Dynamic Search Parameters Added

app.py

@@ -33,24 +33,41 @@ def check_indexes_ready():
         return False
 
 def process_query(query, architecture: Literal["A", "B", "C"]):
-    """Traite la requête de l'utilisateur avec l'architecture A."""
+    """Traite la requête de l'utilisateur avec l'architecture A, B ou C."""
 
     # Reload conversation
     display_chat_history()
 
-    config = { "metadata": { "architecture": architecture }, "tags": ["arch_"+architecture] }
+    config = {"metadata": {"architecture": architecture}, "tags": ["arch_" + architecture]}
 
     if architecture == "A":
         agent = agent_A
-    elif architecture == "B":
-        agent = agent_B
-    elif architecture == "C":
-        agent = agent_C
+        initial_state = {
+            "query": query,
+            "messages": [],
+            "relevant_docs": [],
+            "response": "",
+        }
 
-    st.session_state.chat_history.append({"role": "user", "content": query})
+    elif architecture in ["B", "C"]:
+        agent = agent_B if architecture == "B" else agent_C
+
+        # Récupération des paramètres dynamiques uniquement pour B et C
+        k = st.session_state.get("k", 30)  # Nombre de documents
+        alpha = st.session_state.get("alpha", 0.5)  # Pondération hybride
+        similarity_threshold = st.session_state.get("similarity_threshold", 0.7)  # similarité cosinus
+
+        initial_state = {
+            "query": query,
+            "messages": [],
+            "relevant_docs": [],
+            "response": "",
+            "k": k,  
+            "alpha": alpha,  
+            "similarity_threshold": similarity_threshold,
+        }
 
-    # with st.spinner("Recherche en cours..."):
-    initial_state = {"query": query, "messages": [], "relevant_docs": [], "response": ""}
+    st.session_state.chat_history.append({"role": "user", "content": query})
 
     with st.chat_message("user"):
         st.markdown(query)
@@ -60,21 +77,19 @@ def process_query(query, architecture: Literal["A", "B", "C"]):
     start_time = time.time()  # Start timing
 
     events = agent.stream(initial_state, config=config, stream_mode="messages")
-    
+
     # Ajouter le message du chatbot avec streaming
     with st.chat_message("assistant"):
         response_placeholder = st.empty()
 
-
         for event in events:
             for message in event:
                 if isinstance(message, AIMessageChunk):
-                    # Vérifiez le type de message et affichez son contenu
                     if hasattr(message, 'content'):
                         full_response += message.content
 
-                        if re.search(r"(?i)(MATCH|CREATE|MERGE|DELETE|CALL)[\s\S]+;", full_response):
-                            full_response = full_response = re.sub(r"(?i)(MATCH|CREATE|MERGE|DELETE|CALL)[\s\S]+?;", "", full_response).strip()
+                        # Supprimer les requêtes Cypher avant l'affichage
+                        full_response = re.sub(r"(?i)(MATCH|CREATE|MERGE|DELETE|CALL)[\s\S]+?;", "", full_response).strip()
 
                         response_placeholder.markdown(full_response)
 
@@ -131,6 +146,23 @@ def display_sidebar():
 
         # Token metrics containers
         # st.sidebar.markdown("### Tokens")
+
+        st.markdown("### Paramètres de la recherche hybride")
+
+        # Sélection du nombre de documents (k)
+        k = st.number_input("Nombre de documents à récupérer (k)", min_value=1, max_value=100, value=30, step=1)
+
+        # Sélection du score de similarité cosinus(entre la requête et les documents pertinents)
+        similarity_threshold = st.slider("Score de similarité cosinus entre la requête et les documents pertinents", 0.0, 1.0, 0.7, 0.05)
+        
+        # Sélection du coefficient de pondération alpha
+        alpha = st.slider("Équilibre entre recherche sémantique et syntaxique (pour alpha = 0.0, la recherche est purement syntaxique)", 0.0, 1.0, 0.5, 0.05)
+
+        # Sauvegarde des valeurs dans la session
+        st.session_state['k'] = k
+        st.session_state['alpha'] = alpha
+        st.session_state['similarité'] = similarity_threshold
+
         st.sidebar.markdown("### API Mistral AI")
 
         # HTML pour le tableau sans index ni colonnes inutiles

config.py

@@ -20,6 +20,7 @@ pc = Pinecone(api_key=pinecone_api_key)
 # Configuration des index
 sparse_index_name = "confession-muette-sparse"
 dense_index_name = "confession-muette-dense"
+judge_index_name = "confession-muette-judge"
 
 # LangSmith configuration
 langsmith_endpoint = os.getenv("LANGSMITH_ENDPOINT")
@@ -82,10 +83,25 @@ if dense_index_name not in all_names:
     except Exception as e:
         print("Erreur lors de la création de l'index dense:", e)
 
-print(f"Les index '{sparse_index_name}' et '{dense_index_name}' sont prêts.")
+# Créer l'index judge s'il n'existe pas
+if judge_index_name not in all_names:
+    print(f"L'index '{judge_index_name}' n'existe pas. Création en cours...")
+    try:
+        pc.create_index(
+            name=judge_index_name,
+            metric="dotproduct",
+            dimension=1024,
+            spec=ServerlessSpec(cloud="aws", region="us-east-1"),  
+        )
+        print(f"Index '{judge_index_name}' créé avec succès.")
+    except Exception as e:
+        print("Erreur lors de la création de l'index sparse:", e)
+
+print(f"Les index '{sparse_index_name}' , '{dense_index_name}' et '{judge_index_name}' sont prêts.")
 
 # Connexion aux index
 sparse_index = pc.Index(sparse_index_name)
 dense_index = pc.Index(dense_index_name)
-print(f"Connexion aux index '{sparse_index_name}' et '{dense_index_name}' réussie.")
+judge_index = pc.Index(judge_index_name)
+print(f"Connexion aux index '{sparse_index_name}' , '{dense_index_name}' et '{judge_index_name}' réussie.")
 

emissions.csv.bak

@@ -0,0 +1,18 @@
+timestamp,project_name,run_id,experiment_id,duration,emissions,emissions_rate,cpu_power,gpu_power,ram_power,cpu_energy,gpu_energy,ram_energy,energy_consumed,country_name,country_iso_code,region,cloud_provider,cloud_region,os,python_version,codecarbon_version,cpu_count,cpu_model,gpu_count,gpu_model,longitude,latitude,ram_total_size,tracking_mode,on_cloud,pue
+2025-03-11T09:09:17,codecarbon,6746b1d6-a00b-4535-85f0-5d07699044fc,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,0.343940399994608,1.0596201036586613e-07,3.080824769859177e-07,14.0,0.0,5.9043402671813965,1.3308975555305047e-06,0,5.599643507466613e-07,1.8908619062771664e-06,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:09:28,codecarbon,6746b1d6-a00b-4535-85f0-5d07699044fc,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,11.03492499999993,3.41777160839172e-06,3.0972313888782584e-07,14.0,0.0,5.9043402671813965,4.289979011106576e-05,0,1.8089371158616417e-05,6.098916126968218e-05,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:10:03,codecarbon,6746b1d6-a00b-4535-85f0-5d07699044fc,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,46.45240809999814,1.439103128818192e-05,3.098016201270413e-07,14.0,0.0,5.9043402671813965,0.000180629357611,0,7.617449842101004e-05,0.000256803856032,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:10:14,codecarbon,6746b1d6-a00b-4535-85f0-5d07699044fc,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,57.77152389999537,1.7897503602431466e-05,3.097980180237707e-07,14.0,0.0,5.9043402671813965,0.0002246418070555,0,9.473404908808602e-05,0.0003193758561436,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:10:40,codecarbon,6746b1d6-a00b-4535-85f0-5d07699044fc,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,83.21980689999327,2.57817734186566e-05,3.0980333143092986e-07,14.0,0.0,5.9043402671813965,0.0003236005359999,0,0.0001364678702645,0.0004600684062645,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:23:31,codecarbon,c6623f43-59ad-43c4-86be-5acc04841060,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,0.1962917000055313,6.015137318312737e-08,3.064386990455141e-07,14.0,0.0,5.9043402671813965,7.561085000068285e-07,0,3.172756286023072e-07,1.0733841286091357e-06,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:24:00,codecarbon,c6623f43-59ad-43c4-86be-5acc04841060,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,29.127897900005337,9.023727829842991e-06,3.097967405962838e-07,14.0,0.0,5.9043402671813965,0.0001132624655555,0,4.776338840049357e-05,0.000161025853956,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:28:00,codecarbon,c6623f43-59ad-43c4-86be-5acc04841060,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,269.73810180000146,8.357370895910917e-05,3.0983279114596604e-07,14.0,0.0,5.9043402671813965,0.0010489653354444,0,0.0004423837064568,0.0014913490419013,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:45:29,codecarbon,c6623f43-59ad-43c4-86be-5acc04841060,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,1318.7430936000092,0.0004085957645354,3.0983727347535334e-07,14.0,0.0,5.9043402671813965,0.005128426107,0,0.0021628489788416,0.0072912750858416,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T09:45:33,codecarbon,c6623f43-59ad-43c4-86be-5acc04841060,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,1322.6812570999973,0.0004098154171413,3.098368673038492e-07,14.0,0.0,5.9043402671813965,0.0051437358322777,0,0.0021693036070659,0.0073130394393436,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:08:55,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,0.2247720000013942,6.921696286853276e-08,3.079429949820415e-07,14.0,0.0,5.9043402671813965,8.697468888648372e-07,0,3.654101065942672e-07,1.2351569954591044e-06,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:11:10,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,135.18873959999473,4.1885770293869065e-05,3.0983179825334133e-07,14.0,0.0,5.9043402671813965,0.0005257239851111,0,0.0002217156603834,0.0007474396454945,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:13:22,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,267.0583891999995,8.2743590012402e-05,3.098333299330862e-07,14.0,0.0,5.9043402671813965,0.0010385451210555,0,0.0004379906845869,0.0014765358056425,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:13:27,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,272.09710749999795,8.430420405907759e-05,3.098313129222744e-07,14.0,0.0,5.9043402671813965,0.0010581339039443,0,0.0004462506150348,0.0015043845189792,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:13:35,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,279.8871901999956,8.671737661900644e-05,3.0982974446612534e-07,14.0,0.0,5.9043402671813965,0.0010884237664999,0,0.0004590231297509,0.0015474468962509,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:13:39,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,284.0483774000022,8.800610287334622e-05,3.0982786692498645e-07,14.0,0.0,5.9043402671813965,0.0011045999704444,0,0.0004658438788988,0.0015704438493432,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0
+2025-03-11T10:13:49,codecarbon,0069d3eb-d74a-4e2d-8bd0-0e3492cf6afe,5b0fa12a-3dd7-45bb-9766-cc326314d9f1,293.94846909999615,9.107300579774026e-05,3.0982643344456e-07,14.0,0.0,5.9043402671813965,0.001143094986833312,0,0.0004820768720103549,0.0016251718588436669,France,FRA,nouvelle-aquitaine,,,Windows-10-10.0.19045-SP0,3.10.16,2.8.3,8,11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz,,,-0.5286,44.8567,15.74490737915039,machine,N,1.0

graph_agentB.py

@@ -11,11 +11,19 @@ class GraphState(TypedDict):
     query: str
     relevant_docs: list
     response: str
+    k: int 
+    alpha: float
+    similarity_threshold: float
 
 
 def retrieve_combined(state: GraphState) -> dict:
     """Récupération hybride : Pinecone (sémantique) + BM25 (mots-clés)."""
-    relevant_docs = hybrid_search(state["query"])
+    k = st.session_state.get("k", 30)  # Valeur par défaut 30
+    alpha = st.session_state.get("alpha", 0.5)  # Valeur par défaut 0.5
+    similarity_threshold = st.session_state.get('similarity_threshold', 0.7)  # Valeur par défaut 0.7
+
+
+    relevant_docs = hybrid_search(state["query"], alpha=alpha, k=k, similarity_threshold=similarity_threshold)
     return {"relevant_docs": relevant_docs}
 
 def generate_response(state: GraphState) -> dict:
@@ -30,6 +38,11 @@ def generate_response(state: GraphState) -> dict:
     - Utilisez les mots-clés pertinents de manière naturelle dans votre réponse.
     - Expliquez les concepts en vous appuyant sur le contexte sémantique.
     - Ne mentionnez pas explicitement les termes "recherche sémantique" ou "recherche par mots-clés".
+    - **Si la réponse doit être une liste retournez chaque élément sur **une nouvelle ligne**.
+    - Format correct attendu :
+      1. Élément 1
+      2. Élément 2
+      3. Élément 3
     
     **Informations pertinentes trouvées** :
     {context}

graph_agentC.py

@@ -12,6 +12,9 @@ class GraphState(TypedDict):
     relevant_docs: list  # Résultats de la recherche hybride
     neo4j_results: list  # Résultats de la recherche Neo4j
     response: str
+    k: int 
+    alpha: float
+    similarity_threshold: float
 
 
 def retrieve_unified(state: GraphState) -> dict:

index_documents.py

@@ -1,4 +1,3 @@
-# index_documents.py
 from pinecone_utilsA import index_pdf as index_pdf_A
 from pinecone_utilsB import *
 from pdf_processing import get_existing_pdf, load_and_preprocess_pdf, split_text

neo4j_utils.py

@@ -57,35 +57,27 @@ def generate_cypher_query(user_query):
     prompt = f"""
     Vous êtes un générateur de requêtes Cypher pour une base de données Neo4j. 
     Étant donné une demande de l'utilisateur, générez une requête Cypher correspondante.
-
     **Instructions supplémentaires** :
     - Utilisez `DISTINCT` pour éviter les répétitions.
     - Optimisez la requête pour éviter les chemins redondants.
     - Tous les nœuds ont uniquement la propriété `id`. Utilisez `id` pour filtrer ou retourner des valeurs.
     - Utilisez uniquement les nœuds et relations autorisés suivants :
-
     **Nœuds autorisés** :
     {", ".join(allowed_nodes)}
-
     **Relations autorisées** :
     {", ".join(allowed_relationships)}
-
     **Exemples de requêtes** :
     - Demande : "Trouver tous les lieux mentionnés dans une histoire."
       Requête : MATCH (n:Personnage) RETURN n.id;
-
     - Demande : "Lister tous les personnages qui connaissent Zéphyrine."
       Requête : MATCH (p1:Personnage)-[:CONNAÎT]->(p2:Personnage {{name: "Zéphyrine"}}) RETURN DISTINCT p1.name
-
     **Format de la réponse** :
     - Ne fournissez **que la requête Cypher**, sans explications ni commentaires.
     - La requête doit commencer par `MATCH`, `CREATE`, `MERGE`, `RETURN`, etc.
     - La requête doit se terminer par un point-virgule (`;`).
     - Si vous ne pouvez pas générer de requête, retournez une chaîne vide.
-
     **Demande de l'utilisateur** :
     {user_query}
-
     **Requête Cypher** :
     """
     
@@ -145,7 +137,11 @@ def merge_results(hybrid_results, neo4j_results):
 def unified_search(query):
     """Perform a unified search combining hybrid search and Neo4j knowledge graph search."""
     # Step 1: Perform hybrid search
-    hybrid_results = hybrid_search(query)
+    k = st.session_state.get("k", 30)  # Valeur par défaut 30
+    alpha = st.session_state.get("alpha", 0.5)  # Valeur par défaut 0.5
+    similarity_threshold = st.session_state.get('similarity_threshold', 0.7)  # Valeur par défaut 0.7
+
+    hybrid_results = hybrid_search(query, alpha=alpha, k=k, similarity_threshold=similarity_threshold)
     #print(f"résultat recherche hybride: {hybrid_results}")
 
     # Step 2: Generate and execute Cypher query
@@ -160,4 +156,4 @@ def unified_search(query):
     # Step 3: Merge results
     final_results = merge_results(hybrid_results, neo4j_results)
 
-    return final_results
+    return final_results

pinecone_utilsB.py

@@ -113,44 +113,47 @@ def index_pdf_B(texts):
     st.session_state.indexing_done = True  # Marquer l'indexation comme terminée
     st.success("Indexation terminée sans duplication de contenu.")
 
-def hybrid_search(query):
-        """Récupère les documents pertinents en combinant les résultats de Pinecone et BM25."""
-
-        try:
-            # Générer le vecteur dense pour la requête
-            query_vector = model.encode([query]).tolist()[0]
-            
-            # Générer les valeurs sparse pour la requête
-            sparse_query = sparse_encoder.encode_queries([query])[0]
-            
-            # Effectuer une recherche hybride dans l'indexB
-            results = indexB.query(
-                vector=query_vector,  # Vecteur dense
-                sparse_vector=sparse_query,  # Valeurs sparse
-                top_k=30,  # Nombre de résultats à retourner
-                include_metadata=True  # Inclure les métadonnées
-            )
-            
-            # Récupérer les documents pertinents
-            relevant_docs = []
-            for match in results.get("matches", []):
-                if "metadata" in match and "compressed_text" in match["metadata"]:
+def hybrid_search(query, alpha, k, similarity_threshold):
+    """Récupère les documents pertinents en combinant les résultats de Pinecone et BM25."""
+    try:
+        # Générer le vecteur dense pour la requête
+        query_vector = model.encode([query]).tolist()[0]
+        
+        # Générer les valeurs sparse pour la requête
+        sparse_query = sparse_encoder.encode_queries([query])[0]
+        
+        # Effectuer une recherche hybride dans l'indexB
+        results = indexB.query(
+            vector=query_vector,  # Vecteur dense
+            sparse_vector=sparse_query,  # Valeurs sparse
+            top_k=k,  # Nombre de résultats à retourner
+            include_metadata=True,  # Inclure les métadonnées
+            alpha=alpha  
+        )
+        
+        # Récupérer les documents pertinents
+        relevant_docs = []
+        for match in results.get("matches", []):
+            if "metadata" in match and "compressed_text" in match["metadata"]:
+                score = match.get("score", 0)  # Score de similarité
+                if score >= similarity_threshold:  # Filtrer par seuil
                     compressed_text = match["metadata"]["compressed_text"]
                     sparse_values_json = match["metadata"].get("sparse_values")
 
-                    # Désérialiser les valeurs sparse
+                    # Désérialiser les valeurs sparse si elles existent
                     if sparse_values_json:
                         sparse_values = json.loads(sparse_values_json)
 
                     relevant_docs.append(decompress_text(compressed_text))
-                else:
-                    print(f"Skipping match due to missing metadata or compressed_text: {match}")
-            
-            return relevant_docs
+            else:
+                print(f"Skipping match due to missing metadata or compressed_text: {match}")
+        
+        return relevant_docs
+
+    except Exception as e:
+        st.error(f"Erreur lors de la recherche hybride : {e}")
+        return []
 
-        except Exception as e:
-            st.error(f"Erreur lors de la recherche hybride : {e}")
-            return []
 
 def compress_text(text):
         """Compresse un texte en base64."""