lojol469-cmd commited on
Commit ·
f3a56a5
1
Parent(s): b54b689
Initial commit: Kibali AI with RTX 5090 Blackwell support and CUDA 13.0 Nightly
Browse files- Dockerfile +21 -11
- main.py +8 -3
- requirements.txt +5 -8
- tools/todo.py +24 -156
- tools/web.py +7 -8
Dockerfile
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
# --- STAGE 1 : Build du Frontend
|
| 2 |
FROM node:18-alpine AS build-frontend
|
| 3 |
WORKDIR /app/frontend
|
| 4 |
COPY kibali-ui/package*.json ./
|
|
@@ -6,31 +6,41 @@ RUN npm install
|
|
| 6 |
COPY kibali-ui/ ./
|
| 7 |
RUN npm run build
|
| 8 |
|
| 9 |
-
# --- STAGE 2 : Backend
|
| 10 |
-
|
|
|
|
| 11 |
WORKDIR /app
|
| 12 |
|
| 13 |
-
|
|
|
|
| 14 |
RUN apt-get update && apt-get install -y \
|
| 15 |
python3-pip \
|
| 16 |
python3-dev \
|
|
|
|
|
|
|
| 17 |
&& rm -rf /var/lib/apt/lists/*
|
| 18 |
|
| 19 |
-
#
|
| 20 |
-
|
|
|
|
|
|
|
| 21 |
|
| 22 |
-
# Installation des dépendances
|
| 23 |
COPY requirements.txt .
|
| 24 |
RUN pip3 install --no-cache-dir -r requirements.txt
|
| 25 |
|
| 26 |
-
# On
|
| 27 |
-
|
| 28 |
|
| 29 |
-
|
| 30 |
COPY . .
|
| 31 |
|
|
|
|
|
|
|
|
|
|
| 32 |
ENV PYTHONUNBUFFERED=1
|
|
|
|
|
|
|
| 33 |
EXPOSE 8000
|
| 34 |
|
| 35 |
-
# Commande corrigée pour Ubuntu (python3)
|
| 36 |
CMD ["python3", "-m", "uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]
|
|
|
|
| 1 |
+
# --- STAGE 1 : Build du Frontend ---
|
| 2 |
FROM node:18-alpine AS build-frontend
|
| 3 |
WORKDIR /app/frontend
|
| 4 |
COPY kibali-ui/package*.json ./
|
|
|
|
| 6 |
COPY kibali-ui/ ./
|
| 7 |
RUN npm run build
|
| 8 |
|
| 9 |
+
# --- STAGE 2 : Backend (Base NVIDIA Blackwell Compatible) ---
|
| 10 |
+
# On utilise une base 12.6 qui supporte les drivers de la série 50
|
| 11 |
+
FROM nvidia/cuda:12.6.1-runtime-ubuntu22.04
|
| 12 |
WORKDIR /app
|
| 13 |
|
| 14 |
+
ENV DEBIAN_FRONTEND=noninteractive
|
| 15 |
+
|
| 16 |
RUN apt-get update && apt-get install -y \
|
| 17 |
python3-pip \
|
| 18 |
python3-dev \
|
| 19 |
+
libgomp1 \
|
| 20 |
+
git \
|
| 21 |
&& rm -rf /var/lib/apt/lists/*
|
| 22 |
|
| 23 |
+
# INSTALLATION PYTORCH NIGHTLY CUDA 13.0
|
| 24 |
+
# C'est ici qu'on débloque le support sm_120
|
| 25 |
+
RUN pip3 install --no-cache-dir --upgrade pip
|
| 26 |
+
RUN pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130
|
| 27 |
|
| 28 |
+
# Installation du reste des dépendances
|
| 29 |
COPY requirements.txt .
|
| 30 |
RUN pip3 install --no-cache-dir -r requirements.txt
|
| 31 |
|
| 32 |
+
# On force une version récente de transformers pour le tokenizer Blackwell
|
| 33 |
+
RUN pip3 install --upgrade transformers accelerate bitsandbytes
|
| 34 |
|
| 35 |
+
COPY --from=build-frontend /app/frontend/dist ./static
|
| 36 |
COPY . .
|
| 37 |
|
| 38 |
+
RUN mkdir -p /app/model_cache
|
| 39 |
+
|
| 40 |
+
ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
|
| 41 |
ENV PYTHONUNBUFFERED=1
|
| 42 |
+
ENV MODEL_PATH=/app/model_cache
|
| 43 |
+
|
| 44 |
EXPOSE 8000
|
| 45 |
|
|
|
|
| 46 |
CMD ["python3", "-m", "uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]
|
main.py
CHANGED
|
@@ -55,11 +55,16 @@ app.add_middleware(
|
|
| 55 |
)
|
| 56 |
|
| 57 |
# --- CHARGEMENT DES MODÈLES ---
|
| 58 |
-
MODEL_PATH = "/
|
|
|
|
|
|
|
| 59 |
logger.info("Chargement du modèle d'embedding...")
|
| 60 |
-
|
|
|
|
|
|
|
| 61 |
logger.info("Chargement du tokenizer et du modèle LLM...")
|
| 62 |
-
|
|
|
|
| 63 |
if tokenizer.pad_token is None:
|
| 64 |
tokenizer.pad_token = tokenizer.eos_token
|
| 65 |
|
|
|
|
| 55 |
)
|
| 56 |
|
| 57 |
# --- CHARGEMENT DES MODÈLES ---
|
| 58 |
+
MODEL_PATH = os.getenv("MODEL_PATH", "./model_cache")
|
| 59 |
+
logger.info(f"Utilisation du chemin modèle : {MODEL_PATH}")
|
| 60 |
+
|
| 61 |
logger.info("Chargement du modèle d'embedding...")
|
| 62 |
+
# Utilisation du cache_folder pour que SentenceTransformer stocke aussi dans le volume partagé
|
| 63 |
+
embed_model = SentenceTransformer('paraphrase-multilingual-MiniLM-L12-v2', cache_folder=MODEL_PATH)
|
| 64 |
+
|
| 65 |
logger.info("Chargement du tokenizer et du modèle LLM...")
|
| 66 |
+
# Suppression de local_files_only=True pour permettre la compatibilité initiale avec nouvelles architectures GPU
|
| 67 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
|
| 68 |
if tokenizer.pad_token is None:
|
| 69 |
tokenizer.pad_token = tokenizer.eos_token
|
| 70 |
|
requirements.txt
CHANGED
|
@@ -1,9 +1,7 @@
|
|
| 1 |
-
# --- Core IA
|
| 2 |
-
# Note: On ne met pas de version figée pour torch ici,
|
| 3 |
-
# car on l'installe via l'URL spécifique dans le Dockerfile.
|
| 4 |
transformers==4.41.2
|
| 5 |
-
bitsandbytes>=0.
|
| 6 |
-
accelerate
|
| 7 |
sentence-transformers
|
| 8 |
faiss-gpu
|
| 9 |
|
|
@@ -14,9 +12,8 @@ pydantic
|
|
| 14 |
python-multipart
|
| 15 |
|
| 16 |
# --- Outils & Data ---
|
|
|
|
| 17 |
pypdf>=3.0.0
|
| 18 |
numpy<2.0.0
|
| 19 |
-
folium
|
| 20 |
duckduckgo-search
|
| 21 |
-
huggingface_hub
|
| 22 |
-
spaces
|
|
|
|
| 1 |
+
# --- Core IA ---
|
|
|
|
|
|
|
| 2 |
transformers==4.41.2
|
| 3 |
+
bitsandbytes>=0.43.0
|
| 4 |
+
accelerate>=0.30.0
|
| 5 |
sentence-transformers
|
| 6 |
faiss-gpu
|
| 7 |
|
|
|
|
| 12 |
python-multipart
|
| 13 |
|
| 14 |
# --- Outils & Data ---
|
| 15 |
+
tavily-python
|
| 16 |
pypdf>=3.0.0
|
| 17 |
numpy<2.0.0
|
|
|
|
| 18 |
duckduckgo-search
|
| 19 |
+
huggingface_hub
|
|
|
tools/todo.py
CHANGED
|
@@ -1,7 +1,10 @@
|
|
| 1 |
-
import streamlit as st
|
| 2 |
import time
|
| 3 |
from typing import List, Optional
|
| 4 |
import re
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
|
| 6 |
def analyze_query_type(prompt: str) -> dict:
|
| 7 |
"""Analyse le type de requête pour adapter la stratégie de réflexion"""
|
|
@@ -17,36 +20,30 @@ def analyze_query_type(prompt: str) -> dict:
|
|
| 17 |
"geographical": False
|
| 18 |
}
|
| 19 |
|
| 20 |
-
# Détection de questions temporelles
|
| 21 |
temporal_keywords = ["aujourd'hui", "maintenant", "récent", "actuel", "dernier", "2024", "2025"]
|
| 22 |
if any(kw in prompt_lower for kw in temporal_keywords):
|
| 23 |
analysis["temporal"] = True
|
| 24 |
analysis["needs_web"] = True
|
| 25 |
|
| 26 |
-
# Détection géographique
|
| 27 |
geo_keywords = ["gabon", "libreville", "port-gentil", "franceville", "oyem", "où", "localisation"]
|
| 28 |
if any(kw in prompt_lower for kw in geo_keywords):
|
| 29 |
analysis["geographical"] = True
|
| 30 |
|
| 31 |
-
# Détection de questions sur documents
|
| 32 |
doc_keywords = ["selon le document", "d'après le pdf", "dans le fichier", "uploadé"]
|
| 33 |
if any(kw in prompt_lower for kw in doc_keywords):
|
| 34 |
analysis["needs_docs"] = True
|
| 35 |
analysis["type"] = "document_query"
|
| 36 |
|
| 37 |
-
# Détection de continuation de conversation
|
| 38 |
continuation_keywords = ["ils", "elles", "lui", "leur", "donc", "alors", "ensuite", "aussi", "également"]
|
| 39 |
if any(kw in prompt_lower for kw in continuation_keywords) or len(prompt.split()) < 5:
|
| 40 |
analysis["needs_memory"] = True
|
| 41 |
analysis["type"] = "continuation"
|
| 42 |
|
| 43 |
-
|
| 44 |
-
if len(prompt.split()) > 15 or "?" in prompt and prompt.count("?") > 1:
|
| 45 |
analysis["complexity"] = "complex"
|
| 46 |
-
elif
|
| 47 |
analysis["complexity"] = "medium"
|
| 48 |
|
| 49 |
-
# Questions nécessitant le web
|
| 50 |
web_keywords = ["actualité", "news", "prix", "cours", "météo", "horaire"]
|
| 51 |
if any(kw in prompt_lower for kw in web_keywords):
|
| 52 |
analysis["needs_web"] = True
|
|
@@ -57,71 +54,44 @@ def analyze_query_type(prompt: str) -> dict:
|
|
| 57 |
def detect_subject_shift(prompt: str, current_subject: str, subject_keywords: List[str]) -> dict:
|
| 58 |
"""Détecte un changement de sujet et évalue la force du changement"""
|
| 59 |
if not current_subject or not subject_keywords:
|
| 60 |
-
return {
|
| 61 |
-
"shift_detected": False,
|
| 62 |
-
"shift_strength": 0.0,
|
| 63 |
-
"new_subject_detected": True,
|
| 64 |
-
"reason": "Premier message ou pas de sujet actuel"
|
| 65 |
-
}
|
| 66 |
|
| 67 |
prompt_lower = prompt.lower()
|
| 68 |
-
|
| 69 |
-
# Calcul de l'overlap des mots-clés
|
| 70 |
prompt_words = set(re.findall(r'\b\w{4,}\b', prompt_lower))
|
| 71 |
keyword_overlap = len(prompt_words.intersection(set(subject_keywords)))
|
| 72 |
overlap_ratio = keyword_overlap / max(len(subject_keywords), 1)
|
| 73 |
|
| 74 |
-
# Détection de marqueurs de changement de sujet
|
| 75 |
shift_markers = ["maintenant", "sinon", "autre chose", "parlons de", "passons à", "nouveau sujet"]
|
| 76 |
has_shift_marker = any(marker in prompt_lower for marker in shift_markers)
|
| 77 |
|
| 78 |
-
# Calcul de la force du changement
|
| 79 |
shift_strength = 0.0
|
| 80 |
-
if overlap_ratio < 0.2:
|
| 81 |
-
|
| 82 |
-
if has_shift_marker:
|
| 83 |
-
shift_strength += 0.3
|
| 84 |
-
if len(prompt_words) > 5 and keyword_overlap == 0:
|
| 85 |
-
shift_strength += 0.2
|
| 86 |
-
|
| 87 |
-
shift_detected = shift_strength > 0.4
|
| 88 |
|
| 89 |
return {
|
| 90 |
-
"shift_detected":
|
| 91 |
"shift_strength": shift_strength,
|
| 92 |
"new_subject_detected": shift_strength > 0.6,
|
| 93 |
-
"
|
| 94 |
-
"overlap_ratio": overlap_ratio,
|
| 95 |
-
"reason": f"Overlap: {overlap_ratio:.1%}, Marqueurs: {has_shift_marker}"
|
| 96 |
}
|
| 97 |
|
| 98 |
def generate_search_strategy(analysis: dict, subject_keywords: List[str], geo_info: dict) -> dict:
|
| 99 |
"""Génère une stratégie de recherche optimisée"""
|
| 100 |
strategy = {
|
| 101 |
"use_rag": analysis["needs_docs"],
|
| 102 |
-
"use_memory": analysis["needs_memory"]
|
| 103 |
-
"use_web": analysis["needs_web"]
|
| 104 |
-
"memory_k": 5,
|
| 105 |
-
"
|
| 106 |
-
"web_enhanced": False,
|
| 107 |
-
"search_query_suffix": ""
|
| 108 |
}
|
| 109 |
|
| 110 |
-
# Ajustement selon la complexité
|
| 111 |
if analysis["complexity"] == "complex":
|
| 112 |
-
strategy
|
| 113 |
-
strategy["rag_k"] = 5
|
| 114 |
-
elif analysis["complexity"] == "simple":
|
| 115 |
-
strategy["memory_k"] = 3
|
| 116 |
-
strategy["rag_k"] = 2
|
| 117 |
|
| 118 |
-
# Enrichissement de la recherche web
|
| 119 |
if analysis["needs_web"]:
|
| 120 |
strategy["web_enhanced"] = True
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
if analysis["geographical"]:
|
| 124 |
-
strategy["search_query_suffix"] += f" {geo_info.get('city', 'Gabon')}"
|
| 125 |
|
| 126 |
return strategy
|
| 127 |
|
|
@@ -132,118 +102,16 @@ def execute_reflection_plan(
|
|
| 132 |
current_subject: Optional[str] = None,
|
| 133 |
subject_keywords: Optional[List[str]] = None
|
| 134 |
):
|
| 135 |
-
"""
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
if geo_info is None:
|
| 139 |
-
geo_info = {}
|
| 140 |
-
if messages is None:
|
| 141 |
-
messages = []
|
| 142 |
-
if subject_keywords is None:
|
| 143 |
-
subject_keywords = []
|
| 144 |
-
|
| 145 |
-
# ÉTAPE 1: Analyse du type de requête
|
| 146 |
-
query_analysis = analyze_query_type(prompt)
|
| 147 |
|
| 148 |
-
|
| 149 |
subject_shift = detect_subject_shift(prompt, current_subject, subject_keywords)
|
| 150 |
-
|
| 151 |
-
# ÉTAPE 3: Génération de la stratégie de recherche
|
| 152 |
search_strategy = generate_search_strategy(query_analysis, subject_keywords, geo_info)
|
| 153 |
|
| 154 |
-
#
|
| 155 |
-
|
| 156 |
-
location = f"{geo_info.get('city', 'Libreville')}, {geo_info.get('country', 'Gabon')}"
|
| 157 |
-
|
| 158 |
-
with st.status("🧠 Kibali Thinking Engine", expanded=True) as status:
|
| 159 |
-
st.write(f"🌍 **Localisation active :** {location}")
|
| 160 |
-
st.write("")
|
| 161 |
-
|
| 162 |
-
# Analyse du type de requête
|
| 163 |
-
st.write("### 📊 Analyse de la requête")
|
| 164 |
-
st.write(f"- **Type :** {query_analysis['type'].replace('_', ' ').title()}")
|
| 165 |
-
st.write(f"- **Complexité :** {query_analysis['complexity'].title()}")
|
| 166 |
-
|
| 167 |
-
if query_analysis['temporal']:
|
| 168 |
-
st.write("- ⏰ **Dimension temporelle détectée** → Recherche web activée")
|
| 169 |
-
if query_analysis['geographical']:
|
| 170 |
-
st.write(f"- 🗺️ **Contexte géographique :** {location}")
|
| 171 |
-
|
| 172 |
-
time.sleep(0.2)
|
| 173 |
-
st.write("")
|
| 174 |
-
|
| 175 |
-
# Détection de changement de sujet
|
| 176 |
-
st.write("### 🔄 Analyse du contexte conversationnel")
|
| 177 |
-
if subject_shift['shift_detected']:
|
| 178 |
-
if subject_shift['new_subject_detected']:
|
| 179 |
-
st.write("- 🆕 **Nouveau sujet détecté** → Rafraîchissement du contexte")
|
| 180 |
-
else:
|
| 181 |
-
st.write(f"- ⚠️ **Changement partiel** (force: {subject_shift['shift_strength']:.0%})")
|
| 182 |
-
st.write(f" *Raison : {subject_shift['reason']}*")
|
| 183 |
-
else:
|
| 184 |
-
st.write("- ✅ **Continuité du sujet actuel**")
|
| 185 |
-
if subject_keywords:
|
| 186 |
-
st.write(f" *Mots-clés actifs : {', '.join(subject_keywords[:5])}*")
|
| 187 |
-
st.write(f" *Overlap : {subject_shift['keyword_overlap']}/{len(subject_keywords)} mots-clés*")
|
| 188 |
-
|
| 189 |
-
time.sleep(0.2)
|
| 190 |
-
st.write("")
|
| 191 |
-
|
| 192 |
-
# Stratégie de recherche
|
| 193 |
-
st.write("### 🎯 Stratégie de réponse")
|
| 194 |
-
sources = []
|
| 195 |
-
if search_strategy['use_rag']:
|
| 196 |
-
sources.append(f"📚 Documents PDF (top {search_strategy['rag_k']})")
|
| 197 |
-
if search_strategy['use_memory']:
|
| 198 |
-
sources.append(f"🧠 Mémoire conversationnelle (top {search_strategy['memory_k']})")
|
| 199 |
-
if search_strategy['use_web']:
|
| 200 |
-
web_label = "🌐 Recherche web"
|
| 201 |
-
if search_strategy['web_enhanced']:
|
| 202 |
-
web_label += " (enrichie avec contexte)"
|
| 203 |
-
sources.append(web_label)
|
| 204 |
-
|
| 205 |
-
if not sources:
|
| 206 |
-
sources.append("💭 Connaissance générale du modèle")
|
| 207 |
-
|
| 208 |
-
for i, source in enumerate(sources, 1):
|
| 209 |
-
st.write(f"{i}. {source}")
|
| 210 |
-
time.sleep(0.15)
|
| 211 |
-
|
| 212 |
-
st.write("")
|
| 213 |
-
|
| 214 |
-
# Plan d'action détaillé
|
| 215 |
-
st.write("### ⚙️ Plan d'exécution")
|
| 216 |
-
steps = []
|
| 217 |
-
|
| 218 |
-
if search_strategy['use_rag']:
|
| 219 |
-
steps.append("Extraction des chunks pertinents depuis la base vectorielle PDF")
|
| 220 |
-
|
| 221 |
-
if search_strategy['use_memory']:
|
| 222 |
-
steps.append("Récupération des échanges similaires avec scoring de pertinence")
|
| 223 |
-
|
| 224 |
-
if search_strategy['use_web']:
|
| 225 |
-
query_suffix = search_strategy['search_query_suffix']
|
| 226 |
-
steps.append(f"Requête web : '{prompt[:50]}...{query_suffix}'")
|
| 227 |
-
|
| 228 |
-
steps.append("Synthèse des sources avec priorisation hiérarchique")
|
| 229 |
-
steps.append("Génération de la réponse avec verrouillage contextuel")
|
| 230 |
-
|
| 231 |
-
for i, step in enumerate(steps, 1):
|
| 232 |
-
st.write(f"{i}. {step}")
|
| 233 |
-
time.sleep(0.15)
|
| 234 |
-
|
| 235 |
-
time.sleep(0.3)
|
| 236 |
-
status.update(
|
| 237 |
-
label="✅ Stratégie validée - Génération en cours",
|
| 238 |
-
state="complete",
|
| 239 |
-
expanded=False
|
| 240 |
-
)
|
| 241 |
-
|
| 242 |
-
except Exception as e:
|
| 243 |
-
# Fallback si Streamlit n'est pas disponible
|
| 244 |
-
print(f"[Kibali Thinking] Type: {query_analysis['type']}, Complexité: {query_analysis['complexity']}")
|
| 245 |
-
print(f"[Kibali Thinking] Changement de sujet: {subject_shift['shift_detected']} (force: {subject_shift['shift_strength']:.0%})")
|
| 246 |
-
print(f"[Kibali Thinking] Sources: RAG={search_strategy['use_rag']}, Memory={search_strategy['use_memory']}, Web={search_strategy['use_web']}")
|
| 247 |
|
| 248 |
return {
|
| 249 |
"analysis": query_analysis,
|
|
|
|
|
|
|
| 1 |
import time
|
| 2 |
from typing import List, Optional
|
| 3 |
import re
|
| 4 |
+
import os
|
| 5 |
+
from dotenv import load_dotenv
|
| 6 |
+
|
| 7 |
+
load_dotenv()
|
| 8 |
|
| 9 |
def analyze_query_type(prompt: str) -> dict:
|
| 10 |
"""Analyse le type de requête pour adapter la stratégie de réflexion"""
|
|
|
|
| 20 |
"geographical": False
|
| 21 |
}
|
| 22 |
|
|
|
|
| 23 |
temporal_keywords = ["aujourd'hui", "maintenant", "récent", "actuel", "dernier", "2024", "2025"]
|
| 24 |
if any(kw in prompt_lower for kw in temporal_keywords):
|
| 25 |
analysis["temporal"] = True
|
| 26 |
analysis["needs_web"] = True
|
| 27 |
|
|
|
|
| 28 |
geo_keywords = ["gabon", "libreville", "port-gentil", "franceville", "oyem", "où", "localisation"]
|
| 29 |
if any(kw in prompt_lower for kw in geo_keywords):
|
| 30 |
analysis["geographical"] = True
|
| 31 |
|
|
|
|
| 32 |
doc_keywords = ["selon le document", "d'après le pdf", "dans le fichier", "uploadé"]
|
| 33 |
if any(kw in prompt_lower for kw in doc_keywords):
|
| 34 |
analysis["needs_docs"] = True
|
| 35 |
analysis["type"] = "document_query"
|
| 36 |
|
|
|
|
| 37 |
continuation_keywords = ["ils", "elles", "lui", "leur", "donc", "alors", "ensuite", "aussi", "également"]
|
| 38 |
if any(kw in prompt_lower for kw in continuation_keywords) or len(prompt.split()) < 5:
|
| 39 |
analysis["needs_memory"] = True
|
| 40 |
analysis["type"] = "continuation"
|
| 41 |
|
| 42 |
+
if len(prompt.split()) > 15 or (prompt.count("?") > 1):
|
|
|
|
| 43 |
analysis["complexity"] = "complex"
|
| 44 |
+
elif any(kw in prompt_lower for kw in ["pourquoi", "comment", "expliquer"]):
|
| 45 |
analysis["complexity"] = "medium"
|
| 46 |
|
|
|
|
| 47 |
web_keywords = ["actualité", "news", "prix", "cours", "météo", "horaire"]
|
| 48 |
if any(kw in prompt_lower for kw in web_keywords):
|
| 49 |
analysis["needs_web"] = True
|
|
|
|
| 54 |
def detect_subject_shift(prompt: str, current_subject: str, subject_keywords: List[str]) -> dict:
|
| 55 |
"""Détecte un changement de sujet et évalue la force du changement"""
|
| 56 |
if not current_subject or not subject_keywords:
|
| 57 |
+
return {"shift_detected": False, "shift_strength": 0.0, "new_subject_detected": True, "reason": "Init"}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
prompt_lower = prompt.lower()
|
|
|
|
|
|
|
| 60 |
prompt_words = set(re.findall(r'\b\w{4,}\b', prompt_lower))
|
| 61 |
keyword_overlap = len(prompt_words.intersection(set(subject_keywords)))
|
| 62 |
overlap_ratio = keyword_overlap / max(len(subject_keywords), 1)
|
| 63 |
|
|
|
|
| 64 |
shift_markers = ["maintenant", "sinon", "autre chose", "parlons de", "passons à", "nouveau sujet"]
|
| 65 |
has_shift_marker = any(marker in prompt_lower for marker in shift_markers)
|
| 66 |
|
|
|
|
| 67 |
shift_strength = 0.0
|
| 68 |
+
if overlap_ratio < 0.2: shift_strength += 0.5
|
| 69 |
+
if has_shift_marker: shift_strength += 0.3
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
|
| 71 |
return {
|
| 72 |
+
"shift_detected": shift_strength > 0.4,
|
| 73 |
"shift_strength": shift_strength,
|
| 74 |
"new_subject_detected": shift_strength > 0.6,
|
| 75 |
+
"reason": f"Overlap: {overlap_ratio:.1%}"
|
|
|
|
|
|
|
| 76 |
}
|
| 77 |
|
| 78 |
def generate_search_strategy(analysis: dict, subject_keywords: List[str], geo_info: dict) -> dict:
|
| 79 |
"""Génère une stratégie de recherche optimisée"""
|
| 80 |
strategy = {
|
| 81 |
"use_rag": analysis["needs_docs"],
|
| 82 |
+
"use_memory": analysis["needs_memory"],
|
| 83 |
+
"use_web": analysis["needs_web"],
|
| 84 |
+
"memory_k": 5, "rag_k": 3,
|
| 85 |
+
"web_enhanced": False, "search_query_suffix": ""
|
|
|
|
|
|
|
| 86 |
}
|
| 87 |
|
|
|
|
| 88 |
if analysis["complexity"] == "complex":
|
| 89 |
+
strategy.update({"memory_k": 8, "rag_k": 5})
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
|
|
|
|
| 91 |
if analysis["needs_web"]:
|
| 92 |
strategy["web_enhanced"] = True
|
| 93 |
+
suffix = " ".join(subject_keywords[:3]) if subject_keywords else ""
|
| 94 |
+
strategy["search_query_suffix"] = f"{suffix} {geo_info.get('city', 'Gabon')}"
|
|
|
|
|
|
|
| 95 |
|
| 96 |
return strategy
|
| 97 |
|
|
|
|
| 102 |
current_subject: Optional[str] = None,
|
| 103 |
subject_keywords: Optional[List[str]] = None
|
| 104 |
):
|
| 105 |
+
"""Phase de réflexion structurée compatible FastAPI (sans Streamlit)"""
|
| 106 |
+
geo_info = geo_info or {}
|
| 107 |
+
subject_keywords = subject_keywords or []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
+
query_analysis = analyze_query_type(prompt)
|
| 110 |
subject_shift = detect_subject_shift(prompt, current_subject, subject_keywords)
|
|
|
|
|
|
|
| 111 |
search_strategy = generate_search_strategy(query_analysis, subject_keywords, geo_info)
|
| 112 |
|
| 113 |
+
# Logs internes (visibles dans Docker)
|
| 114 |
+
print(f"🧠 [REFLECTION] Type: {query_analysis['type']} | Web: {search_strategy['use_web']}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
return {
|
| 117 |
"analysis": query_analysis,
|
tools/web.py
CHANGED
|
@@ -1,6 +1,5 @@
|
|
| 1 |
from tavily import TavilyClient
|
| 2 |
from duckduckgo_search import DDGS
|
| 3 |
-
import streamlit as st
|
| 4 |
import os
|
| 5 |
from dotenv import load_dotenv
|
| 6 |
|
|
@@ -12,9 +11,10 @@ def web_search(query: str):
|
|
| 12 |
results = []
|
| 13 |
images = []
|
| 14 |
|
| 15 |
-
# 1. TENTATIVE AVEC TAVILY
|
| 16 |
if TAVILY_API_KEY:
|
| 17 |
try:
|
|
|
|
| 18 |
tavily = TavilyClient(api_key=TAVILY_API_KEY)
|
| 19 |
res = tavily.search(
|
| 20 |
query=query,
|
|
@@ -28,12 +28,12 @@ def web_search(query: str):
|
|
| 28 |
if len(results) >= 2:
|
| 29 |
return {"results": results, "images": images, "query": query, "source": "Tavily"}
|
| 30 |
except Exception as e:
|
| 31 |
-
print(f"Tavily Error: {e}")
|
| 32 |
|
| 33 |
-
# 2. FALLBACK AVEC DUCKDUCKGO
|
|
|
|
| 34 |
try:
|
| 35 |
with DDGS() as ddgs:
|
| 36 |
-
# Texte - Utilisation d'un timeout implicite par le context manager
|
| 37 |
ddg_gen = ddgs.text(query, max_results=5)
|
| 38 |
if ddg_gen:
|
| 39 |
for r in ddg_gen:
|
|
@@ -43,16 +43,15 @@ def web_search(query: str):
|
|
| 43 |
"url": r.get('href')
|
| 44 |
})
|
| 45 |
|
| 46 |
-
# Images - Séparé pour éviter de tout bloquer en cas de 403
|
| 47 |
try:
|
| 48 |
ddg_img_gen = ddgs.images(query, max_results=3)
|
| 49 |
if ddg_img_gen:
|
| 50 |
images = [img.get('image') for img in ddg_img_gen if img.get('image')]
|
| 51 |
except Exception:
|
| 52 |
-
pass
|
| 53 |
|
| 54 |
except Exception as e:
|
| 55 |
-
print(f"DuckDuckGo Error: {e}")
|
| 56 |
|
| 57 |
return {
|
| 58 |
"results": results,
|
|
|
|
| 1 |
from tavily import TavilyClient
|
| 2 |
from duckduckgo_search import DDGS
|
|
|
|
| 3 |
import os
|
| 4 |
from dotenv import load_dotenv
|
| 5 |
|
|
|
|
| 11 |
results = []
|
| 12 |
images = []
|
| 13 |
|
| 14 |
+
# 1. TENTATIVE AVEC TAVILY
|
| 15 |
if TAVILY_API_KEY:
|
| 16 |
try:
|
| 17 |
+
print(f"🔍 Recherche Tavily pour : {query}")
|
| 18 |
tavily = TavilyClient(api_key=TAVILY_API_KEY)
|
| 19 |
res = tavily.search(
|
| 20 |
query=query,
|
|
|
|
| 28 |
if len(results) >= 2:
|
| 29 |
return {"results": results, "images": images, "query": query, "source": "Tavily"}
|
| 30 |
except Exception as e:
|
| 31 |
+
print(f"⚠️ Tavily Error: {e}")
|
| 32 |
|
| 33 |
+
# 2. FALLBACK AVEC DUCKDUCKGO
|
| 34 |
+
print(f"🦆 Fallback DuckDuckGo pour : {query}")
|
| 35 |
try:
|
| 36 |
with DDGS() as ddgs:
|
|
|
|
| 37 |
ddg_gen = ddgs.text(query, max_results=5)
|
| 38 |
if ddg_gen:
|
| 39 |
for r in ddg_gen:
|
|
|
|
| 43 |
"url": r.get('href')
|
| 44 |
})
|
| 45 |
|
|
|
|
| 46 |
try:
|
| 47 |
ddg_img_gen = ddgs.images(query, max_results=3)
|
| 48 |
if ddg_img_gen:
|
| 49 |
images = [img.get('image') for img in ddg_img_gen if img.get('image')]
|
| 50 |
except Exception:
|
| 51 |
+
pass
|
| 52 |
|
| 53 |
except Exception as e:
|
| 54 |
+
print(f"❌ DuckDuckGo Error: {e}")
|
| 55 |
|
| 56 |
return {
|
| 57 |
"results": results,
|