update the LLM Docker description

2025-09-06 17:47:24 +02:00
parent 2a0d69e6b8
commit a0260b8520
6 changed files with 325 additions and 533 deletions
--- a/LLM/Docker/Dockerfile
+++ b/LLM/Docker/Dockerfile
@@ -34,10 +34,11 @@ COPY start.sh .
 RUN chmod +x start.sh

 # Variables d'environnement par défaut
-ENV OLLAMA_API_URL="http://localhost:11434/api/generate"
-ENV OLLAMA_MODEL="deepseek-coder:33b"
+ENV OLLAMA_HOST="http://host.docker.internal:11434"
+ENV GENERATE_MODEL="deepseek-coder:6.7b-instruct"
 ENV CHROMA_PERSIST_DIR="/chroma_db"
 ENV RESPONSE_CACHE_DIR="/response_cache"
+ENV EMBED_MODEL="nomic-embed-text"

 # Exposer FastAPI
 EXPOSE 8000
--- a/LLM/Docker/app.py
+++ b/LLM/Docker/app.py
@@ -7,9 +7,12 @@ import textwrap
 import requests
 import httpx
 import socket
+import json
+import random
+import traceback
 from fastapi.responses import StreamingResponse
 from fastapi import FastAPI, Query, HTTPException, Request, Body
-from langchain_community.vectorstores import Chroma
+from langchain_chroma import Chroma
 from langchain_ollama import OllamaEmbeddings
 from langchain_community.document_loaders import DirectoryLoader, TextLoader
 from langchain_text_splitters import RecursiveCharacterTextSplitter
@@ -17,55 +20,42 @@ from ddgs import DDGS
 from unstructured.cleaners.core import clean_extra_whitespace, clean_non_ascii_chars, replace_unicode_quotes
 from datetime import datetime, timezone, timedelta
 from pydantic import BaseModel, Field, PrivateAttr, ValidationError
-from typing import Optional, List, Dict, Any
+from typing import Optional, List, Dict, Any, Union, Literal
+from collections import Counter


 # --- Configuration via variables d'environnement ---
+
 PERSIST_DIR = os.environ.get("CHROMA_PERSIST_DIR", "/chroma_db")
 CACHE_DIR = os.environ.get("RESPONSE_CACHE_DIR", "/response_cache")
-MODEL_NAME = os.environ.get("OLLAMA_MODEL", "llama3:13b")
-EMBED_MODEL = os.environ.get("EMBED_MODEL", "mxbai-embed-large:latest")
+GENERATE_MODEL = os.environ.get("GENERATE_MODEL", "deepseek-coder:6.7b-instruct")
+CHAT_MODEL = os.environ.get("CHAT_MODEL", GENERATE_MODEL)
+EMBED_MODEL = os.environ.get("EMBED_MODEL", "nomic-embed-text:latest")
 SRC_PATH=os.environ.get("SRC_PATH", ".")
+OLLAMA_HOST = os.environ.get("OLLAMA_HOST", "http://host.docker.internal:11434")
+PROG_LANG = os.environ.get("PROG_LANG", "go")
+
+try:
+    CHUNCK_SIZE = int(os.environ.get("CHUNCK_SIZE", "300"))
+except (ValueError, TypeError):
+    CHUNCK_SIZE = 300
+
+try:
+    CHUNCK_OVERLAP = int(os.environ.get("CHUNCK_OVERLAP", "50"))
+except (ValueError, TypeError):
+    CHUNCK_OVERLAP = 50
+
+try:
+    QUERY_TIMEOUT = int(os.environ.get("QUERY_TIMEOUT", "120"))
+except (ValueError, TypeError):
+    QUERY_TIMEOUT = 120

-# Configuration
-OLLAMA_BASE_URL = "http://127.0.0.1:11434"
-VECTORSTORE = None  # Initialisé ailleurs
 DDGS_SEARCH_ENABLED = True


 os.makedirs(PERSIST_DIR, exist_ok=True)
 os.makedirs(CACHE_DIR, exist_ok=True)

-# Configuration de l'URL de base d'Ollama
-def get_ollama_base_url():
-    """Détermine dynamiquement l'URL d'Ollama"""
-    # 1. Vérifier la variable d'environnement
-    if "OLLAMA_BASE_URL" in os.environ:
-        return os.environ["OLLAMA_BASE_URL"]
-    
-    # 2. Tester la connectivité locale
-    try:
-        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-            s.settimeout(1)
-            s.connect(("localhost", 11434))
-        return "http://localhost:11434"
-    except (socket.timeout, ConnectionRefusedError):
-        pass
-    
-    # 3. Essayer l'adresse spéciale Docker
-    try:
-        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
-            s.settimeout(1)
-            s.connect(("host.docker.internal", 11434))
-        return "http://host.docker.internal:11434"
-    except (socket.timeout, ConnectionRefusedError):
-        pass
-    # 4. Fallback pour les environnements cloud
-    return "http://127.0.0.1:11434"
-
-
-OLLAMA_BASE_URL = get_ollama_base_url()
-

 # --- Nettoyage du code ---
 def clean_code_content(content: str) -> str:
@@ -131,35 +121,39 @@ app = FastAPI()
 # --- Traitement des chemins ---


-paths = SRC_PATH.split(":")
-if not paths:
-    paths = ["."]
+src_paths_directories = SRC_PATH.split(":")
+if not src_paths_directories:
+    src_paths_directories = ["."]

 # --- Initialisation ---
 vectorstore = None
 code_hash = ""

 def build_vectorstore():
-    global vectorstore, code_hash, paths
+    global vectorstore, code_hash, src_paths_directories
    print("🔹 Construction du vectorstore ...", file=sys.stderr)
    
+
    # Hash du code pour hot-reload
-    new_hash = hash_code_dir(paths)
+    new_hash = hash_code_dir(src_paths_directories)
    if vectorstore and new_hash == code_hash:
        print("🔹 Pas de changement dans /code, utilisation du vectorstore existant", file=sys.stderr)
        return
    code_hash = new_hash

+    print(f"  🔹 Programmation language: {PROG_LANG}", file=sys.stderr)
+    print(f"  🔹 Chunck size: {CHUNCK_SIZE}", file=sys.stderr)
+    print(f"  🔹 Chunck overlap: {CHUNCK_OVERLAP}", file=sys.stderr)
    # Text splitter optimisé Go
-    go_splitter = RecursiveCharacterTextSplitter.from_language(
-        language="go",
-        chunk_size=800,
-        chunk_overlap=150 #,
-        #separators=["\n\n", "\nfunc ", "}\n\n", "\n//", "\n/*", "\t"]
+    code_splitter = RecursiveCharacterTextSplitter.from_language(
+        language=PROG_LANG,
+        chunk_size=CHUNCK_SIZE,
+        chunk_overlap=CHUNCK_OVERLAP, 
+        keep_separator=True 
    )

    all_docs = []
-    for path in paths:
+    for path in src_paths_directories:
        abs_path = os.path.join("/code", path) if path != "." else "/code"
        print(f"   🔹 Chargement du code Go depuis: {abs_path}", file=sys.stderr)
        loader = DirectoryLoader(
@@ -176,32 +170,69 @@ def build_vectorstore():
        all_docs.extend(loaded_docs)

    print(f"🔹 {len(all_docs)} documents après chargement", file=sys.stderr)
-    splits = go_splitter.split_documents(all_docs)
-    print(f"🔹 {len(splits)} chunks créés", file=sys.stderr)

-    embedding = NomicEmbeddingsWrapper(model=EMBED_MODEL, base_url=OLLAMA_BASE_URL)
+    embedding = NomicEmbeddingsWrapper(model=EMBED_MODEL, base_url=OLLAMA_HOST)
+
+    chat_collection = Chroma.from_documents(
+        documents=all_docs,
+        embedding=embedding,
+        persist_directory=PERSIST_DIR,
+        collection_metadata={"hnsw:space": "cosine"},
+        collection_name="chat_context"
+    )
+
+    splits = code_splitter.split_documents(all_docs)
+    print(f"🔹 {len(splits)} chunks créés", file=sys.stderr)

    splits = [doc for doc in splits if doc.page_content.strip()]

+    # Ajout des statistiques de longueur du split
+    bins=50
+    counter = Counter(int(len(split.page_content) / bins) * bins for split in splits)
+    print("🔹 Histogramme de la longueur des splits :", file=sys.stderr)
+    max_value = max(counter.values())
+    for length, count in sorted(counter.items()):
+        normalized_count = int((count / max_value) * 50)
+        print(f"   {length+1:6}-{length+bins:-6}: { '#' * normalized_count}", file=sys.stderr)
+
    print(f"🔹 {len(splits)} fragments non vides à intégrer", file=sys.stderr)
    # Créer ou recharger Chroma
-    vectorstore = Chroma.from_documents(
-        documents=splits,
+
+    # 2. Vectorstore pour la Génération (splits courts)
+    gen_collection = Chroma.from_documents(
+        documents=splits,  # Morceaux de 200-400 tokens
        embedding=embedding,
+        collection_name="code_completion",
        persist_directory=PERSIST_DIR,
        collection_metadata={"hnsw:space": "cosine"}
    )

+    vectorstore = {
+        "chat": chat_collection,
+        "generate": gen_collection
+       }
+
    print("🔹 Vectorstore créé", file=sys.stderr)

 # --- Formatage du contexte ---
 def format_context(docs: list) -> str:
    context = []
+            
+    extraits = {}
+    print("🔹 Les fichiers suivants ont été selectionnés:", file=sys.stderr)
+
    for i, doc in enumerate(docs):
        source = doc.metadata.get('source', 'unknown')
        filename = os.path.basename(source)
+        print(f"  🔹 {filename} -- extrait {i+1} --", file=sys.stderr)
+        if doc.page_content not in extraits:
+            extraits[doc.page_content] = True
            context.append(f"### Fichier: {filename} (Extrait {i+1}) ###")
            context.append(textwrap.indent(doc.page_content, '    '))
+            print(f"  🔹 {filename} -- fin extrait {i+1} --", file=sys.stderr)
+        else:
+            print(f"  🔸 {filename} -- extrait {i+1} duppliqué et éliminé --", file=sys.stderr)
+
    return "\n\n".join(context)

 def format_iso_time_with_ns():
@@ -250,8 +281,8 @@ class ChatRequest(BaseModel):
    messages: List[ChatMessage]
    format: Optional[str] = None
    options: Optional[Dict[str, Any]] = None
-    stream: bool = False
-    keep_alive: Optional[str] = None
+    stream: bool = True
+    keep_alive: Optional[Union[str, int]] = None  # Modification ici

 class EmbeddingRequest(BaseModel):
    model: str
@@ -262,11 +293,11 @@ class EmbeddingResponse(BaseModel):
    embedding: List[float]

 # Fonctions utilitaires
-async def perform_rag_search(prompt: str, k: int = 4) -> str:
+async def perform_rag_search(mode: Literal["generate", "chat"], prompt: str, k: int = 4) -> str:
    """Effectue une recherche RAG et retourne le contexte"""
    build_vectorstore()
    
-    rag_docs = vectorstore.similarity_search(prompt, k=k)
+    rag_docs = vectorstore[mode].similarity_search(prompt, k=k)
    return format_context(rag_docs) if rag_docs else "Aucun contexte trouvé."

 async def perform_web_search(prompt: str, k: int = 2) -> str:
@@ -275,171 +306,265 @@ async def perform_web_search(prompt: str, k: int = 2) -> str:
        return "Recherche web désactivée"
    
    try:
-        from duckduckgo_search import DDGS
+        from ddgs import DDGS
        with DDGS() as ddgs:
            results = list(ddgs.text(prompt, max_results=k))
-            web_info = "\n".join(f"- [{r['title']}]({r['href']}): {r['body'][:150]}..." for r in results) if results else "Aucun résultat web trouvé."
+            print(f"🔹 {len(results)} résultats trouvés sur le web", file=sys.stderr)
+            for i, r in enumerate(results):  
+                print(f" - {i+1}. {r['title']} : {r['href']}", file=sys.stderr)
+            
+            web_info = "\n".join(
+                f"- [{r['title']}]({r['href']}): {r['body'][:150]}..." 
+                for r in results
+            ) if results else "Aucun résultat web trouvé."
    except Exception as e:
        return f"Erreur recherche web: {str(e)}"

-def build_enhanced_prompt(original_prompt: str, rag_context: str, web_context: str) -> str:
-    """Construit un prompt enrichi avec les contextes"""
-    return f"""
-### CONTEXTE RAG (Code) ###
-{rag_context or "Aucun contexte code disponible"}
+def build_enhanced_prompt(
+                 mode: Literal["generate", "chat"], 
+                 question: str,
+                 rag_context: str,
+                 web_context: str):
    
-### CONTEXTE WEB ###
-{web_context or "Aucune information web disponible"}
+    if mode == "chat":
+        prompt = f"""
+# Consigne

-### QUESTION UTILISATEUR ###
-{original_prompt}
+Vous êtes un expert en programmation {PROG_LANG}. Répondez à la question en utilisant le contexte fourni (extraits de code) et les informations web si disponibles.
+Autant que possible tu indiqueras tes sources, url, nom du fichier source...
+
+# Contexte de la question:
+
+## **Contexte Code (extraits pertinents):**
+
+{rag_context}
+
+## **Informations Web:**
+
+{web_context}
+
+## **Instructions:**
+- Répondez de manière concise et précise à la question
+- Si la réponse se trouve dans le contexte code, citez le fichier et l'extrait correspondant
+- Si vous utilisez les informations web, citez la source
+- Si la question est en anglais, répondez en anglais. Sinon, en français
+- Pour les extraits de code, conservez le formatage et l'indentation
+
+# **Question** à laquelle tu dois répondre
+
+{question}
+"""
+    else:
+        prompt=f"""
+# Consigne
+
+Vous êtes un expert en programmation {PROG_LANG}. Essayer de concevoir un petit bout de code permetant de résoudre la question
+
+# Contexte de la question:
+
+## **Contexte Code (extraits pertinents):**
+
+{rag_context}
+
+## **Instructions:**
+
+- Rédigez les commentaires de code dans la même langue que le code qui vous est fourni. À défaut en anglais.
+- Nommez les variables dans la même langue que le code qui vous est fourni. À défaut en anglais.
+- Si la question est en anglais, répondez en anglais. Sinon, en français
+- Pour les extraits de code, conservez le formatage et l'indentation
+
+# **Question** à laquelle tu dois répondre
+
+{question}
 """

+    return prompt

-# Endpoints compatibles Ollama
-@app.post("/api/generate")
-async def generate_endpoint(request_data: GenerateRequest = Body(...)):
-    """Endpoint pour la génération avec gestion du streaming"""
-    try:
-        # Récupération des contextes RAG et web
-        rag_context = await perform_rag_search(request_data.prompt)
-        web_context = await perform_web_search(request_data.prompt)
+# --- Fonctions utilitaires factorisées ---
+async def build_augmented_prompt(
+                 mode: Literal["generate", "chat"], 
+                 question: str
+        ) -> str:
+    """Construit un prompt enrichi avec contextes RAG et web"""

-        # Construction du prompt enrichi
-        enhanced_prompt = build_enhanced_prompt(
-            original_prompt=request_data.prompt,
+    rag_context = await perform_rag_search(mode,question,k= 2 if mode=='chat' else 8)
+
+    if mode == "chat":
+        web_context = await perform_web_search(question)
+    else:
+        web_context = ""
+
+    return build_enhanced_prompt(
+        mode=mode,
+        question=question,
        rag_context=rag_context,
        web_context=web_context
    )

-        # Préparation du payload pour Ollama
-        ollama_payload = {
-            "model": request_data.model,
-            "prompt": enhanced_prompt,
-            "stream": request_data.stream,
-            "options": request_data.options or {}
-        }
-        
-        # Appel à Ollama
-        async with httpx.AsyncClient() as client:
-            response = await client.post(
-                f"{OLLAMA_BASE_URL}/api/generate",
-                json=ollama_payload,
-                timeout=120.0
-            )
-            response.raise_for_status()
-            
-            # Gestion des réponses NON-STREAMING
-            if not request_data.stream:
-                result = response.json()
-                return {
-                    "model": result["model"],
-                    "response": result["response"],
-                    "done": result["done"],
-                    "context": result.get("context"),
-                    "total_duration": result.get("total_duration")
-                }
-            
-            # Gestion des réponses STREAMING
-            else:
-                async def generate():
-                    """Générateur pour le streaming des résultats"""
-                    full_response = ""
+async def _stream_ollama_response(response: httpx.Response, model_name: str):
+    """Générateur pour le streaming de la réponse de chat au format Ollama"""
+    start_time = datetime.now(timezone.utc).isoformat()
    async for line in response.aiter_lines():
        if line.strip():
            try:
                chunk = json.loads(line)
                
-                                # 1. Format SSE valide avec double newline
-                                yield f"data: {json.dumps(chunk)}\n\n"
+                # Construction du message conforme à l'API Ollama
+                message_chunk = {
+                    "model": model_name,
+                    "created_at": start_time,
+                    "message": {
+                        "role": "assistant",
+                        "content": chunk.get("message", {}).get("content", "") if "message" in chunk else chunk.get("content", "")
+                    },
+                    "done": chunk.get("done", False)
+                }
                
-                                # 2. Accumuler la réponse complète pour les logs
-                                full_response += chunk.get("response", "")
+                # Ajout des champs optionnels
+                for field in ["total_duration", "load_duration", "prompt_eval_count", "eval_count"]:
+                    if field in chunk:
+                        message_chunk[field] = chunk[field]
                
-                                # 3. Envoyer périodiquement un keep-alive
-                                if random.random() < 0.1:  # 10% des chunks
-                                    yield ": keep-alive\n\n"
+                yield f"data: {json.dumps(message_chunk)}\n\n"
                
-                                # 4. Fin du stream
-                                if chunk.get("done", False):
-                                    break
            except json.JSONDecodeError:
-                                print(f"⚠️ Ligne JSON invalide: {line}")
-                                yield f"event: error\ndata: Invalid JSON line\n\n"
+                yield "event: error\ndata: Invalid JSON chunk\n\n"
    
-                    # 5. Envoyer un message de fin explicite
    yield "event: end\ndata: Stream completed\n\n"

-                    # 6. Log de la réponse complète
-                    print(f"🔹 Réponse complète ({len(full_response)} caractères): {full_response[:200]}...", file=sys.stderr)

-                    # Log de la réponse complète (optionnel)
-                    print(f"🔹 Réponse complète: {full_response}")
+# --- Endpoints API ---
+@app.post("/api/generate")
+async def generate_endpoint(request_data: GenerateRequest = Body(...)):
+    try:
+        # Construction du prompt enrichi
+        enhanced_prompt = await build_augmented_prompt("generate",request_data.prompt)
        
-                # 7. Configuration de la réponse avec des headers spécifiques
+        # Appel à Ollama
+        ollama_payload = {
+            "model": GENERATE_MODEL,
+            "prompt": enhanced_prompt,
+            "stream": request_data.stream,
+            "options": request_data.options or {}
+        }
+        
+        async with httpx.AsyncClient() as client:
+            response = await client.post(
+                f"{OLLAMA_HOST}/api/generate",
+                json=ollama_payload,
+                timeout=QUERY_TIMEOUT
+            )
+            response.raise_for_status()
+            
+            if not request_data.stream:
+                return response.json()
+            else:
+                # --- CORRECTION DU STREAMING ---
+                async def generate_stream():
+                    """Générateur pour le streaming des résultats"""
+                    async for chunk in response.aiter_text():
+                        # Transférer directement les chunks
+                        yield chunk
+                        
+                    # Fermeture propre du stream
+                    # yield "data: [DONE]\n\n"
+                
+                # Utilisez text/plain au lieu de text/event-stream
                return StreamingResponse(
-                    generate(),
-                    media_type="text/event-stream",
+                    generate_stream(),
+                    media_type="text/plain",
                    headers={
                        "Cache-Control": "no-cache",
                        "Connection": "keep-alive",
-                        "X-Accel-Buffering": "no"  # Important pour Nginx
+                        "X-Accel-Buffering": "no"
                    }
                )
                        
    except httpx.RequestError as e:
        raise HTTPException(500, f"Erreur de connexion à Ollama: {str(e)}")
-    except json.JSONDecodeError as e:
-        raise HTTPException(500, f"Erreur de décodage JSON: {str(e)}")
    except Exception as e:
        raise HTTPException(500, f"Erreur interne: {str(e)}")

-
@app.post("/api/chat")
-async def chat(request: ChatRequest):
-    """Endpoint /api/chat avec enrichissement du dernier message"""
-    # Copie profonde des messages
-    processed_messages = [msg.dict() for msg in request.messages]
+async def chat_endpoint(request_data: ChatRequest):
+    try:
+        messages = [msg.dict() for msg in request_data.messages]
        
-    # Enrichissement uniquement du dernier message utilisateur
-    if processed_messages and processed_messages[-1]["role"] == "user":
-        last_msg = processed_messages[-1]["content"]
+        if messages and messages[-1]["role"] == "user":
+            original_question = messages[-1]["content"]
+            try:
+                # Limiter la taille du contexte
+                augmented_prompt = await build_augmented_prompt("chat",original_question)
+                messages[-1]["content"] = augmented_prompt[-8000:]  # Truncate to the last 8000 chars
+                print(f"🔹 Prompt enrichi ({len(augmented_prompt)} caractères)", file=sys.stderr)
+            except Exception as e:
+                print(f"⚠️ Erreur d'enrichissement: {str(e)}", file=sys.stderr)
+                messages[-1]["content"] = original_question  # Fallback to original
        
-        rag_context = await perform_rag_search(last_msg)
-        web_context = await perform_web_search(last_msg)
+        # Préparation du payload pour Ollama
+        ollama_payload = {
+            "model": CHAT_MODEL,
+            "messages": messages,
+            "stream": request_data.stream,
+            "options": request_data.options or {}
+        }

-        enhanced_content = build_enhanced_prompt(
-            original_prompt=last_msg,
-            rag_context=rag_context,
-            web_context=web_context
-        )
+        if request_data.keep_alive is not None:
+            if isinstance(request_data.keep_alive, int):
+                ollama_payload["keep_alive"] = f"{request_data.keep_alive}s"
+            else:
+                ollama_payload["keep_alive"] = request_data.keep_alive
        
-        processed_messages[-1]["content"] = enhanced_content
+        sopload = json.dumps(ollama_payload, indent=2)
+        print(f" 🔹 Taille du Payload vers Ollama : {len(sopload)} octets...", file=sys.stderr)
+        print(f" 🔹 Début du payload : {sopload}...", file=sys.stderr)

-    # Appel au vrai Ollama
+        # Appel à Ollama
        async with httpx.AsyncClient() as client:
            try:
                response = await client.post(
-                f"{OLLAMA_BASE_URL}/api/chat",
-                json={
-                    "model": request.model,
-                    "messages": processed_messages,
-                    "format": request.format,
-                    "options": request.options,
-                    "stream": request.stream,
-                    "keep_alive": request.keep_alive
-                },
-                timeout=120.0
+                    f"{OLLAMA_HOST}/api/chat",
+                    json=ollama_payload,
+                    timeout=QUERY_TIMEOUT
                )
                response.raise_for_status()
                
-            if request.stream:
-                return response.iter_lines()
-            
+                # Gestion des réponses NON-STREAMING
+                if not request_data.stream:
                    return response.json()
                
+                # --- CORRECTION DU STREAMING ---
+                async def generate_stream():
+                    """Générateur pour le streaming des résultats"""
+                    async for chunk in response.aiter_text():
+                        # Transférer directement les chunks
+                        yield chunk
+                        
+                    # Fermeture propre du stream
+                    # yield "data: [DONE]\n\n"
+                
+                # Utilisez text/plain au lieu de text/event-stream
+                return StreamingResponse(
+                    generate_stream(),
+                    media_type="text/plain",
+                    headers={
+                        "Cache-Control": "no-cache",
+                        "Connection": "keep-alive",
+                        "X-Accel-Buffering": "no"
+                    }
+                )
+    # ... [gestion des erreurs existante] ...
+            except httpx.HTTPStatusError as e:
+                error_detail = e.response.text if e.response else str(e)
+                print(f"🚨 Erreur HTTP Ollama ({e.response.status_code}): {error_detail}", file=sys.stderr)
+                raise HTTPException(502, f"Erreur Ollama: {error_detail}")
            except httpx.RequestError as e:
-            raise HTTPException(500, f"Erreur de connexion à Ollama: {str(e)}")
+                print(f"🚨 Erreur réseau Ollama: {str(e)}", file=sys.stderr)
+                raise HTTPException(503, f"Ollama non disponible: {str(e)}")
+                
+    except Exception as e:
+        print(f"🚨 Erreur interne: {traceback.format_exc()}", file=sys.stderr)
+        raise HTTPException(500, f"Erreur interne: {str(e)}")

@app.post("/api/embeddings")
 async def embeddings(request: EmbeddingRequest):
@@ -447,7 +572,7 @@ async def embeddings(request: EmbeddingRequest):
    async with httpx.AsyncClient() as client:
        try:
            response = await client.post(
-                f"{OLLAMA_BASE_URL}/api/embeddings",
+                f"{OLLAMA_HOST}/api/embeddings",
                json=request.dict()
            )
            response.raise_for_status()
@@ -461,7 +586,7 @@ async def list_models():
    """Proxy pour lister les modèles disponibles"""
    async with httpx.AsyncClient() as client:
        try:
-            response = await client.get(f"{OLLAMA_BASE_URL}/api/tags")
+            response = await client.get(f"{OLLAMA_HOST}/api/tags")
            response.raise_for_status()
            return response.json()
            
--- a/LLM/Docker/app_old.py
+++ b/LLM/Docker/app_old.py
@@ -1,338 +0,0 @@
-import os
-import sys
-import glob
-import time
-import hashlib
-import textwrap
-import requests
-import httpx
-from fastapi.responses import StreamingResponse
-from fastapi import FastAPI, Query, HTTPException, Request
-from langchain_community.vectorstores import Chroma
-from langchain_ollama import OllamaEmbeddings
-from langchain_community.document_loaders import DirectoryLoader, TextLoader
-from langchain_text_splitters import RecursiveCharacterTextSplitter
-from duckduckgo_search import DDGS
-from unstructured.cleaners.core import clean_extra_whitespace, clean_non_ascii_chars, replace_unicode_quotes
-from datetime import datetime, timezone, timedelta
-
-
-# --- Configuration via variables d'environnement ---
-PERSIST_DIR = os.environ.get("CHROMA_PERSIST_DIR", "/chroma_db")
-CACHE_DIR = os.environ.get("RESPONSE_CACHE_DIR", "/response_cache")
-OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434")
-MODEL_NAME = os.environ.get("OLLAMA_MODEL", "llama3:13b")
-
-os.makedirs(PERSIST_DIR, exist_ok=True)
-os.makedirs(CACHE_DIR, exist_ok=True)
-
-# --- Nettoyage du code ---
-def clean_code_content(content: str) -> str:
-    cleaned = replace_unicode_quotes(content)
-    cleaned = clean_non_ascii_chars(cleaned)
-    cleaned = clean_extra_whitespace(cleaned)
-    return cleaned
-
-# --- Cache simple ---
-def get_cache_key(question: str) -> str:
-    return hashlib.md5(question.encode()).hexdigest()
-
-# --- Hot-reload : hash du code ---
-def hash_code_dir(paths: list) -> str:
-    m = hashlib.md5()
-    for path in paths:
-        abs_path = os.path.join("/code", path) if path != "." else "/code"
-        for f in glob.glob(f"{abs_path}/**/*.go", recursive=True):
-            try:
-                with open(f, "rb") as file:
-                    m.update(file.read())
-            except Exception:
-                continue
-    return m.hexdigest()
-
-# --- Wrapper Nomic Embeddings ---
-from typing import List
-
-class NomicEmbeddingsWrapper(OllamaEmbeddings):
-    """Wrapper automatique pour les préfixes Nomic"""
-    
-    def _prefix_text(self, text: str, is_document: bool) -> str:
-        prefix = "search_document: " if is_document else "search_query: "
-        return prefix + text
-    
-    def embed_documents(self, texts: List[str]) -> List[List[float]]:
-        prefixed_texts = [self._prefix_text(t, is_document=True) for t in texts]
-        return super().embed_documents(prefixed_texts)
-    
-    def embed_query(self, text: str) -> List[float]:
-        return super().embed_query(self._prefix_text(text, is_document=False))
-
-# --- FastAPI ---
-app = FastAPI()
-
-# --- Traitement des chemins ---
-paths = sys.argv[1:] if len(sys.argv) > 1 else ["."]
-if not paths:
-    paths = ["."]
-
-# --- Initialisation ---
-vectorstore = None
-code_hash = ""
-
-def build_vectorstore():
-    global vectorstore, code_hash
-    print("🔹 Construction du vectorstore...", file=sys.stderr)
-
-    # Hash du code pour hot-reload
-    new_hash = hash_code_dir(paths)
-    if vectorstore and new_hash == code_hash:
-        print("🔹 Pas de changement dans /code, utilisation du vectorstore existant", file=sys.stderr)
-        return
-    code_hash = new_hash
-
-    # Text splitter optimisé Go
-    go_splitter = RecursiveCharacterTextSplitter.from_language(
-        language="go",
-        chunk_size=800,
-        chunk_overlap=150,
-        separators=["\n\n", "\nfunc ", "}\n\n", "\n//", "\n/*", "\t"]
-    )
-
-    all_docs = []
-    for path in paths:
-        abs_path = os.path.join("/code", path) if path != "." else "/code"
-        print(f"   🔹 Chargement du code Go depuis: {abs_path}", file=sys.stderr)
-        loader = DirectoryLoader(
-            abs_path,
-            glob="**/*.go",
-            loader_cls=TextLoader,
-            use_multithreading=True,
-            loader_kwargs={'autodetect_encoding': True},
-            max_files=500
-        )
-        loaded_docs = loader.load()
-        print(f"   🔸 {len(loaded_docs)} fichiers chargés", file=sys.stderr)
-        for doc in loaded_docs:
-            doc.page_content = clean_code_content(doc.page_content)
-        all_docs.extend(loaded_docs)
-
-    print(f"🔹 {len(all_docs)} documents après chargement", file=sys.stderr)
-    splits = go_splitter.split_documents(all_docs)
-    print(f"🔹 {len(splits)} chunks créés", file=sys.stderr)
-
-    embedding = NomicEmbeddingsWrapper(model="nomic-embed-text", api_base=OLLAMA_URL)
-
-    # Créer ou recharger Chroma
-    vectorstore = Chroma.from_documents(
-        documents=splits,
-        embedding=embedding,
-        persist_directory=PERSIST_DIR,
-        collection_metadata={"hnsw:space": "cosine"}
-    )
-    vectorstore.persist()
-    print("🔹 Vectorstore créé et persisté", file=sys.stderr)
-
-# --- Formatage du contexte ---
-def format_context(docs: list) -> str:
-    context = []
-    for i, doc in enumerate(docs):
-        source = doc.metadata.get('source', 'unknown')
-        filename = os.path.basename(source)
-        context.append(f"### Fichier: {filename} (Extrait {i+1}) ###")
-        context.append(textwrap.indent(doc.page_content, '    '))
-    return "\n\n".join(context)
-
-def format_iso_time_with_ns():
-    # 1. Obtenir le timestamp actuel avec nanosecondes
-    current_time_ns = time.time_ns()
-    
-    # 2. Convertir en datetime avec timezone locale
-    dt = datetime.fromtimestamp(current_time_ns / 1e9).astimezone()
-    
-    # 3. Formater avec les nanosecondes et décalage horaire
-    # - Extraire les nanosecondes
-    nanoseconds = current_time_ns % 10**9
-    
-    # - Formater la partie datetime de base
-    base_format = dt.strftime("%Y-%m-%dT%H:%M:%S")
-    
-    # - Ajouter les nanosecondes (9 chiffres)
-    nano_format = f".{nanoseconds:09d}"
-    
-    # - Formater le décalage horaire
-    utc_offset = dt.utcoffset()
-    offset_hours = utc_offset.total_seconds() // 3600
-    offset_minutes = (utc_offset.total_seconds() % 3600) // 60
-    offset_sign = '-' if offset_hours < 0 else '+'
-    offset_format = f"{offset_sign}{abs(int(offset_hours)):02d}:{int(offset_minutes):02d}"
-    
-    return base_format + nano_format + offset_format
-
-def build_prompt(question: str,
-                 k_rag: int,
-                 k_web: int):
-    build_vectorstore()
-
-    rag_docs = vectorstore.similarity_search(question, k=k_rag)
-    context_str = format_context(rag_docs) if rag_docs else "Aucun contexte trouvé."
-
-    # Recherche web
-    web_info = ""
-    if k_web > 0:
-        try:
-            with DDGS(timeout=10) as ddgs:
-                results = list(ddgs.text(question, max_results=k_web))
-                web_info = "\n".join(f"- [{r['title']}]({r['href']}): {r['body'][:150]}..." for r in results) if results else "Aucun résultat web trouvé."
-        except Exception as e:
-            web_info = f"⚠️ Erreur recherche web: {str(e)}"
-    else:
-        web_info = "Recherche web désactivée."
-
-    prompt = f"""
-Vous êtes un expert en programmation Go. Répondez à la question en utilisant le contexte fourni (extraits de code) et les informations web si disponibles.
-
-**Contexte Code (extraits pertinents):**
-{context_str}
-
-**Informations Web:**
-{web_info}
-
-**Question:**
-{question}
-
-**Instructions:**
- Répondez de manière concise et précise
- Si la réponse se trouve dans le contexte code, citez le fichier et l'extrait correspondant
- Si vous utilisez les informations web, citez la source
- Si la question est en anglais, répondez en anglais. Sinon, en français
- Pour les extraits de code, conservez le formatage et l'indentation
-"""
-
-    return prompt
-
-@app.post("/api/chat")
-async def chat(
-    question: str = Query(..., min_length=3),
-    history: list = Query(default=[]),  # liste d'anciens messages [{role, content}]
-    k_rag: int = Query(4, ge=1, le=10),
-    k_web: int = Query(2, ge=0, le=5),
-):
-    start_time = time.time()
-
-    # Construire le prompt enrichi (RAG + Web)
-    prompt = build_prompt(question=question, k_rag=k_rag, k_web=k_web)
-
-    # Construire la conversation pour Ollama
-    messages = history + [
-        {"role": "user", "content": prompt}
-    ]
-
-    try:
-        r = requests.post(
-            f"{OLLAMA_URL}/api/chat",
-            json={
-                "model": MODEL_NAME,
-                "messages": messages,
-                "options": {
-                    "temperature": 0.3,
-                    "num_predict": 1024,
-                    "top_k": 50,
-                    "top_p": 0.9
-                }
-            },
-            timeout=120
-        )
-        r.raise_for_status()
-        result = r.json()
-        answer = result.get("message", {}).get("content", "Pas de réponse générée.")
-
-        return {
-            "answer": answer,
-            "processing_time": f"{time.time() - start_time:.2f}s",
-            "model": MODEL_NAME,
-            "cached": False,
-            "history": messages + [{"role": "assistant", "content": answer}]
-        }
-    except requests.exceptions.RequestException as e:
-        detail = f"Erreur API Ollama: {str(e)}"
-        if hasattr(e, 'response') and e.response:
-            detail += f" | Status: {e.response.status_code} | Response: {e.response.text[:200]}"
-        raise HTTPException(status_code=500, detail=detail)
-
-# --- Endpoint /ask ---
-@app.get("/api/generate")
-async def ask_question(
-    question: str = Query(..., min_length=3),
-    k_rag: int = Query(4, ge=1, le=10),
-    k_web: int = Query(2, ge=0, le=5),
-    use_cache: bool = Query(True)
-):
-    start_time = time.time_ns()
-
-    cache_path = os.path.join(CACHE_DIR, f"{get_cache_key(question)}.txt")
-    if use_cache and os.path.exists(cache_path):
-        with open(cache_path, "r") as f:
-            return {"answer": f.read(), "cached": True}
-
-    prompt = build_prompt(question=question, k_rag=k_rag, k_web=k_web)
-
-    try:
-        r = requests.post(f"{OLLAMA_URL}/api/generate", json={
-            "model": MODEL_NAME,
-            "prompt": prompt,
-            "stream": False,
-            "options": {"temperature": 0.3, "num_predict": 1024, "top_k": 50, "top_p": 0.9}
-        }, timeout=120)
-        r.raise_for_status()
-        result = r.json()
-        answer = result.get("response") or result.get("text") or "Pas de réponse générée."
-
-        with open(cache_path, "w") as f:
-            f.write(answer)
-
-        return {
-                "model": MODEL_NAME,
-                "eval_duration": f"{time.time_ns() - start_time:.0f}",
-                "created_at": format_iso_time_with_ns(),
-                "response": answer,
-                "done": false
-               }
-        
-        
-        {"answer": answer, 
-                "processing_time": f"{time.time() - start_time:.2f}s", 
-                "model": MODEL_NAME, 
-                "cached": False}
-    except requests.exceptions.RequestException as e:
-        detail = f"Erreur API Ollama: {str(e)}"
-        if hasattr(e, 'response') and e.response:
-            detail += f" | Status: {e.response.status_code} | Response: {e.response.text[:200]}"
-        raise HTTPException(status_code=500, detail=detail)
-
-# --- Endpoint /status ---
-@app.get("/status")
-def status_check():
-    try:
-        count = vectorstore._collection.count() if vectorstore else 0
-        return {
-            "status": "OK",
-            "vectorstore_items": count,
-            "model": MODEL_NAME,
-            "persist_dir": PERSIST_DIR,
-            "cache_dir": CACHE_DIR
-        }
-    except Exception as e:
-        raise HTTPException(500, f"Erreur: {str(e)}")
-
-@app.api_route("/{path:path}", methods=["GET", "POST", "PUT", "DELETE", "PATCH"])
-async def proxy(request: Request, path: str):
-    async with httpx.AsyncClient() as client:
-        url = f"{OLLAMA_URL}/{path}"
-        body = await request.body()
-        r = await client.request(
-            method=request.method,
-            url=url,
-            headers=request.headers,
-            content=body
-        )
-        return StreamingResponse(r.aiter_bytes(), status_code=r.status_code, headers=dict(r.headers))
--- a/LLM/Docker/docker-compose.yml
+++ b/LLM/Docker/docker-compose.yml
@@ -7,8 +7,9 @@ services:
      - "8000:8000"
    environment:
      # Variables Ollama / FastAPI
-      OLLAMA_API_URL: "http://localhost:11434/api/generate"
-      OLLAMA_MODEL: "deepseek-coder:33b"
+      OLLAMA_HOST: "http://host.docker.internal:11434"
+      GENERATE_MODEL: "deepseek-coder:6.7b-instruct"
+      CHAT_MODEL: "gpt-oss:20b"
      CHROMA_PERSIST_DIR: "/chroma_db"
      RESPONSE_CACHE_DIR: "/response_cache"
      PYTHONUNBUFFERED: 1
--- a/LLM/Docker/requirements.txt
+++ b/LLM/Docker/requirements.txt
@@ -1,10 +1,11 @@
 fastapi==0.116.1
 uvicorn[standard]==0.23.2
 chromadb==1.0.15
-duckduckgo-search==8.1.1
+ddgs==9.5.4
 langchain==0.3.27
 langchain-community==0.3.27
 sentence-transformers==5.1.0
+langchain-chroma==0.2.5
 unstructured==0.7.10
 langchain-ollama==0.3.6
 requests==2.32.1
--- a/LLM/Docker/start.sh
+++ b/LLM/Docker/start.sh
@@ -2,20 +2,22 @@
 set -e

 # Lancer Ollama en arrière-plan
-export OLLAMA_MODELS=/models
-echo "🔹 Démarrage de Ollama..." 1>&2
-ollama serve | sed 's/^/ 🔹[Ollama server] /'  &
+# export OLLAMA_MODELS=/models
+# echo "🔹 Démarrage de Ollama..." 1>&2
+# ollama serve 2>&1 \
+# | grep -vF "decode: cannot decode batches with this context (use llama_encode() instead)" \
+# | sed 's/^/ 🔹[Ollama server] /' 1>&2 &

-sleep 10
+# sleep 10
+
+if [[ -n "$1" ]] ; then 
+    eval $*
+fi

 echo "🔹 Preaload Ollama models: "
 ollama ls  | sed 's/^/ 🔹 /' 1>&2

-# Attendre Ollama
-sleep 5
-
-# Vérifier / précharger le modèle Nomic Embed Text
-EMBED_MODEL="nomic-embed-text:latest"
+# Vérifier / précharger le modèle d'embedding

 echo "🔹 Vérification du modèle d'embedding: $EMBED_MODEL" 1>&2
 if ! ollama list | grep -q "$EMBED_MODEL"; then