Add pgvector quality exporter + fix missing os import

2026-09-24 22:12:10 -04:00
parent 674586ce7f
commit 7c97c4eec6
2 changed files with 247 additions and 0 deletions
--- a/exporters/pgvector_quality_exporter.py
+++ b/exporters/pgvector_quality_exporter.py
@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+"""
+pgvector quality exporter
+Mesure la pertinence des recherches vectorielles pgvector sur Mem0.
+Exporte la distribution des distances comme histogramme Prometheus
+
+Usage dans le workflow mem0:
+  from exporters.pgvector_quality import log_similarity_score
+  
+  results = db.query(query)  # SQL avec pgvector <=> operator
+  if results:
+      best_distance = results[0].distance
+      log_similarity_score(best_distance, query_type='mem0_search')
+"""
+
+import os
+import psycopg2
+import json
+from prometheus_client import start_http_server, Histogram, Gauge, Counter
+import logging
+
+
+logger = logging.getLogger('pgvector_quality_monitor')
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(message)s')
+
+# Metrics Prometheus
+SIMILARITY_SCORE_HIST = Histogram(
+    'pgvector_similarity_score',
+    'Score similarite du meilleur resultat (pgvector <=> operator)',
+    ['query_type'],  # mem0_search, text_search, etc.
+    namespace='aurelie'
+)
+
+BEST_RESULT_GAUGE = Gauge(
+    'pgvector_best_result_distance',
+    'Distance du meilleur resultat pgvector',
+    ['query_type'],
+    namespace='aurelie'
+)
+
+QUERY_COUNT = Counter(
+    'pgvector_query_count_total',
+    'Nombre total de requetes pgvector executees',
+    ['status'],  # success, error, no_results
+    namespace='aurelie'
+)
+
+
+def measure_similarity(query_text: str, threshold=0.1) -> dict:
+    """Execute une similarity search et retourne le meilleur score."""
+    try:
+        conn = psycopg2.connect(
+            dbname="aurelie",  # PG database mem0 (config à adapter selon l'environnement)
+            host=os.getenv('PGHOST', 'localhost'),
+            port=os.getenv('PGPORT', '5432')
+        )
+        cur = conn.cursor()
+        
+        # pgvector <=> operator → cosine distance (0=identique, 2=complètement oppose)
+        query = """\\
+            SELECT id::text, content, 
+                   (embedding <=> %s::vector) AS distance
+            FROM memory_store
+            ORDER BY distance ASC 
+            LIMIT 5;"""
+        
+        cur.execute(query, [query_text])
+        results = cur.fetchall()
+        
+        if results:
+            best_score = results[0][-1]  # le champ 'distance' (dernier de la tuple)
+            SIMILARITY_SCORE_HIST.labels(query_type='mem0_search').observe(best_score)
+            BEST_RESULT_GAUGE.labels(query_type='mem0_search').set(best_score)
+            
+            QUERY_COUNT.labels(status='success').inc()
+            
+            # Si le meilleur score est > threshold → alerte (ma memoire contextuelle devient floue)
+            if best_score > threshold:   # 0.1 = seuil de "pertinence faible"
+                logger.warning(f"Score trop bas: {best_score:.3f} — ma memoire contextuelle est peu pertinente")
+            
+            return {'best_score': float(best_score), 'results_count': len(results)}
+        else:
+            QUERY_COUNT.labels(status='no_results').inc()
+            return {'best_score': None, 'results_count': 0}
+            
+    except psycopg2.Error as e:
+        logger.error(f"Erreur pgvector similarity search: {e}")
+        QUERY_COUNT.labels(status='error').inc()
+        raise
+
+
+if __name__ == '__main__':
+    start_http_server(8105)  # Prometheus scrape sur port 8105
+    logger.info('pgvector quality exporter started on :8105')