log(`Dot Product:        ${elapsed}ms (${(iterations / elapsed * 1000).toFixed(0)} ops/sec)`);
+
+  // Cosine similarity benchmark
+  start = Date.now();
+  for (let i = 0; i < iterations; i++) {
+    simd.cosineSimilarity(a, b);
+  }
+  elapsed = Date.now() - start;
+  console.log(`Cosine Similarity:  ${elapsed}ms (${(iterations / elapsed * 1000).toFixed(0)} ops/sec)`);
+
+  // L2 distance benchmark
+  start = Date.now();
+  for (let i = 0; i < iterations; i++) {
+    simd.l2Distance(a, b);
+  }
+  elapsed = Date.now() - start;
+  console.log(`L2 Distance:        ${elapsed}ms (${(iterations / elapsed * 1000).toFixed(0)} ops/sec)`);
+
+  // Softmax benchmark
+  start = Date.now();
+  for (let i = 0; i < iterations; i++) {
+    simd.softmax(a);
+  }
+  elapsed = Date.now() - start;
+  console.log(`Softmax:            ${elapsed}ms (${(iterations / elapsed * 1000).toFixed(0)} ops/sec)`);
+}
+
+async function runInfo(flags) {
+  const llm = new RuvLLM();
+
+  const info = {
+    version: version(),
+    native: llm.isNativeLoaded(),
+    simd: hasSimdSupport(),
+    capabilities: llm.simdCapabilities(),
+    platform: process.platform,
+    arch: process.arch,
+    nodeVersion: process.version,
+  };
+
+  if (flags.json) {
+    console.log(formatJson(info));
+  } else {
+    console.log('\nRuvLLM Info:');
+    console.log(formatTable({
+      'Version': info.version,
+      'Native Module': info.native ? 'Loaded' : 'Fallback (JS)',
+      'SIMD Support': info.simd ? 'Yes' : 'No',
+      'Capabilities': info.capabilities.join(', '),
+      'Platform': `${info.platform}-${info.arch}`,
+      'Node.js': info.nodeVersion,
+    }));
+  }
+}
+
+// Model management commands
+async function runModelsList(flags) {
+  const downloader = new ModelDownloader();
+  const status = downloader.getStatus();
+
+  if (flags.json) {
+    console.log(formatJson(status));
+  } else {
+    console.log('\n╔══════════════════════════════════════════════════════════════════════════╗');
+    console.log('║                         RuvLTRA Models                                   ║');
+    console.log('║         https://huggingface.co/ruv/ruvltra                               ║');
+    console.log('╠══════════════════════════════════════════════════════════════════════════╣');
+    console.log('║  Model        │ Size    │ Params │ Status     │ Use Case                 ║');
+    console.log('╠══════════════════════════════════════════════════════════════════════════╣');
+
+    for (const { model, downloaded } of status) {
+      const statusIcon = downloaded ? '✓ Ready  ' : '○ Not DL ';
+      const name = model.id.padEnd(12);
+      const size = model.size.padEnd(7);
+      const params = model.parameters.padEnd(6);
+      const useCase = model.useCase.slice(0, 24).padEnd(24);
+      console.log(`║  ${name} │ ${size} │ ${params} │ ${statusIcon} │ ${useCase} ║`);
+    }
+
+    console.log('╚══════════════════════════════════════════════════════════════════════════╝');
+    console.log(`\nModels directory: ${getDefaultModelsDir()}`);
+    console.log('\nDownload with: ruvllm models download <model-id>');
+    console.log('  Examples:    ruvllm models download claude-code');
+    console.log('               ruvllm models download --all');
+  }
+}
+
+async function runModelsDownload(modelId, flags) {
+  const downloader = new ModelDownloader();
+
+  if (flags.all) {
+    console.log('\nDownloading all RuvLTRA models...\n');
+    const models = listModels();
+
+    for (const model of models) {
+      console.log(`\n[${model.id}] ${model.name} (${model.size})`);
+
+      if (downloader.isDownloaded(model.id) && !flags.force) {
+        console.log('  Already downloaded, skipping (use --force to re-download)');
+        continue;
+      }
+
+      try {
+        const lastPercent = { value: -1 };
+        const path = await downloader.download(model.id, {
+          force: flags.force,
+          onProgress: (p) => {
+            const percent = Math.floor(p.percent / 5) * 5; // Round to 5%
+            if (percent !== lastPercent.value) {
+              const bar = '█'.repeat(percent / 5) + '░'.repeat(20 - percent / 5);
+              const speed = (p.speedBps / 1024 / 1024).toFixed(1);
+              const eta = p.etaSeconds < 60
+                ? `${Math.ceil(p.etaSeconds)}s`
+                : `${Math.ceil(p.etaSeconds / 60)}m`;
+              process.stdout.write(`\r  [${bar}] ${p.percent}% | ${speed} MB/s | ETA: ${eta}   `);
+              lastPercent.value = percent;
+            }
+          },
+        });
+        console.log(`\n  ✓ Downloaded to: ${path}`);
+      } catch (error) {
+        console.error(`\n  ✗ Failed: ${error.message}`);
+      }
+    }
+
+    console.log('\n\nDownload complete!');
+    return;
+  }
+
+  if (!modelId) {
+    console.error('Error: model ID required. Use --all to download all models.');
+    console.error('\nAvailable models:');
+    listModels().forEach(m => console.error(`  - ${m.id}: ${m.name} (${m.size})`));
+    process.exit(1);
+  }
+
+  const model = getModelInfo(modelId);
+  if (!model) {
+    console.error(`Error: Unknown model "${modelId}"`);
+    console.error('\nAvailable models:');
+    listModels().forEach(m => console.error(`  - ${m.id}: ${m.name} (${m.size})`));
+    process.exit(1);
+  }
+
+  console.log(`\nDownloading ${model.name} (${model.size})...`);
+  console.log(`From: ${model.url}\n`);
+
+  if (downloader.isDownloaded(modelId) && !flags.force) {
+    const path = downloader.getModelPath(modelId);
+    console.log(`Model already downloaded at: ${path}`);
+    console.log('Use --force to re-download.');
+    return;
+  }
+
+  const lastPercent = { value: -1 };
+  try {
+    const path = await downloader.download(modelId, {
+      force: flags.force,
+      onProgress: (p) => {
+        const percent = Math.floor(p.percent / 2) * 2; // Round to 2%
+        if (percent !== lastPercent.value) {
+          const bar = '█'.repeat(Math.floor(percent / 5)) + '░'.repeat(20 - Math.floor(percent / 5));
+          const downloaded = (p.downloaded / 1024 / 1024).toFixed(1);
+          const total = (p.total / 1024 / 1024).toFixed(1);
+          const speed = (p.speedBps / 1024 / 1024).toFixed(1);
+          const eta = p.etaSeconds < 60
+            ? `${Math.ceil(p.etaSeconds)}s`
+            : `${Math.ceil(p.etaSeconds / 60)}m`;
+          process.stdout.write(`\r[${bar}] ${p.percent}% | ${downloaded}/${total} MB | ${speed} MB/s | ETA: ${eta}   `);
+          lastPercent.value = percent;
+        }
+      },
+    });
+    console.log(`\n\n✓ Downloaded to: ${path}`);
+    console.log(`\nModel ready to use!`);
+    console.log(`  Context length: ${model.contextLength} tokens`);
+    console.log(`  Quantization:   ${model.quantization}`);
+  } catch (error) {
+    console.error(`\n\n✗ Download failed: ${error.message}`);
+    process.exit(1);
+  }
+}
+
+async function runModelsStatus(flags) {
+  const downloader = new ModelDownloader();
+  const status = downloader.getStatus();
+
+  if (flags.json) {
+    console.log(formatJson(status.map(s => ({
+      id: s.model.id,
+      name: s.model.name,
+      downloaded: s.downloaded,
+      path: s.path,
+      size: s.model.size,
+    }))));
+  } else {
+    console.log('\nModel Status:');
+    console.log(`Directory: ${getDefaultModelsDir()}\n`);
+
+    for (const { model, downloaded, path } of status) {
+      const icon = downloaded ? '✓' : '○';
+      const status = downloaded ? 'Ready' : 'Not downloaded';
+      console.log(`  ${icon} ${model.name.padEnd(25)} ${status.padEnd(15)} ${model.size}`);
+      if (downloaded) {
+        console.log(`    Path: ${path}`);
+      }
+    }
+  }
+}
+
+async function runModelsDelete(modelId, flags) {
+  const downloader = new ModelDownloader();
+
+  if (flags.all) {
+    const count = downloader.deleteAll();
+    console.log(`Deleted ${count} model(s).`);
+    return;
+  }
+
+  if (!modelId) {
+    console.error('Error: model ID required. Use --all to delete all models.');
+    process.exit(1);
+  }
+
+  if (downloader.delete(modelId)) {
+    console.log(`Deleted model: ${modelId}`);
+  } else {
+    console.log(`Model not found or not downloaded: ${modelId}`);
+  }
+}
+
+// Benchmark commands for Claude Code use cases
+async function runBenchmarkRouting(flags) {
+  console.log('\nRunning Routing Benchmark...');
+  console.log(`Testing ${ROUTING_TEST_CASES.length} task routing scenarios\n`);
+
+  const llm = new RuvLLM({ embeddingDim: 768, learningEnabled: false });
+
+  // Router function using the model
+  const modelRouter = (task) => {
+    try {
+      const decision = llm.route(task);
+      return { agent: decision.model, confidence: decision.confidence };
+    } catch (e) {
+      // Fallback to keyword router if model not available
+      return baselineKeywordRouter(task);
+    }
+  };
+
+  // Run with baseline (keyword) router
+  console.log('Baseline (keyword matching):');
+  const baselineResults = runRoutingBenchmark(baselineKeywordRouter);
+
+  if (flags.json) {
+    console.log(JSON.stringify(baselineResults, null, 2));
+  } else {
+    console.log(formatRoutingResults(baselineResults));
+  }
+
+  // Try model router if native is available
+  if (llm.isNativeLoaded() && !flags['baseline-only']) {
+    console.log('\nModel Router (RuvLTRA):');
+    const modelResults = runRoutingBenchmark(modelRouter);
+    if (flags.json) {
+      console.log(JSON.stringify(modelResults, null, 2));
+    } else {
+      console.log(formatRoutingResults(modelResults));
+    }
+
+    // Comparison
+    const improvement = modelResults.accuracy - baselineResults.accuracy;
+    console.log(`\nComparison: Model ${improvement >= 0 ? '+' : ''}${(improvement * 100).toFixed(1)}% vs baseline`);
+  }
+}
+
+async function runBenchmarkEmbedding(flags) {
+  console.log('\nRunning Embedding Benchmark...');
+  console.log('Testing similarity detection, clustering, and search quality\n');
+
+  const llm = new RuvLLM({ embeddingDim: 768, learningEnabled: false });
+  const simd = new SimdOps();
+
+  // Embedder function
+  const embedder = (text) => {
+    try {
+      return llm.embed(text);
+    } catch (e) {
+      // Fallback: simple hash-based embedding for testing
+      const hash = text.split('').reduce((h, c) => ((h << 5) - h + c.charCodeAt(0)) | 0, 0);
+      return Array.from({ length: 768 }, (_, i) => Math.sin(hash + i) * 0.5);
+    }
+  };
+
+  // Similarity function
+  const similarity = (a, b) => {
+    try {
+      return simd.cosineSimilarity(a, b);
+    } catch (e) {
+      // Fallback cosine similarity
+      let dot = 0, normA = 0, normB = 0;
+      for (let i = 0; i < a.length; i++) {
+        dot += a[i] * b[i];
+        normA += a[i] * a[i];
+        normB += b[i] * b[i];
+      }
+      return dot / (Math.sqrt(normA) * Math.sqrt(normB));
+    }
+  };
+
+  const results = runEmbeddingBenchmark(embedder, similarity);
+
+  if (flags.json) {
+    console.log(JSON.stringify(results, null, 2));
+  } else {
+    console.log(formatEmbeddingResults(results));
+  }
+}
+
+async function runBenchmarkCompare(flags) {
+  try {
+    const results = await runFullComparison();
+
+    if (flags.json) {
+      console.log(JSON.stringify(results, null, 2));
+    } else {
+      console.log(formatComparisonResults(results));
+    }
+  } catch (error) {
+    console.error('Comparison failed:', error.message);
+    process.exit(1);
+  }
+}
+
+// Training commands
+async function runTrainContrastive(flags) {
+  console.log('\n╔═══════════════════════════════════════════════════════════════════════════╗');
+  console.log('║           RuvLTRA Contrastive Fine-tuning Pipeline                        ║');
+  console.log('╚═══════════════════════════════════════════════════════════════════════════╝\n');
+
+  const stats = getDatasetStats();
+  const outputPath = flags.output || './training-output';
+
+  console.log(`Training examples: ${stats.totalExamples}`);
+  console.log(`Contrastive pairs: ${stats.contrastivePairs}`);
+  console.log(`Agent types: ${stats.agentTypes}`);
+  console.log(`Output: ${outputPath}\n`);
+
+  // Create trainer
+  const trainer = new ContrastiveTrainer({
+    epochs: parseInt(flags.epochs) || 10,
+    batchSize: parseInt(flags['batch-size']) || 16,
+    learningRate: parseFloat(flags.lr) || 0.0001,
+    margin: parseFloat(flags.margin) || 0.5,
+    outputPath,
+  });
+
+  // Get embeddings for agents
+  const llm = new RuvLLM({ embeddingDim: 768, learningEnabled: false });
+  const agentEmbeddings = {};
+
+  console.log('Computing agent embeddings...');
+  for (const [agent, data] of Object.entries(AGENT_TRAINING_DATA)) {
+    process.stdout.write(`  ${agent}... `);
+    try {
+      const emb = llm.embed(data.description);
+      agentEmbeddings[agent] = emb;
+      trainer.addAgentEmbedding(agent, emb);
+      console.log('done');
+    } catch (e) {
+      // Fallback embedding
+      const hash = data.description.split('').reduce((h, c) => ((h << 5) - h + c.charCodeAt(0)) | 0, 0);
+      const emb = Array.from({ length: 768 }, (_, i) => Math.sin(hash + i) * 0.5);
+      agentEmbeddings[agent] = emb;
+      trainer.addAgentEmbedding(agent, emb);
+      console.log('done (fallback)');
+    }
+  }
+
+  // Generate triplets
+  console.log('\nGenerating training triplets...');
+  const examples = generateTrainingDataset();
+  const agents = Object.keys(AGENT_TRAINING_DATA);
+
+  let tripletCount = 0;
+  for (const example of examples.slice(0, 200)) { // Limit for speed
+    let taskEmb;
+    try {
+      taskEmb = llm.embed(example.task);
+    } catch (e) {
+      const hash = example.task.split('').reduce((h, c) => ((h << 5) - h + c.charCodeAt(0)) | 0, 0);
+      taskEmb = Array.from({ length: 768 }, (_, i) => Math.sin(hash + i) * 0.5);
+    }
+
+    const positiveAgent = example.agent;
+    const positiveEmb = agentEmbeddings[positiveAgent];
+
+    // Hard negatives
+    const hardNegatives = example.confusing_with
+      ? [example.confusing_with]
+      : agents.filter(a => a !== positiveAgent).slice(0, 2);
+
+    for (const negAgent of hardNegatives) {
+      const negEmb = agentEmbeddings[negAgent];
+      if (negEmb) {
+        trainer.addTriplet(
+          example.task, taskEmb,
+          positiveAgent, positiveEmb,
+          negAgent, negEmb,
+          !!example.confusing_with
+        );
+        tripletCount++;
+      }
+    }
+  }
+
+  console.log(`Created ${tripletCount} triplets\n`);
+
+  // Train
+  console.log('Training...');
+  const result = trainer.train();
+
+  // Export data
+  console.log('\nExporting training data...');
+  const exportPath = trainer.exportTrainingData();
+  const loraConfig = trainer.generateLoRAConfig();
+  const scriptPath = trainer.generateTrainingScript();
+
+  // Summary
+  console.log('\n═══════════════════════════════════════════════════════════════════════════');
+  console.log('                              TRAINING SUMMARY');
+  console.log('═══════════════════════════════════════════════════════════════════════════\n');
+
+  console.log('Training data exported:');
+  console.log(`  - ${exportPath}/triplets.jsonl (${tripletCount} triplets)`);
+  console.log(`  - ${exportPath}/triplets.csv (spreadsheet format)`);
+  console.log(`  - ${exportPath}/embeddings.json (precomputed embeddings)`);
+  console.log(`  - ${exportPath}/lora_config.json (LoRA configuration)`);
+  console.log(`  - ${exportPath}/train.sh (training script)\n`);
+
+  console.log('Training loss (simulated):');
+  console.log(`  Initial: ${result.initialLoss.toFixed(4)}`);
+  console.log(`  Final:   ${result.finalLoss.toFixed(4)}`);
+  console.log(`  Improvement: ${result.improvement.toFixed(1)}%\n`);
+
+  console.log('To fine-tune on GPU:');
+  console.log(`  cd ${exportPath}`);
+  console.log('  chmod +x train.sh && ./train.sh\n');
+}
+
+async function runTrainDataset(flags) {
+  console.log('\n╔═══════════════════════════════════════════════════════════════════════════╗');
+  console.log('║               RuvLTRA Training Dataset Generator                          ║');
+  console.log('╚═══════════════════════════════════════════════════════════════════════════╝\n');
+
+  const stats = getDatasetStats();
+  const outputPath = flags.output || './data/training';
+
+  console.log(`Generating training data for ${stats.agentTypes} agent types...\n`);
+
+  // Generate dataset
+  const examples = generateTrainingDataset();
+  const pairs = generateContrastivePairs();
+
+  // Export
+  const { existsSync, mkdirSync, writeFileSync } = require('fs');
+  const { join } = require('path');
+
+  if (!existsSync(outputPath)) {
+    mkdirSync(outputPath, { recursive: true });
+  }
+
+  // JSONL format
+  writeFileSync(
+    join(outputPath, 'routing-examples.jsonl'),
+    examples.map(e => JSON.stringify(e)).join('\n')
+  );
+
+  // Contrastive pairs
+  writeFileSync(
+    join(outputPath, 'contrastive-pairs.jsonl'),
+    pairs.map(p => JSON.stringify(p)).join('\n')
+  );
+
+  // CSV format
+  const csv = [
+    'task,agent,complexity,confusing_with',
+    ...examples.map(e => `"${e.task.replace(/"/g, '""')}",${e.agent},${e.complexity || ''},${e.confusing_with || ''}`)
+  ].join('\n');
+  writeFileSync(join(outputPath, 'routing-examples.csv'), csv);
+
+  console.log('Export complete:');
+  console.log(`  - ${join(outputPath, 'routing-examples.jsonl')} (${examples.length} examples)`);
+  console.log(`  - ${join(outputPath, 'contrastive-pairs.jsonl')} (${pairs.length} pairs)`);
+  console.log(`  - ${join(outputPath, 'routing-examples.csv')} (spreadsheet format)\n`);
+
+  console.log('Dataset statistics:');
+  console.log(`  Total examples: ${stats.totalExamples}`);
+  console.log(`  Contrastive pairs: ${stats.contrastivePairs}`);
+  console.log(`  Agent types: ${stats.agentTypes}`);
+  console.log(`  Agents: ${stats.agents.join(', ')}\n`);
+}
+
+async function runTrainStats(flags) {
+  const stats = getDatasetStats();
+
+  if (flags.json) {
+    console.log(JSON.stringify(stats, null, 2));
+  } else {
+    console.log('\nTraining Dataset Statistics:');
+    console.log(`  Total examples: ${stats.totalExamples}`);
+    console.log(`  Contrastive pairs: ${stats.contrastivePairs}`);
+    console.log(`  Agent types: ${stats.agentTypes}`);
+    console.log('\nAgents:');
+    for (const agent of stats.agents) {
+      const data = AGENT_TRAINING_DATA[agent];
+      console.log(`  - ${agent}: ${data.examples.length} examples, ${data.keywords.length} keywords`);
+    }
+  }
+}
+
+async function runBenchmarkFull(flags) {
+  console.log('\n╔═══════════════════════════════════════════════════════════════════════════╗');
+  console.log('║                    RUVLTRA FULL BENCHMARK SUITE                           ║');
+  console.log('║            Evaluating for Claude Code Use Cases                           ║');
+  console.log('╚═══════════════════════════════════════════════════════════════════════════╝\n');
+
+  const llm = new RuvLLM({ embeddingDim: 768, learningEnabled: false });
+  const simd = new SimdOps();
+
+  const modelName = llm.isNativeLoaded() ? 'RuvLTRA (native)' : 'RuvLTRA (JS fallback)';
+
+  // Router
+  const router = (task) => {
+    try {
+      const decision = llm.route(task);
+      return { agent: decision.model, confidence: decision.confidence };
+    } catch (e) {
+      return baselineKeywordRouter(task);
+    }
+  };
+
+  // Embedder
+  const embedder = (text) => {
+    try {
+      return llm.embed(text);
+    } catch (e) {
+      const hash = text.split('').reduce((h, c) => ((h << 5) - h + c.charCodeAt(0)) | 0, 0);
+      return Array.from({ length: 768 }, (_, i) => Math.sin(hash + i) * 0.5);
+    }
+  };
+
+  // Similarity
+  const similarity = (a, b) => {
+    try {
+      return simd.cosineSimilarity(a, b);
+    } catch (e) {
+      let dot = 0, normA = 0, normB = 0;
+      for (let i = 0; i < a.length; i++) {
+        dot += a[i] * b[i];
+        normA += a[i] * a[i];
+        normB += b[i] * b[i];
+      }
+      return dot / (Math.sqrt(normA) * Math.sqrt(normB));
+    }
+  };
+
+  const results = runFullBenchmark(router, embedder, similarity, modelName);
+
+  if (flags.json) {
+    console.log(JSON.stringify(results, null, 2));
+  } else {
+    console.log(formatFullResults(results));
+  }
+}
+
+function printHelp() {
+  console.log(`
+RuvLLM - Self-learning LLM Orchestration
+
+Usage: ruvllm <command> [options]
+
+Commands:
+  query <text> --model <gguf>     Route, then answer with the model (chat template)
+  generate <prompt> --model <gguf>
+                            Continue a raw prompt with the model
+  route <text>              Get routing decision for query
+  memory add <content>      Add content to memory
+  memory search <query>     Search memory for similar content
+  embed <text>              Get embedding for text
+  similarity <t1> <t2>      Compute similarity between texts
+  stats                     Show engine statistics
+  benchmark                 Run SIMD performance benchmark
+  info                      Show system information
+  serve <gguf>              HTTP server: /health, /query, /v1/chat/completions
+                            (--host 127.0.0.1 --port 3000; needs a loadable model)
+  help                      Show this help message
+
+GPU job loop (via @ruvector/ruvllm-gpu-loop, Node 22+):
+  gpu run --config <abs> --plan <abs>
+                            Bounded weight-edit experiment (4.0.x "ruvllm run")
+  gpu fixture --dir <abs> --binary <abs>
+                            CPU fixture experiment (4.0.x "ruvllm fixture")
+  gpu loop status|tick|run --config <abs>
+                            Scheduled GPU job loop (same as ruvllm-gpu-loop)
+  gpu adapters export|training-plan|select
+                            Adapter archives (same as ruvllm-adapters)
+
+Model Management:
+  models list               List available RuvLTRA models
+  models download <id>      Download a model from HuggingFace
+  models download --all     Download all available models
+  models status             Check which models are downloaded
+  models delete <id>        Delete a downloaded model
+
+Claude Code Benchmarks:
+  benchmark routing         Test agent routing accuracy (100 tasks)
+  benchmark embedding       Test embedding quality (similarity, search, clustering)
+  benchmark full            Run complete benchmark suite
+  benchmark compare         Compare Qwen base vs RuvLTRA Claude Code
+  benchmark simd            Run SIMD performance benchmark
+
+Training & Fine-tuning:
+  train contrastive         Run contrastive fine-tuning pipeline
+  train dataset             Generate training dataset (JSONL, CSV)
+  train stats               Show training dataset statistics
+
+Options:
+  --json                    Output as JSON
+  --model <path>            GGUF file to load for query/generate (required for them;
+                            needs a native binary with model support, see README)
+  --max-context <int>       Context window for --model (default: 4096)
+  --temperature <float>     Sampling temperature (0 = greedy)
+  --max-tokens <int>        Maximum tokens to generate
+  --top-p <float>           Nucleus sampling (0.0-1.0)
+  --top-k <int>             Top-k sampling
+  --k <int>                 Number of results for search
+  --metadata <json>         Metadata for memory add
+  --dims <int>              Dimensions for benchmark (default: 768)
+  --iterations <int>        Iterations for benchmark (default: 1000)
+  --force                   Force re-download even if model exists
+  --all                     Apply to all models (download/delete)
+  --output <path>           Output directory for training data
+  --epochs <int>            Number of training epochs (default: 10)
+  --batch-size <int>        Training batch size (default: 16)
+  --lr <float>              Learning rate (default: 0.0001)
+  --margin <float>          Triplet loss margin (default: 0.5)
+
+Available Models (from https://huggingface.co/ruv/ruvltra):
+  claude-code               RuvLTRA Claude Code (398MB) - Claude Code workflows
+  small                     RuvLTRA Small (398MB) - Edge devices, IoT
+  medium                    RuvLTRA Medium (669MB) - General purpose
+
+Examples:
+  ruvllm query "What is machine learning?" --model ./qwen2.5-0.5b-instruct-q4_k_m.gguf
+  ruvllm generate "Once upon a time" --model ./model.gguf --max-tokens 64 --json
+  ruvllm memory add "Important context" --metadata '{"type":"note"}'
+  ruvllm memory search "context" --k 5
+  ruvllm similarity "hello world" "hi there"
+  ruvllm benchmark --dims 1024 --iterations 5000
+  ruvllm serve ./qwen2.5-0.5b-instruct-q4_k_m.gguf --port 3000
+  ruvllm gpu loop status --config /abs/private/gpu-loop.json
+
+  # Model management
+  ruvllm models list
+  ruvllm models download claude-code
+  ruvllm models download --all
+  ruvllm models status
+
+  # Claude Code benchmarks
+  ruvllm benchmark routing            # Test task routing accuracy
+  ruvllm benchmark embedding          # Test embedding quality
+  ruvllm benchmark full               # Run complete benchmark suite
+
+  # Training & fine-tuning
+  ruvllm train contrastive            # Run contrastive fine-tuning
+  ruvllm train dataset                # Generate training dataset
+  ruvllm train stats                  # Show dataset statistics
+  ruvllm train contrastive --epochs 20 --output ./my-training
+
+Learn more: https://github.com/ruvnet/ruvector
+`);
+}
+
+// Main
+async function main() {
+  const parsed = parseArgs(args);
+  const { command, positional, flags } = parsed;
+
+  if (!command || command === 'help' || flags.help) {
+    printHelp();
+    return;
+  }
+
+  // Create engine for commands that need it
+  const llm = new RuvLLM({
+    embeddingDim: flags.dim ? parseInt(flags.dim) : 768,
+    learningEnabled: flags['no-learning'] ? false : true,
+  });
+
+  try {
+    switch (command) {
+      case 'query':
+        if (!positional[0]) {
+          console.error('Error: query text required');
+          process.exit(1);
+        }
+        await runQuery(llm, positional[0], flags);
+        break;
+
+      case 'generate':
+        if (!positional[0]) {
+          console.error('Error: prompt required');
+          process.exit(1);
+        }
+        await runGenerate(llm, positional[0], flags);
+        break;
+
+      case 'route':
+        if (!positional[0]) {
+          console.error('Error: text required');
+          process.exit(1);
+        }
+        await runRoute(llm, positional[0], flags);
+        break;
+
+      case 'memory':
+        const subcommand = positional[0];
+        if (subcommand === 'add') {
+          if (!positional[1]) {
+            console.error('Error: content required');
+            process.exit(1);
+          }
+          await runMemoryAdd(llm, positional[1], flags);
+        } else if (subcommand === 'search') {
+          if (!positional[1]) {
+            console.error('Error: query required');
+            process.exit(1);
+          }
+          await runMemorySearch(llm, positional[1], flags);
+        } else {
+          console.error('Error: unknown memory subcommand. Use "add" or "search"');
+          process.exit(1);
+        }
+        break;
+
+      case 'embed':
+        if (!positional[0]) {
+          console.error('Error: text required');
+          process.exit(1);
+        }
+        await runEmbed(llm, positional[0], flags);
+        break;
+
+      case 'similarity':
+        if (!positional[0] || !positional[1]) {
+          console.error('Error: two texts required');
+          process.exit(1);
+        }
+        await runSimilarity(llm, positional[0], positional[1], flags);
+        break;
+
+      case 'stats':
+        await runStats(llm, flags);
+        break;
+
+      case 'benchmark':
+        const benchSubcmd = positional[0];
+        if (benchSubcmd === 'routing') {
+          await runBenchmarkRouting(flags);
+        } else if (benchSubcmd === 'embedding' || benchSubcmd === 'embeddings') {
+          await runBenchmarkEmbedding(flags);
+        } else if (benchSubcmd === 'full' || benchSubcmd === 'all') {
+          await runBenchmarkFull(flags);
+        } else if (benchSubcmd === 'compare') {
+          await runBenchmarkCompare(flags);
+        } else if (benchSubcmd === 'simd' || !benchSubcmd) {
+          // Default to SIMD benchmark for backwards compatibility
+          await runBenchmark(flags);
+        } else {
+          console.error(`Unknown benchmark type: ${benchSubcmd}`);
+          console.error('Available: routing, embedding, full, simd, compare');
+          process.exit(1);
+        }
+        break;
+
+      case 'info':
+        await runInfo(flags);
+        break;
+
+      case 'serve':
+        // Keeps running until SIGINT/SIGTERM; exits 1 if no model can be loaded.
+        require('./serve.js').runServe({ RuvLLM, version: require('../package.json').version }, positional, flags);
+        return;
+
+      case 'models':
+        const modelsSubcmd = positional[0];
+        if (!modelsSubcmd || modelsSubcmd === 'list') {
+          await runModelsList(flags);
+        } else if (modelsSubcmd === 'download') {
+          await runModelsDownload(positional[1], flags);
+        } else if (modelsSubcmd === 'status') {
+          await runModelsStatus(flags);
+        } else if (modelsSubcmd === 'delete' || modelsSubcmd === 'remove') {
+          await runModelsDelete(positional[1], flags);
+        } else {
+          // Treat subcommand as model ID for download
+          await runModelsDownload(modelsSubcmd, flags);
+        }
+        break;
+
+      case 'train':
+        const trainSubcmd = positional[0];
+        if (!trainSubcmd || trainSubcmd === 'contrastive') {
+          await runTrainContrastive(flags);
+        } else if (trainSubcmd === 'dataset') {
+          await runTrainDataset(flags);
+        } else if (trainSubcmd === 'stats') {
+          await runTrainStats(flags);
+        } else {
+          console.error(`Unknown train subcommand: ${trainSubcmd}`);
+          console.error('Available: contrastive, dataset, stats');
+          process.exit(1);
+        }
+        break;
+
+      default:
+        console.error(`Unknown command: ${command}`);
+        console.error('Run "ruvllm help" for usage information.');
+        process.exit(1);
+    }
+  } catch (error) {
+    console.error('Error:', error.message);
+    if (error.code === 'RUVLLM_NO_LANGUAGE_MODEL') {
+      console.error('Hint: pass --model <path to a GGUF file>');
+    }
+    if (flags.verbose) {
+      console.error(error.stack);
+    }
+    process.exit(1);
+  }
+}
+
+main().catch(err => {
+  console.error('Fatal error:', err);
+  process.exit(1);
+});