wevia-brain/wevia-rag-engine.php

<?php
/**
 * ╔══════════════════════════════════════════════════════════════════╗
 * ║  WEVIA RAG ENGINE v1.0 — Qdrant Semantic Search for Master Router ║
 * ║  Standalone — ZERO impact on existing codebase                    ║
 * ║  Created: 2026-04-04                                              ║
 * ╚══════════════════════════════════════════════════════════════════╝
 *
 * FLOW: Query → Embed (Ollama) → Search (Qdrant) → Context Injection
 */

define('RAG_QDRANT_URL', 'http://127.0.0.1:6333');
define('RAG_OLLAMA_URL', 'http://127.0.0.1:11435');
define('RAG_EMBED_MODEL', 'all-minilm:latest');  // 384 dim, matches weval_skills
define('RAG_TOP_K', 5);
define('RAG_MIN_SCORE', 0.35);

/**
 * Generate embedding vector via Ollama
 */
function rag_embed($text, $timeout = 10) {
    $payload = json_encode([
        'model' => RAG_EMBED_MODEL,
        'input' => mb_substr($text, 0, 1000),
    ], JSON_UNESCAPED_UNICODE);

    $ch = curl_init(RAG_OLLAMA_URL . '/api/embed');
    curl_setopt_array($ch, [
        CURLOPT_POST => true,
        CURLOPT_POSTFIELDS => $payload,
        CURLOPT_HTTPHEADER => ['Content-Type: application/json'],
        CURLOPT_RETURNTRANSFER => true,
        CURLOPT_TIMEOUT => $timeout,
        CURLOPT_CONNECTTIMEOUT => 3,
    ]);

    $result = curl_exec($ch);
    $httpCode = curl_getinfo($ch, CURLINFO_HTTP_CODE);
    curl_close($ch);

    if ($httpCode !== 200 || !$result) return null;

    $data = json_decode($result, true);
    // Ollama returns embeddings array
    $embeddings = $data['embeddings'] ?? $data['embedding'] ?? null;
    if (is_array($embeddings) && !empty($embeddings)) {
        // If nested array, take first
        return is_array($embeddings[0]) ? $embeddings[0] : $embeddings;
    }
    return null;
}

/**
 * Search Qdrant collection with vector
 */
function rag_searchQdrant($collection, $vector, $topK = 5, $minScore = 0.35) {
    $payload = json_encode([
        'vector' => $vector,
        'limit' => $topK,
        'score_threshold' => $minScore,
        'with_payload' => true,
    ], JSON_UNESCAPED_UNICODE);

    $ch = curl_init(RAG_QDRANT_URL . "/collections/$collection/points/query");
    curl_setopt_array($ch, [
        CURLOPT_POST => true,
        CURLOPT_POSTFIELDS => $payload,
        CURLOPT_HTTPHEADER => ['Content-Type: application/json'],
        CURLOPT_RETURNTRANSFER => true,
        CURLOPT_TIMEOUT => 5,
        CURLOPT_CONNECTTIMEOUT => 2,
    ]);

    $result = curl_exec($ch);
    $httpCode = curl_getinfo($ch, CURLINFO_HTTP_CODE);
    curl_close($ch);

    if ($httpCode !== 200 || !$result) {
        // Fallback: try /points/search (older Qdrant API)
        $ch2 = curl_init(RAG_QDRANT_URL . "/collections/$collection/points/search");
        curl_setopt_array($ch2, [
            CURLOPT_POST => true,
            CURLOPT_POSTFIELDS => $payload,
            CURLOPT_HTTPHEADER => ['Content-Type: application/json'],
            CURLOPT_RETURNTRANSFER => true,
            CURLOPT_TIMEOUT => 5,
        ]);
        $result = curl_exec($ch2);
        $httpCode = curl_getinfo($ch2, CURLINFO_HTTP_CODE);
        curl_close($ch2);
        
        if ($httpCode !== 200 || !$result) return [];
    }

    $data = json_decode($result, true);
    $points = $data['result'] ?? $data['result']['points'] ?? [];
    if (!is_array($points)) return [];
    
    // Handle nested result format
    if (isset($points['points'])) $points = $points['points'];

    $results = [];
    foreach ($points as $point) {
        $score = $point['score'] ?? 0;
        $payload_data = $point['payload'] ?? [];
        $text = $payload_data['text'] ?? $payload_data['content'] ?? $payload_data['description'] ?? '';
        $title = $payload_data['title'] ?? $payload_data['name'] ?? $payload_data['key'] ?? '';
        $source = $payload_data['source'] ?? $payload_data['category'] ?? '';
        
        if (!empty($text)) {
            $results[] = [
                'text' => mb_substr($text, 0, 800),
                'title' => $title,
                'source' => $source,
                'score' => round($score, 3),
            ];
        }
    }

    return $results;
}

/**
 * Multi-collection RAG search
 * Searches weval_skills (main) + wevia_kb + wevia_memory
 */
function rag_search($query, $options = []) {
    $start = microtime(true);
    $topK = $options['top_k'] ?? RAG_TOP_K;
    $minScore = $options['min_score'] ?? RAG_MIN_SCORE;
    $collections = $options['collections'] ?? ['weval_skills', 'wevia_kb'];

    // Step 1: Generate embedding
    $vector = rag_embed($query);
    if (!$vector) {
        error_log("RAG: embedding failed for: " . mb_substr($query, 0, 50));
        return ['context' => '', 'results' => [], 'error' => 'embedding_failed'];
    }

    $embedLatency = round((microtime(true) - $start) * 1000);

    // Step 2: Search each collection
    $allResults = [];
    foreach ($collections as $collection) {
        $results = rag_searchQdrant($collection, $vector, $topK, $minScore);
        foreach ($results as &$r) {
            $r['collection'] = $collection;
        }
        $allResults = array_merge($allResults, $results);
    }

    // Step 3: Sort by score descending
    usort($allResults, fn($a, $b) => $b['score'] <=> $a['score']);

    // Step 4: Take top K
    $allResults = array_slice($allResults, 0, $topK);

    // Step 5: Build context string
    $context = '';
    if (!empty($allResults)) {
        $context = "\n\n## CONTEXTE WEVAL (RAG — " . count($allResults) . " résultats pertinents)\n";
        foreach ($allResults as $i => $r) {
            $title = $r['title'] ? "**{$r['title']}**" : '';
            $src = $r['source'] ? " [{$r['source']}]" : '';
            $score = $r['score'];
            $context .= "\n### Résultat " . ($i + 1) . " (score: $score)$src\n$title\n{$r['text']}\n";
        }
    }

    $totalLatency = round((microtime(true) - $start) * 1000);

    error_log(sprintf(
        "RAG: query=%s results=%d embed=%dms total=%dms",
        mb_substr($query, 0, 40),
        count($allResults),
        $embedLatency,
        $totalLatency
    ));

    return [
        'context' => $context,
        'results' => $allResults,
        'embed_latency_ms' => $embedLatency,
        'total_latency_ms' => $totalLatency,
        'vector_dim' => count($vector),
        'collections_searched' => $collections,
    ];
}

/**
 * Enrich system prompt with RAG context
 */
function rag_enrichPrompt($systemPrompt, $userMessage, $options = []) {
    $rag = rag_search($userMessage, $options);
    
    if (!empty($rag['context'])) {
        $systemPrompt .= $rag['context'];
        $systemPrompt .= "\n\n**INSTRUCTION:** Utilise le contexte ci-dessus pour enrichir ta réponse. Si le contexte est pertinent, intègre-le naturellement. Sinon, réponds normalement.";
    }

    return [
        'system_prompt' => $systemPrompt,
        'rag' => $rag,
    ];
}