Update app.py

app.py

@@ -2,15 +2,40 @@ from fastapi import FastAPI
 from pydantic import BaseModel
 from typing import Optional
 
+# ✅ Modules de LlamaIndex
 from llama_index.core.settings import Settings
 from llama_index.core import Document
-from llama_index.embeddings.huggingface import HuggingFaceEmbedding
 from llama_index.llms.llama_cpp import LlamaCPP
 from llama_index.core.node_parser import SemanticSplitterNodeParser
 
+# ✅ Pour l'embedding LOCAL via transformers
+from transformers import AutoTokenizer, AutoModel
+import torch
+import torch.nn.functional as F
+import os
+
 app = FastAPI()
 
-# 📥 Modèle de la requête JSON envoyée à /chunk
+# ✅ Configuration locale du cache HF pour Hugging Face
+CACHE_DIR = "/data"
+os.environ["HF_HOME"] = CACHE_DIR
+os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
+os.environ["HF_MODULES_CACHE"] = CACHE_DIR
+os.environ["HF_HUB_CACHE"] = CACHE_DIR
+
+# ✅ Configuration du modèle d’embedding local (ex: BGE / Nomic / GTE etc.)
+MODEL_NAME = "BAAI/bge-small-en-v1.5"
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
+model = AutoModel.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
+
+def get_embedding(text: str):
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
+        outputs = model(**inputs)
+        embeddings = outputs.last_hidden_state[:, 0]
+        return F.normalize(embeddings, p=2, dim=1).squeeze().tolist()
+
+# ✅ Données entrantes du POST
 class ChunkRequest(BaseModel):
     text: str
     source_id: Optional[str] = None
@@ -20,38 +45,26 @@ class ChunkRequest(BaseModel):
 
 @app.post("/chunk")
 async def chunk_text(data: ChunkRequest):
-    # ✅ Chargement direct d’un modèle hébergé sur Hugging Face (pas de fichier local .gguf)
-    llm = LlamaCPP(
-        model_url="https://huggingface.co/leafspark/Mistral-7B-Instruct-v0.2-Q4_K_M-GGUF/resolve/main/mistral-7b-instruct-v0.2.Q4_K_M.gguf",
-        temperature=0.1,
-        max_new_tokens=512,
-        context_window=2048,
-        generate_kwargs={"top_p": 0.95},
-        model_kwargs={"n_gpu_layers": 1},  # Laisse 1 si CPU
-    )
-
-    # ✅ Embedding open-source via Hugging Face
-    embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-small-en-v1.5")
-
-    # ✅ Configuration du service IA
-    # service_context = ServiceContext.from_defaults(
-        # llm=llm,
-        # embed_model=embed_model
-    # )
-
+    try:
+        # ✅ Chargement du modèle LLM depuis Hugging Face en ligne (pas de .gguf local)
+        llm = LlamaCPP(
+            model_url="https://huggingface.co/leafspark/Mistral-7B-Instruct-v0.2-Q4_K_M-GGUF/resolve/main/mistral-7b-instruct-v0.2.Q4_K_M.gguf",
+            temperature=0.1,
+            max_new_tokens=512,
+            context_window=2048,
+            generate_kwargs={"top_p": 0.95},
+            model_kwargs={"n_gpu_layers": 1},
+        )
 
-    # ✅ Nouvelle méthode recommandée : paramétrer Settings globalement
-    Settings.llm = llm
-    Settings.embed_model = embed_model
+        # ✅ Intégration manuelle de l'embedding local dans Settings
+        class SimpleEmbedding:
+            def get_text_embedding(self, text: str):
+                return get_embedding(text)
 
+        Settings.llm = llm
+        Settings.embed_model = SimpleEmbedding()
 
-    
-    try:
         # ✅ Découpage sémantique intelligent
-        # parser = SemanticSplitterNodeParser.from_defaults(service_context=service_context)
-        
-        # ✅ Appel du parser sans service_context
-        
         parser = SemanticSplitterNodeParser.from_defaults()
         nodes = parser.get_nodes_from_documents([Document(text=data.text)])
 
@@ -61,14 +74,12 @@ async def chunk_text(data: ChunkRequest):
             "source_id": data.source_id,
             "titre": data.titre,
             "source": data.source,
-            "type": data.type
+            "type": data.type,
         }
+    
     except Exception as e:
         return {"error": str(e)}
 
-
 if __name__ == "__main__":
     import uvicorn
     uvicorn.run("app:app", host="0.0.0.0", port=7860)
-
-