web-agent-bridge 3.0.0 → 3.3.0

package/server/services/local-ai.js

@@ -1,389 +1,389 @@
-/**
- * Local AI — Sovereign Intelligence Runtime
- *
- * Manages local AI models running on the user's own hardware.
- * Auto-detects Ollama, llama.cpp, and any OpenAI-compatible local endpoint.
- * Routes inference requests to the best available model based on capability,
- * context window, and current load.
- *
- * Supported Providers:
- *   - Ollama (http://localhost:11434)
- *   - llama.cpp server (http://localhost:8080)
- *   - Custom OpenAI-compatible endpoints
- *
- * All inference happens locally. No data leaves the device.
- */
-
-const crypto = require('crypto');
-const { db } = require('../models/db');
-
-// ─── Schema ──────────────────────────────────────────────────────────
-
-db.exec(`
-  CREATE TABLE IF NOT EXISTS local_models (
-    id TEXT PRIMARY KEY,
-    site_id TEXT NOT NULL,
-    provider TEXT NOT NULL,
-    model_name TEXT NOT NULL,
-    endpoint TEXT NOT NULL,
-    capabilities TEXT DEFAULT '["text"]',
-    context_window INTEGER DEFAULT 4096,
-    parameters TEXT DEFAULT '{}',
-    status TEXT DEFAULT 'available',
-    total_requests INTEGER DEFAULT 0,
-    total_tokens INTEGER DEFAULT 0,
-    avg_latency_ms REAL DEFAULT 0,
-    last_used TEXT,
-    last_probe TEXT,
-    created_at TEXT DEFAULT (datetime('now')),
-    UNIQUE(site_id, provider, model_name)
-  );
-
-  CREATE TABLE IF NOT EXISTS local_inference_log (
-    id TEXT PRIMARY KEY,
-    model_id TEXT NOT NULL,
-    task_type TEXT,
-    prompt_tokens INTEGER DEFAULT 0,
-    completion_tokens INTEGER DEFAULT 0,
-    latency_ms INTEGER DEFAULT 0,
-    success INTEGER DEFAULT 1,
-    created_at TEXT DEFAULT (datetime('now'))
-  );
-
-  CREATE INDEX IF NOT EXISTS idx_local_models_site ON local_models(site_id);
-  CREATE INDEX IF NOT EXISTS idx_local_models_status ON local_models(status);
-  CREATE INDEX IF NOT EXISTS idx_local_inference_model ON local_inference_log(model_id);
-`);
-
-// ─── Prepared Statements ─────────────────────────────────────────────
-
-const stmts = {
-  upsertModel: db.prepare("INSERT INTO local_models (id, site_id, provider, model_name, endpoint, capabilities, context_window, parameters) VALUES (?, ?, ?, ?, ?, ?, ?, ?) ON CONFLICT(site_id, provider, model_name) DO UPDATE SET endpoint = ?, capabilities = ?, context_window = ?, parameters = ?, status = 'available', last_probe = datetime('now')"),
-  getModel: db.prepare('SELECT * FROM local_models WHERE id = ?'),
-  getModels: db.prepare('SELECT * FROM local_models WHERE site_id = ? ORDER BY provider, model_name'),
-  getAvailableModels: db.prepare("SELECT * FROM local_models WHERE site_id = ? AND status = 'available' ORDER BY avg_latency_ms ASC"),
-  getModelsByCapability: db.prepare("SELECT * FROM local_models WHERE site_id = ? AND status = 'available' AND capabilities LIKE ? ORDER BY avg_latency_ms ASC"),
-  updateModelStatus: db.prepare('UPDATE local_models SET status = ?, last_probe = datetime(\'now\') WHERE id = ?'),
-  updateModelStats: db.prepare("UPDATE local_models SET total_requests = total_requests + 1, total_tokens = total_tokens + ?, avg_latency_ms = (avg_latency_ms * total_requests + ?) / (total_requests + 1), last_used = datetime('now') WHERE id = ?"),
-  insertLog: db.prepare('INSERT INTO local_inference_log (id, model_id, task_type, prompt_tokens, completion_tokens, latency_ms, success) VALUES (?, ?, ?, ?, ?, ?, ?)'),
-  getStats: db.prepare(`SELECT
-    (SELECT COUNT(*) FROM local_models WHERE site_id = ? AND status = 'available') as available_models,
-    (SELECT COUNT(*) FROM local_models WHERE site_id = ?) as total_models,
-    (SELECT SUM(total_requests) FROM local_models WHERE site_id = ?) as total_requests,
-    (SELECT SUM(total_tokens) FROM local_models WHERE site_id = ?) as total_tokens,
-    (SELECT AVG(avg_latency_ms) FROM local_models WHERE site_id = ? AND status = 'available') as avg_latency`),
-};
-
-// ─── Default Provider Endpoints ──────────────────────────────────────
-
-const PROVIDERS = {
-  ollama: { name: 'ollama', baseUrl: 'http://localhost:11434', tagsPath: '/api/tags', chatPath: '/api/chat', generatePath: '/api/generate' },
-  llamacpp: { name: 'llamacpp', baseUrl: 'http://localhost:8080', chatPath: '/v1/chat/completions', modelsPath: '/v1/models' },
-};
-
-// ─── Model Discovery ─────────────────────────────────────────────────
-
-/**
- * Probe local endpoints and register discovered models.
- */
-async function discoverModels(siteId, customEndpoints = []) {
-  const discovered = [];
-
-  // Probe Ollama
-  try {
-    const ollamaModels = await _probeOllama(PROVIDERS.ollama.baseUrl);
-    for (const m of ollamaModels) {
-      const result = _registerModel(siteId, 'ollama', m.name, PROVIDERS.ollama.baseUrl, m.capabilities, m.contextWindow, m.parameters);
-      discovered.push(result);
-    }
-  } catch (_) { /* Ollama not running */ }
-
-  // Probe llama.cpp
-  try {
-    const lcModels = await _probeLlamaCpp(PROVIDERS.llamacpp.baseUrl);
-    for (const m of lcModels) {
-      const result = _registerModel(siteId, 'llamacpp', m.name, PROVIDERS.llamacpp.baseUrl, m.capabilities, m.contextWindow, m.parameters);
-      discovered.push(result);
-    }
-  } catch (_) { /* llama.cpp not running */ }
-
-  // Probe custom endpoints
-  for (const ep of customEndpoints) {
-    try {
-      const models = await _probeOpenAICompatible(ep.url);
-      for (const m of models) {
-        const result = _registerModel(siteId, ep.name || 'custom', m.name, ep.url, m.capabilities, m.contextWindow, m.parameters);
-        discovered.push(result);
-      }
-    } catch (_) { /* endpoint not available */ }
-  }
-
-  return { discovered: discovered.length, models: discovered };
-}
-
-/**
- * Register a model manually.
- */
-function registerModel(siteId, provider, modelName, endpoint, capabilities = ['text'], contextWindow = 4096) {
-  return _registerModel(siteId, provider, modelName, endpoint, capabilities, contextWindow, {});
-}
-
-function _registerModel(siteId, provider, modelName, endpoint, capabilities, contextWindow, parameters) {
-  const id = crypto.randomUUID();
-  const caps = JSON.stringify(capabilities);
-  const params = JSON.stringify(parameters);
-
-  stmts.upsertModel.run(id, siteId, provider, modelName, endpoint, caps, contextWindow, params, endpoint, caps, contextWindow, params);
-  return { id, provider, modelName, endpoint, capabilities, contextWindow };
-}
-
-// ─── Inference ───────────────────────────────────────────────────────
-
-/**
- * Run inference on the best available local model.
- * @param {string} siteId
- * @param {string} prompt - The user prompt
- * @param {object} options - { capability, model, systemPrompt, temperature, maxTokens, stream }
- */
-async function infer(siteId, prompt, options = {}) {
-  const capability = options.capability || 'text';
-
-  // Select model
-  let model;
-  if (options.modelId) {
-    model = stmts.getModel.get(options.modelId);
-    if (!model || model.status !== 'available') throw new Error('Selected model unavailable');
-  } else {
-    const candidates = stmts.getModelsByCapability.all(siteId, `%${capability}%`);
-    if (candidates.length === 0) throw new Error(`No local model available for capability: ${capability}`);
-    model = candidates[0]; // Fastest by avg latency
-  }
-
-  const start = Date.now();
-  let result;
-
-  try {
-    const parsed = JSON.parse(model.parameters || '{}');
-    if (model.provider === 'ollama') {
-      result = await _inferOllama(model, prompt, options);
-    } else if (model.provider === 'llamacpp') {
-      result = await _inferLlamaCpp(model, prompt, options);
-    } else {
-      result = await _inferOpenAICompatible(model, prompt, options);
-    }
-  } catch (err) {
-    const latency = Date.now() - start;
-    stmts.insertLog.run(crypto.randomUUID(), model.id, capability, 0, 0, latency, 0);
-    throw err;
-  }
-
-  const latency = Date.now() - start;
-  const totalTokens = (result.promptTokens || 0) + (result.completionTokens || 0);
-
-  stmts.updateModelStats.run(totalTokens, latency, model.id);
-  stmts.insertLog.run(crypto.randomUUID(), model.id, capability, result.promptTokens || 0, result.completionTokens || 0, latency, 1);
-
-  return {
-    modelId: model.id,
-    provider: model.provider,
-    model: model.model_name,
-    response: result.text,
-    promptTokens: result.promptTokens || 0,
-    completionTokens: result.completionTokens || 0,
-    latencyMs: latency,
-  };
-}
-
-// ─── Model Management ────────────────────────────────────────────────
-
-function getModels(siteId) {
-  return stmts.getModels.all(siteId).map(_deserializeModel);
-}
-
-function getAvailableModels(siteId) {
-  return stmts.getAvailableModels.all(siteId).map(_deserializeModel);
-}
-
-function getModel(modelId) {
-  const row = stmts.getModel.get(modelId);
-  return row ? _deserializeModel(row) : null;
-}
-
-function updateModelStatus(modelId, status) {
-  stmts.updateModelStatus.run(status, modelId);
-}
-
-function getStats(siteId) {
-  const row = stmts.getStats.get(siteId, siteId, siteId, siteId, siteId);
-  return {
-    availableModels: row.available_models || 0,
-    totalModels: row.total_models || 0,
-    totalRequests: row.total_requests || 0,
-    totalTokens: row.total_tokens || 0,
-    avgLatency: row.avg_latency ? Math.round(row.avg_latency) : 0,
-  };
-}
-
-// ─── Provider-Specific Inference ─────────────────────────────────────
-
-async function _inferOllama(model, prompt, options) {
-  const body = {
-    model: model.model_name,
-    messages: [],
-    stream: false,
-    options: {},
-  };
-
-  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
-  body.messages.push({ role: 'user', content: prompt });
-  if (options.temperature != null) body.options.temperature = options.temperature;
-
-  const res = await fetch(`${model.endpoint}/api/chat`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify(body),
-    signal: AbortSignal.timeout(options.timeout || 120000),
-  });
-
-  if (!res.ok) throw new Error(`Ollama error: ${res.status}`);
-  const data = await res.json();
-
-  return {
-    text: data.message?.content || '',
-    promptTokens: data.prompt_eval_count || 0,
-    completionTokens: data.eval_count || 0,
-  };
-}
-
-async function _inferLlamaCpp(model, prompt, options) {
-  const body = {
-    model: model.model_name,
-    messages: [],
-    max_tokens: options.maxTokens || 2048,
-    stream: false,
-  };
-
-  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
-  body.messages.push({ role: 'user', content: prompt });
-  if (options.temperature != null) body.temperature = options.temperature;
-
-  const res = await fetch(`${model.endpoint}/v1/chat/completions`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify(body),
-    signal: AbortSignal.timeout(options.timeout || 120000),
-  });
-
-  if (!res.ok) throw new Error(`llama.cpp error: ${res.status}`);
-  const data = await res.json();
-
-  return {
-    text: data.choices?.[0]?.message?.content || '',
-    promptTokens: data.usage?.prompt_tokens || 0,
-    completionTokens: data.usage?.completion_tokens || 0,
-  };
-}
-
-async function _inferOpenAICompatible(model, prompt, options) {
-  const body = {
-    model: model.model_name,
-    messages: [],
-    max_tokens: options.maxTokens || 2048,
-    stream: false,
-  };
-
-  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
-  body.messages.push({ role: 'user', content: prompt });
-  if (options.temperature != null) body.temperature = options.temperature;
-
-  const res = await fetch(`${model.endpoint}/v1/chat/completions`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify(body),
-    signal: AbortSignal.timeout(options.timeout || 120000),
-  });
-
-  if (!res.ok) throw new Error(`Inference error: ${res.status}`);
-  const data = await res.json();
-
-  return {
-    text: data.choices?.[0]?.message?.content || '',
-    promptTokens: data.usage?.prompt_tokens || 0,
-    completionTokens: data.usage?.completion_tokens || 0,
-  };
-}
-
-// ─── Provider Probing ────────────────────────────────────────────────
-
-async function _probeOllama(baseUrl) {
-  const res = await fetch(`${baseUrl}/api/tags`, { signal: AbortSignal.timeout(5000) });
-  if (!res.ok) return [];
-  const data = await res.json();
-  return (data.models || []).map(m => ({
-    name: m.name,
-    capabilities: _detectCapabilities(m.name),
-    contextWindow: m.details?.parameter_size ? _estimateContext(m.details.parameter_size) : 4096,
-    parameters: { size: m.size, family: m.details?.family },
-  }));
-}
-
-async function _probeLlamaCpp(baseUrl) {
-  const res = await fetch(`${baseUrl}/v1/models`, { signal: AbortSignal.timeout(5000) });
-  if (!res.ok) return [];
-  const data = await res.json();
-  return (data.data || []).map(m => ({
-    name: m.id,
-    capabilities: _detectCapabilities(m.id),
-    contextWindow: 4096,
-    parameters: {},
-  }));
-}
-
-async function _probeOpenAICompatible(baseUrl) {
-  const res = await fetch(`${baseUrl}/v1/models`, { signal: AbortSignal.timeout(5000) });
-  if (!res.ok) return [];
-  const data = await res.json();
-  return (data.data || []).map(m => ({
-    name: m.id,
-    capabilities: _detectCapabilities(m.id),
-    contextWindow: 4096,
-    parameters: {},
-  }));
-}
-
-// ─── Helpers ─────────────────────────────────────────────────────────
-
-function _detectCapabilities(modelName) {
-  const n = modelName.toLowerCase();
-  const caps = ['text'];
-  if (n.includes('vision') || n.includes('llava') || n.includes('bakllava')) caps.push('vision');
-  if (n.includes('code') || n.includes('codellama') || n.includes('deepseek-coder') || n.includes('starcoder')) caps.push('code');
-  if (n.includes('embed') || n.includes('nomic')) caps.push('embedding');
-  if (n.includes('mistral') || n.includes('mixtral')) caps.push('reasoning');
-  return caps;
-}
-
-function _estimateContext(paramSize) {
-  // Rough estimate: smaller models typically have smaller context
-  if (typeof paramSize === 'string') {
-    const num = parseFloat(paramSize);
-    if (num >= 70) return 32768;
-    if (num >= 13) return 8192;
-    return 4096;
-  }
-  return 4096;
-}
-
-function _deserializeModel(row) {
-  return {
-    ...row,
-    capabilities: JSON.parse(row.capabilities || '["text"]'),
-    parameters: JSON.parse(row.parameters || '{}'),
-  };
-}
-
-module.exports = {
-  discoverModels, registerModel, infer,
-  getModels, getAvailableModels, getModel, updateModelStatus,
-  getStats,
-};
+/**
+ * Local AI — Sovereign Intelligence Runtime
+ *
+ * Manages local AI models running on the user's own hardware.
+ * Auto-detects Ollama, llama.cpp, and any OpenAI-compatible local endpoint.
+ * Routes inference requests to the best available model based on capability,
+ * context window, and current load.
+ *
+ * Supported Providers:
+ *   - Ollama (http://localhost:11434)
+ *   - llama.cpp server (http://localhost:8080)
+ *   - Custom OpenAI-compatible endpoints
+ *
+ * All inference happens locally. No data leaves the device.
+ */
+
+const crypto = require('crypto');
+const { db } = require('../models/db');
+
+// ─── Schema ──────────────────────────────────────────────────────────
+
+db.exec(`
+  CREATE TABLE IF NOT EXISTS local_models (
+    id TEXT PRIMARY KEY,
+    site_id TEXT NOT NULL,
+    provider TEXT NOT NULL,
+    model_name TEXT NOT NULL,
+    endpoint TEXT NOT NULL,
+    capabilities TEXT DEFAULT '["text"]',
+    context_window INTEGER DEFAULT 4096,
+    parameters TEXT DEFAULT '{}',
+    status TEXT DEFAULT 'available',
+    total_requests INTEGER DEFAULT 0,
+    total_tokens INTEGER DEFAULT 0,
+    avg_latency_ms REAL DEFAULT 0,
+    last_used TEXT,
+    last_probe TEXT,
+    created_at TEXT DEFAULT (datetime('now')),
+    UNIQUE(site_id, provider, model_name)
+  );
+
+  CREATE TABLE IF NOT EXISTS local_inference_log (
+    id TEXT PRIMARY KEY,
+    model_id TEXT NOT NULL,
+    task_type TEXT,
+    prompt_tokens INTEGER DEFAULT 0,
+    completion_tokens INTEGER DEFAULT 0,
+    latency_ms INTEGER DEFAULT 0,
+    success INTEGER DEFAULT 1,
+    created_at TEXT DEFAULT (datetime('now'))
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_local_models_site ON local_models(site_id);
+  CREATE INDEX IF NOT EXISTS idx_local_models_status ON local_models(status);
+  CREATE INDEX IF NOT EXISTS idx_local_inference_model ON local_inference_log(model_id);
+`);
+
+// ─── Prepared Statements ─────────────────────────────────────────────
+
+const stmts = {
+  upsertModel: db.prepare("INSERT INTO local_models (id, site_id, provider, model_name, endpoint, capabilities, context_window, parameters) VALUES (?, ?, ?, ?, ?, ?, ?, ?) ON CONFLICT(site_id, provider, model_name) DO UPDATE SET endpoint = ?, capabilities = ?, context_window = ?, parameters = ?, status = 'available', last_probe = datetime('now')"),
+  getModel: db.prepare('SELECT * FROM local_models WHERE id = ?'),
+  getModels: db.prepare('SELECT * FROM local_models WHERE site_id = ? ORDER BY provider, model_name'),
+  getAvailableModels: db.prepare("SELECT * FROM local_models WHERE site_id = ? AND status = 'available' ORDER BY avg_latency_ms ASC"),
+  getModelsByCapability: db.prepare("SELECT * FROM local_models WHERE site_id = ? AND status = 'available' AND capabilities LIKE ? ORDER BY avg_latency_ms ASC"),
+  updateModelStatus: db.prepare('UPDATE local_models SET status = ?, last_probe = datetime(\'now\') WHERE id = ?'),
+  updateModelStats: db.prepare("UPDATE local_models SET total_requests = total_requests + 1, total_tokens = total_tokens + ?, avg_latency_ms = (avg_latency_ms * total_requests + ?) / (total_requests + 1), last_used = datetime('now') WHERE id = ?"),
+  insertLog: db.prepare('INSERT INTO local_inference_log (id, model_id, task_type, prompt_tokens, completion_tokens, latency_ms, success) VALUES (?, ?, ?, ?, ?, ?, ?)'),
+  getStats: db.prepare(`SELECT
+    (SELECT COUNT(*) FROM local_models WHERE site_id = ? AND status = 'available') as available_models,
+    (SELECT COUNT(*) FROM local_models WHERE site_id = ?) as total_models,
+    (SELECT SUM(total_requests) FROM local_models WHERE site_id = ?) as total_requests,
+    (SELECT SUM(total_tokens) FROM local_models WHERE site_id = ?) as total_tokens,
+    (SELECT AVG(avg_latency_ms) FROM local_models WHERE site_id = ? AND status = 'available') as avg_latency`),
+};
+
+// ─── Default Provider Endpoints ──────────────────────────────────────
+
+const PROVIDERS = {
+  ollama: { name: 'ollama', baseUrl: 'http://localhost:11434', tagsPath: '/api/tags', chatPath: '/api/chat', generatePath: '/api/generate' },
+  llamacpp: { name: 'llamacpp', baseUrl: 'http://localhost:8080', chatPath: '/v1/chat/completions', modelsPath: '/v1/models' },
+};
+
+// ─── Model Discovery ─────────────────────────────────────────────────
+
+/**
+ * Probe local endpoints and register discovered models.
+ */
+async function discoverModels(siteId, customEndpoints = []) {
+  const discovered = [];
+
+  // Probe Ollama
+  try {
+    const ollamaModels = await _probeOllama(PROVIDERS.ollama.baseUrl);
+    for (const m of ollamaModels) {
+      const result = _registerModel(siteId, 'ollama', m.name, PROVIDERS.ollama.baseUrl, m.capabilities, m.contextWindow, m.parameters);
+      discovered.push(result);
+    }
+  } catch (_) { /* Ollama not running */ }
+
+  // Probe llama.cpp
+  try {
+    const lcModels = await _probeLlamaCpp(PROVIDERS.llamacpp.baseUrl);
+    for (const m of lcModels) {
+      const result = _registerModel(siteId, 'llamacpp', m.name, PROVIDERS.llamacpp.baseUrl, m.capabilities, m.contextWindow, m.parameters);
+      discovered.push(result);
+    }
+  } catch (_) { /* llama.cpp not running */ }
+
+  // Probe custom endpoints
+  for (const ep of customEndpoints) {
+    try {
+      const models = await _probeOpenAICompatible(ep.url);
+      for (const m of models) {
+        const result = _registerModel(siteId, ep.name || 'custom', m.name, ep.url, m.capabilities, m.contextWindow, m.parameters);
+        discovered.push(result);
+      }
+    } catch (_) { /* endpoint not available */ }
+  }
+
+  return { discovered: discovered.length, models: discovered };
+}
+
+/**
+ * Register a model manually.
+ */
+function registerModel(siteId, provider, modelName, endpoint, capabilities = ['text'], contextWindow = 4096) {
+  return _registerModel(siteId, provider, modelName, endpoint, capabilities, contextWindow, {});
+}
+
+function _registerModel(siteId, provider, modelName, endpoint, capabilities, contextWindow, parameters) {
+  const id = crypto.randomUUID();
+  const caps = JSON.stringify(capabilities);
+  const params = JSON.stringify(parameters);
+
+  stmts.upsertModel.run(id, siteId, provider, modelName, endpoint, caps, contextWindow, params, endpoint, caps, contextWindow, params);
+  return { id, provider, modelName, endpoint, capabilities, contextWindow };
+}
+
+// ─── Inference ───────────────────────────────────────────────────────
+
+/**
+ * Run inference on the best available local model.
+ * @param {string} siteId
+ * @param {string} prompt - The user prompt
+ * @param {object} options - { capability, model, systemPrompt, temperature, maxTokens, stream }
+ */
+async function infer(siteId, prompt, options = {}) {
+  const capability = options.capability || 'text';
+
+  // Select model
+  let model;
+  if (options.modelId) {
+    model = stmts.getModel.get(options.modelId);
+    if (!model || model.status !== 'available') throw new Error('Selected model unavailable');
+  } else {
+    const candidates = stmts.getModelsByCapability.all(siteId, `%${capability}%`);
+    if (candidates.length === 0) throw new Error(`No local model available for capability: ${capability}`);
+    model = candidates[0]; // Fastest by avg latency
+  }
+
+  const start = Date.now();
+  let result;
+
+  try {
+    const parsed = JSON.parse(model.parameters || '{}');
+    if (model.provider === 'ollama') {
+      result = await _inferOllama(model, prompt, options);
+    } else if (model.provider === 'llamacpp') {
+      result = await _inferLlamaCpp(model, prompt, options);
+    } else {
+      result = await _inferOpenAICompatible(model, prompt, options);
+    }
+  } catch (err) {
+    const latency = Date.now() - start;
+    stmts.insertLog.run(crypto.randomUUID(), model.id, capability, 0, 0, latency, 0);
+    throw err;
+  }
+
+  const latency = Date.now() - start;
+  const totalTokens = (result.promptTokens || 0) + (result.completionTokens || 0);
+
+  stmts.updateModelStats.run(totalTokens, latency, model.id);
+  stmts.insertLog.run(crypto.randomUUID(), model.id, capability, result.promptTokens || 0, result.completionTokens || 0, latency, 1);
+
+  return {
+    modelId: model.id,
+    provider: model.provider,
+    model: model.model_name,
+    response: result.text,
+    promptTokens: result.promptTokens || 0,
+    completionTokens: result.completionTokens || 0,
+    latencyMs: latency,
+  };
+}
+
+// ─── Model Management ────────────────────────────────────────────────
+
+function getModels(siteId) {
+  return stmts.getModels.all(siteId).map(_deserializeModel);
+}
+
+function getAvailableModels(siteId) {
+  return stmts.getAvailableModels.all(siteId).map(_deserializeModel);
+}
+
+function getModel(modelId) {
+  const row = stmts.getModel.get(modelId);
+  return row ? _deserializeModel(row) : null;
+}
+
+function updateModelStatus(modelId, status) {
+  stmts.updateModelStatus.run(status, modelId);
+}
+
+function getStats(siteId) {
+  const row = stmts.getStats.get(siteId, siteId, siteId, siteId, siteId);
+  return {
+    availableModels: row.available_models || 0,
+    totalModels: row.total_models || 0,
+    totalRequests: row.total_requests || 0,
+    totalTokens: row.total_tokens || 0,
+    avgLatency: row.avg_latency ? Math.round(row.avg_latency) : 0,
+  };
+}
+
+// ─── Provider-Specific Inference ─────────────────────────────────────
+
+async function _inferOllama(model, prompt, options) {
+  const body = {
+    model: model.model_name,
+    messages: [],
+    stream: false,
+    options: {},
+  };
+
+  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
+  body.messages.push({ role: 'user', content: prompt });
+  if (options.temperature != null) body.options.temperature = options.temperature;
+
+  const res = await fetch(`${model.endpoint}/api/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(options.timeout || 120000),
+  });
+
+  if (!res.ok) throw new Error(`Ollama error: ${res.status}`);
+  const data = await res.json();
+
+  return {
+    text: data.message?.content || '',
+    promptTokens: data.prompt_eval_count || 0,
+    completionTokens: data.eval_count || 0,
+  };
+}
+
+async function _inferLlamaCpp(model, prompt, options) {
+  const body = {
+    model: model.model_name,
+    messages: [],
+    max_tokens: options.maxTokens || 2048,
+    stream: false,
+  };
+
+  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
+  body.messages.push({ role: 'user', content: prompt });
+  if (options.temperature != null) body.temperature = options.temperature;
+
+  const res = await fetch(`${model.endpoint}/v1/chat/completions`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(options.timeout || 120000),
+  });
+
+  if (!res.ok) throw new Error(`llama.cpp error: ${res.status}`);
+  const data = await res.json();
+
+  return {
+    text: data.choices?.[0]?.message?.content || '',
+    promptTokens: data.usage?.prompt_tokens || 0,
+    completionTokens: data.usage?.completion_tokens || 0,
+  };
+}
+
+async function _inferOpenAICompatible(model, prompt, options) {
+  const body = {
+    model: model.model_name,
+    messages: [],
+    max_tokens: options.maxTokens || 2048,
+    stream: false,
+  };
+
+  if (options.systemPrompt) body.messages.push({ role: 'system', content: options.systemPrompt });
+  body.messages.push({ role: 'user', content: prompt });
+  if (options.temperature != null) body.temperature = options.temperature;
+
+  const res = await fetch(`${model.endpoint}/v1/chat/completions`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(options.timeout || 120000),
+  });
+
+  if (!res.ok) throw new Error(`Inference error: ${res.status}`);
+  const data = await res.json();
+
+  return {
+    text: data.choices?.[0]?.message?.content || '',
+    promptTokens: data.usage?.prompt_tokens || 0,
+    completionTokens: data.usage?.completion_tokens || 0,
+  };
+}
+
+// ─── Provider Probing ────────────────────────────────────────────────
+
+async function _probeOllama(baseUrl) {
+  const res = await fetch(`${baseUrl}/api/tags`, { signal: AbortSignal.timeout(5000) });
+  if (!res.ok) return [];
+  const data = await res.json();
+  return (data.models || []).map(m => ({
+    name: m.name,
+    capabilities: _detectCapabilities(m.name),
+    contextWindow: m.details?.parameter_size ? _estimateContext(m.details.parameter_size) : 4096,
+    parameters: { size: m.size, family: m.details?.family },
+  }));
+}
+
+async function _probeLlamaCpp(baseUrl) {
+  const res = await fetch(`${baseUrl}/v1/models`, { signal: AbortSignal.timeout(5000) });
+  if (!res.ok) return [];
+  const data = await res.json();
+  return (data.data || []).map(m => ({
+    name: m.id,
+    capabilities: _detectCapabilities(m.id),
+    contextWindow: 4096,
+    parameters: {},
+  }));
+}
+
+async function _probeOpenAICompatible(baseUrl) {
+  const res = await fetch(`${baseUrl}/v1/models`, { signal: AbortSignal.timeout(5000) });
+  if (!res.ok) return [];
+  const data = await res.json();
+  return (data.data || []).map(m => ({
+    name: m.id,
+    capabilities: _detectCapabilities(m.id),
+    contextWindow: 4096,
+    parameters: {},
+  }));
+}
+
+// ─── Helpers ─────────────────────────────────────────────────────────
+
+function _detectCapabilities(modelName) {
+  const n = modelName.toLowerCase();
+  const caps = ['text'];
+  if (n.includes('vision') || n.includes('llava') || n.includes('bakllava')) caps.push('vision');
+  if (n.includes('code') || n.includes('codellama') || n.includes('deepseek-coder') || n.includes('starcoder')) caps.push('code');
+  if (n.includes('embed') || n.includes('nomic')) caps.push('embedding');
+  if (n.includes('mistral') || n.includes('mixtral')) caps.push('reasoning');
+  return caps;
+}
+
+function _estimateContext(paramSize) {
+  // Rough estimate: smaller models typically have smaller context
+  if (typeof paramSize === 'string') {
+    const num = parseFloat(paramSize);
+    if (num >= 70) return 32768;
+    if (num >= 13) return 8192;
+    return 4096;
+  }
+  return 4096;
+}
+
+function _deserializeModel(row) {
+  return {
+    ...row,
+    capabilities: JSON.parse(row.capabilities || '["text"]'),
+    parameters: JSON.parse(row.parameters || '{}'),
+  };
+}
+
+module.exports = {
+  discoverModels, registerModel, infer,
+  getModels, getAvailableModels, getModel, updateModelStatus,
+  getStats,
+};