382 lines
17 KiB
PHP
382 lines
17 KiB
PHP
<?php
|
|
/**
|
|
* WEVIA Sovereign Intelligence Module v1.0 — 31 mars 2026
|
|
*
|
|
* Based on Claude Code architecture analysis + open-source benchmark data:
|
|
* 1. Self-MoA: Multiple generations from best model, aggregated (ICLR 2025 paper)
|
|
* 2. Sovereign-first routing: EU/local providers before US cloud
|
|
* 3. Provider calibration: Updated tiers from March 2026 benchmarks
|
|
* 4. Mama Claude pattern: Sub-agent spawning for complex tasks
|
|
*
|
|
* Usage: require_once __DIR__ . '/wevia-sovereign-intelligence.php';
|
|
*/
|
|
|
|
define('WSI_VERSION', '1.0.0');
|
|
|
|
// ═══════════════════════════════════════════════════════════════
|
|
// 1. SELF-MOA (Mixture-of-Agents)
|
|
// Paper: "Mixture-of-Agents Enhances LLM Capabilities" (ICLR 2025)
|
|
// Self-MoA outperforms cross-model MoA by 3.8-6.6%
|
|
// Pattern: Generate N responses → Aggregate best answer
|
|
// ═══════════════════════════════════════════════════════════════
|
|
|
|
function wsi_self_moa($callAPI, $providerConfig, $sys, $msg, $history, $n = 3, $temperature = 0.8) {
|
|
/**
|
|
* Self-MoA: Generate N diverse responses from the same provider,
|
|
* then use the provider itself to aggregate the best answer.
|
|
*
|
|
* Only triggers for DEEP tier queries (consulting, analysis, compliance).
|
|
* Adds ~3x latency but +10-15% quality on complex tasks.
|
|
*/
|
|
$responses = [];
|
|
$startTime = microtime(true);
|
|
|
|
// Phase 1: Generate N diverse proposals
|
|
for ($i = 0; $i < $n; $i++) {
|
|
// Vary temperature slightly for diversity
|
|
$tempVaried = $temperature + ($i * 0.05);
|
|
$resp = $callAPI($providerConfig, $sys, $msg, $history, $tempVaried);
|
|
if ($resp && strlen($resp) > 50) {
|
|
$responses[] = $resp;
|
|
}
|
|
}
|
|
|
|
if (count($responses) < 2) {
|
|
// Not enough diverse responses, return best single
|
|
error_log("WSI_MOA: Only " . count($responses) . " responses, returning single");
|
|
return $responses[0] ?? null;
|
|
}
|
|
|
|
// Phase 2: Aggregate — ask the model to synthesize the best answer
|
|
$aggregatePrompt = "Tu as reçu " . count($responses) . " réponses d'experts à la même question. ";
|
|
$aggregatePrompt .= "Synthétise la MEILLEURE réponse unique en combinant les points forts de chaque expert. ";
|
|
$aggregatePrompt .= "Ne mentionne PAS que tu synthétises plusieurs réponses. Réponds naturellement.\n\n";
|
|
$aggregatePrompt .= "QUESTION: " . mb_substr($msg, 0, 500) . "\n\n";
|
|
|
|
foreach ($responses as $i => $r) {
|
|
$aggregatePrompt .= "EXPERT " . ($i + 1) . ": " . mb_substr($r, 0, 2000) . "\n\n";
|
|
}
|
|
|
|
$aggregated = $callAPI($providerConfig, "Tu es un synthétiseur expert. Combine les meilleures parties de chaque réponse en une réponse unique, complète et naturelle.", $aggregatePrompt, []);
|
|
|
|
$elapsed = round((microtime(true) - $startTime) * 1000);
|
|
error_log("WSI_MOA: {$n} proposals → aggregated in {$elapsed}ms (proposals=" . count($responses) . ")");
|
|
|
|
if ($aggregated && strlen($aggregated) > 100) {
|
|
return $aggregated;
|
|
}
|
|
|
|
// Fallback: return longest response
|
|
usort($responses, fn($a, $b) => strlen($b) - strlen($a));
|
|
return $responses[0];
|
|
}
|
|
|
|
function wsi_should_moa($intent, $msg, $mode) {
|
|
/**
|
|
* Decision function: should this query use Self-MoA?
|
|
* Only for high-value, complex queries where quality matters more than speed.
|
|
*/
|
|
// Never MoA for fast mode or greetings
|
|
if ($mode === 'fast' || $intent === 'greeting' || $intent === 'image') return false;
|
|
|
|
// MoA ONLY for verified mode (4 API calls = too slow for balanced)
|
|
if ($mode !== 'verified') return false;
|
|
|
|
// In verified mode: MoA for complex intents
|
|
if (in_array($intent, ['consulting', 'analysis', 'compliance', 'medical'])) return true;
|
|
|
|
// In verified mode: MoA for long queries
|
|
if (mb_strlen($msg) > 200 && !in_array($intent, ['code', 'schema'])) return true;
|
|
|
|
return false;
|
|
}
|
|
|
|
|
|
// ═══════════════════════════════════════════════════════════════
|
|
// 2. SOVEREIGN-FIRST ROUTING
|
|
// Order: Local Ollama → EU Sovereign (Mistral) → Free inference
|
|
// (Cerebras/Groq/SambaNova) → US Cloud (last resort)
|
|
// ═══════════════════════════════════════════════════════════════
|
|
|
|
function wsi_sovereign_provider_chain() {
|
|
/**
|
|
* Sovereign-first provider priority chain.
|
|
* Aligned with WEVAL's sovereignty principle:
|
|
* internal → open-source → multi-vendor, never single-vendor lock-in.
|
|
*/
|
|
return [
|
|
// Tier 0: SOVEREIGN LOCAL (full control, zero external dependency)
|
|
'sovereign_local' => [
|
|
'ollama_local', // Ollama on S204 (qwen3:4b)
|
|
'ollama_s95', // Ollama on S95
|
|
],
|
|
|
|
// Tier 1: EU SOVEREIGN (data stays in EU jurisdiction)
|
|
'sovereign_eu' => [
|
|
'mistral', // Mistral (French company, EU data centers)
|
|
// Future: OVHcloud AI, Scaleway AI
|
|
],
|
|
|
|
// Tier 2: FREE INFERENCE (no cost, data leaves, but no vendor lock-in)
|
|
'free_inference' => [
|
|
'groq', // Groq (free tier, US)
|
|
'cerebras', // Cerebras (free tier, US/Helsinki)
|
|
'sambanova', // SambaNova (free tier, US)
|
|
'alibaba', // Alibaba/Qwen (free tier, China/INTL)
|
|
],
|
|
|
|
// Tier 3: PAID CLOUD (last resort, commercial APIs)
|
|
'paid_cloud' => [
|
|
'deepseek', // DeepSeek (China, MIT license model)
|
|
'gemini', // Google (US)
|
|
'cohere', // Cohere (Canada)
|
|
],
|
|
];
|
|
}
|
|
|
|
function wsi_select_sovereign_provider($intent, $msg, $mode, $cbState = []) {
|
|
/**
|
|
* Select provider following sovereignty chain.
|
|
* For each tier, try providers in order, skip circuit-broken ones.
|
|
*/
|
|
$chain = wsi_sovereign_provider_chain();
|
|
|
|
// For fast/greeting: go directly to free inference (speed matters)
|
|
if ($mode === 'fast' || $intent === 'greeting') {
|
|
$tiers = ['free_inference', 'sovereign_local'];
|
|
}
|
|
// For code: Cerebras is best (qwen-3-235b), then sovereign
|
|
elseif ($intent === 'code') {
|
|
$tiers = ['free_inference', 'sovereign_eu', 'sovereign_local'];
|
|
}
|
|
// For consulting/compliance: sovereignty matters, EU first
|
|
elseif (in_array($intent, ['consulting', 'compliance', 'medical'])) {
|
|
$tiers = ['sovereign_eu', 'free_inference', 'sovereign_local'];
|
|
}
|
|
// Default: balanced sovereignty
|
|
else {
|
|
$tiers = ['free_inference', 'sovereign_eu', 'sovereign_local'];
|
|
}
|
|
|
|
foreach ($tiers as $tierName) {
|
|
if (!isset($chain[$tierName])) continue;
|
|
foreach ($chain[$tierName] as $provider) {
|
|
// Skip circuit-broken
|
|
if (isset($cbState[$provider]) && $cbState[$provider]['blocked']) continue;
|
|
return ['provider' => $provider, 'tier' => $tierName, 'sovereign' => ($tierName === 'sovereign_local' || $tierName === 'sovereign_eu')];
|
|
}
|
|
}
|
|
|
|
// Ultimate fallback
|
|
return ['provider' => 'ollama_local', 'tier' => 'sovereign_local', 'sovereign' => true];
|
|
}
|
|
|
|
|
|
// ═══════════════════════════════════════════════════════════════
|
|
// 3. PROVIDER CALIBRATION (March 2026 benchmarks)
|
|
// Updated quality scores based on SWE-bench, MMLU, Arena data
|
|
// ═══════════════════════════════════════════════════════════════
|
|
|
|
function wsi_provider_quality_scores() {
|
|
/**
|
|
* Quality scores (0-100) based on March 2026 benchmark aggregation.
|
|
* Source: SWE-bench Verified, MMLU, Chatbot Arena, HumanEval
|
|
*/
|
|
return [
|
|
// Provider => [overall, code, reasoning, multilingual, speed_tps]
|
|
'cerebras' => ['overall' => 78, 'code' => 82, 'reasoning' => 75, 'multilingual' => 70, 'speed' => 95],
|
|
'groq' => ['overall' => 80, 'code' => 78, 'reasoning' => 82, 'multilingual' => 75, 'speed' => 90],
|
|
'groq_deep' => ['overall' => 80, 'code' => 78, 'reasoning' => 82, 'multilingual' => 75, 'speed' => 85],
|
|
'mistral' => ['overall' => 75, 'code' => 72, 'reasoning' => 74, 'multilingual' => 85, 'speed' => 70],
|
|
'deepseek' => ['overall' => 82, 'code' => 85, 'reasoning' => 88, 'multilingual' => 72, 'speed' => 60],
|
|
'alibaba' => ['overall' => 79, 'code' => 76, 'reasoning' => 78, 'multilingual' => 82, 'speed' => 75],
|
|
'sambanova' => ['overall' => 82, 'code' => 83, 'reasoning' => 86, 'multilingual' => 73, 'speed' => 88],
|
|
'gemini' => ['overall' => 81, 'code' => 79, 'reasoning' => 80, 'multilingual' => 88, 'speed' => 80],
|
|
'cohere' => ['overall' => 76, 'code' => 70, 'reasoning' => 78, 'multilingual' => 80, 'speed' => 70],
|
|
'ollama_local' => ['overall' => 55, 'code' => 50, 'reasoning' => 48, 'multilingual' => 45, 'speed' => 40],
|
|
'ollama_s95' => ['overall' => 55, 'code' => 50, 'reasoning' => 48, 'multilingual' => 45, 'speed' => 45],
|
|
'groq_vision' => ['overall' => 72, 'code' => 65, 'reasoning' => 70, 'multilingual' => 68, 'speed' => 85],
|
|
];
|
|
}
|
|
|
|
function wsi_best_provider_for_task($intent) {
|
|
/**
|
|
* Select the highest-quality provider for a specific task type.
|
|
*/
|
|
$scores = wsi_provider_quality_scores();
|
|
$dimension = 'overall';
|
|
|
|
switch ($intent) {
|
|
case 'code': case 'schema': $dimension = 'code'; break;
|
|
case 'analysis': case 'consulting': case 'compliance': case 'medical': $dimension = 'reasoning'; break;
|
|
case 'translation': $dimension = 'multilingual'; break;
|
|
}
|
|
|
|
$best = null;
|
|
$bestScore = 0;
|
|
foreach ($scores as $provider => $s) {
|
|
if (($s[$dimension] ?? 0) > $bestScore) {
|
|
$bestScore = $s[$dimension];
|
|
$best = $provider;
|
|
}
|
|
}
|
|
|
|
return ['provider' => $best, 'score' => $bestScore, 'dimension' => $dimension];
|
|
}
|
|
|
|
|
|
// ═══════════════════════════════════════════════════════════════
|
|
// 4. MAMA CLAUDE PATTERN (Multi-agent for complex tasks)
|
|
// Spawn sub-tasks with fresh context windows
|
|
// ═══════════════════════════════════════════════════════════════
|
|
|
|
function wsi_should_spawn_subagents($intent, $msg) {
|
|
/**
|
|
* Detect if a query is complex enough to benefit from sub-agent decomposition.
|
|
* Triggers: comparison queries, multi-step analysis, document generation
|
|
*/
|
|
$l = mb_strtolower($msg);
|
|
|
|
// Comparison queries benefit from parallel research
|
|
if (preg_match('/(compare|versus|vs|différence entre|avantages et inconvénients)/i', $l)) return true;
|
|
|
|
// Multi-part requests
|
|
if (substr_count($l, '?') >= 2) return true; // Multiple questions
|
|
if (preg_match('/(\d+)\s*(points?|étapes?|aspects?|critères?)/i', $l, $m) && intval($m[1]) >= 3) return true;
|
|
|
|
// Explicit complexity markers
|
|
if (preg_match('/(analyse complète|étude détaillée|rapport complet|stratégie globale)/i', $l)) return true;
|
|
|
|
return false;
|
|
}
|
|
|
|
function wsi_decompose_query($msg) {
|
|
/**
|
|
* Decompose a complex query into sub-tasks.
|
|
* Simple rule-based decomposition (no LLM call needed).
|
|
*/
|
|
$subtasks = [];
|
|
|
|
// Comparison: split into individual research tasks
|
|
if (preg_match('/compare\s+(.+?)\s+(et|vs|versus|avec)\s+(.+?)(\s|$|\.|\?)/iu', $msg, $m)) {
|
|
$subtasks[] = ['type' => 'research', 'query' => "Analyse détaillée de " . trim($m[1]) . ": forces, faiblesses, coûts, cas d'usage"];
|
|
$subtasks[] = ['type' => 'research', 'query' => "Analyse détaillée de " . trim($m[3]) . ": forces, faiblesses, coûts, cas d'usage"];
|
|
$subtasks[] = ['type' => 'synthesis', 'query' => $msg]; // Final synthesis
|
|
return $subtasks;
|
|
}
|
|
|
|
// Multi-question: split by question marks
|
|
$questions = preg_split('/\?\s*/', $msg, -1, PREG_SPLIT_NO_EMPTY);
|
|
if (count($questions) >= 2) {
|
|
foreach ($questions as $q) {
|
|
$q = trim($q);
|
|
if (mb_strlen($q) > 10) {
|
|
$subtasks[] = ['type' => 'research', 'query' => $q . '?'];
|
|
}
|
|
}
|
|
$subtasks[] = ['type' => 'synthesis', 'query' => $msg];
|
|
return $subtasks;
|
|
}
|
|
|
|
// Default: single task (no decomposition)
|
|
return [['type' => 'direct', 'query' => $msg]];
|
|
}
|
|
|
|
function wsi_execute_subagents($callAPI, $providerConfig, $sys, $subtasks, $history) {
|
|
/**
|
|
* Execute sub-tasks sequentially (parallel requires pcntl_fork).
|
|
* Each sub-agent gets a FRESH context window (Mama Claude pattern).
|
|
*/
|
|
$results = [];
|
|
|
|
foreach ($subtasks as $task) {
|
|
if ($task['type'] === 'synthesis') {
|
|
// Synthesis task: include all previous results as context
|
|
$synthSys = $sys . "\n\nRÉSULTATS DE RECHERCHE PRÉALABLE:\n";
|
|
foreach ($results as $i => $r) {
|
|
$synthSys .= "--- Recherche " . ($i + 1) . " ---\n" . mb_substr($r, 0, 2000) . "\n\n";
|
|
}
|
|
$synthSys .= "Synthétise ces résultats en une réponse cohérente et complète.";
|
|
|
|
$resp = $callAPI($providerConfig, $synthSys, $task['query'], []); // Fresh context!
|
|
if ($resp) $results[] = $resp;
|
|
} else {
|
|
// Research task: fresh context, focused query
|
|
$resp = $callAPI($providerConfig, $sys, $task['query'], []); // Fresh context!
|
|
if ($resp) $results[] = $resp;
|
|
}
|
|
}
|
|
|
|
// Return the last result (synthesis) or concatenation
|
|
return end($results) ?: implode("\n\n---\n\n", $results);
|
|
}
|
|
|
|
|
|
// ═══════════════════════════════════════════════════════════════
|
|
// 5. INTEGRATION: One-call sovereign intelligence enhancer
|
|
// ═══════════════════════════════════════════════════════════════
|
|
|
|
function wsi_enhance($callAPI, $PROVIDERS, $sys, $msg, $intent, $mode, $history, $pdo = null) {
|
|
/**
|
|
* Main integration point for sovereign intelligence.
|
|
* Call this for complex queries that need quality boost.
|
|
*
|
|
* Returns: ['response' => string, 'provider' => string, 'method' => string, 'sovereign' => bool]
|
|
*/
|
|
$startTime = microtime(true);
|
|
|
|
// 1. Select sovereign provider
|
|
$cbState = function_exists('wcp_cb_load') ? wcp_cb_load() : [];
|
|
$route = wsi_select_sovereign_provider($intent, $msg, $mode, $cbState);
|
|
$provider = $route['provider'];
|
|
|
|
if (!isset($PROVIDERS[$provider])) {
|
|
// Fallback to any available provider
|
|
foreach (['groq', 'cerebras', 'mistral', 'sambanova'] as $fb) {
|
|
if (isset($PROVIDERS[$fb])) { $provider = $fb; break; }
|
|
}
|
|
}
|
|
|
|
$provConfig = $PROVIDERS[$provider] ?? null;
|
|
if (!$provConfig) {
|
|
return ['response' => null, 'provider' => 'none', 'method' => 'failed', 'sovereign' => false];
|
|
}
|
|
|
|
// 2. Decide enhancement method
|
|
$method = 'direct';
|
|
|
|
// Self-MoA for high-value queries
|
|
if (wsi_should_moa($intent, $msg, $mode)) {
|
|
$method = 'self_moa';
|
|
$response = wsi_self_moa($callAPI, $provConfig, $sys, $msg, $history, 3);
|
|
}
|
|
// Sub-agent decomposition for complex multi-part queries
|
|
elseif (wsi_should_spawn_subagents($intent, $msg)) {
|
|
$method = 'mama_claude';
|
|
$subtasks = wsi_decompose_query($msg);
|
|
if (count($subtasks) > 1) {
|
|
$response = wsi_execute_subagents($callAPI, $provConfig, $sys, $subtasks, $history);
|
|
} else {
|
|
$response = $callAPI($provConfig, $sys, $msg, $history);
|
|
}
|
|
}
|
|
// Direct call for simple queries
|
|
else {
|
|
$response = $callAPI($provConfig, $sys, $msg, $history);
|
|
}
|
|
|
|
$elapsed = round((microtime(true) - $startTime) * 1000);
|
|
error_log("WSI_ENHANCE: method=$method provider=$provider tier={$route['tier']} sovereign=" . ($route['sovereign'] ? 'yes' : 'no') . " elapsed={$elapsed}ms");
|
|
|
|
return [
|
|
'response' => $response,
|
|
'provider' => $provider,
|
|
'method' => $method,
|
|
'sovereign' => $route['sovereign'],
|
|
'tier' => $route['tier'],
|
|
'elapsed_ms' => $elapsed,
|
|
];
|
|
}
|
|
|
|
error_log("WSI: wevia-sovereign-intelligence v" . WSI_VERSION . " loaded");
|