Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import sys
|
| 3 |
+
import pandas as pd
|
| 4 |
+
import numpy as np
|
| 5 |
+
import faiss
|
| 6 |
+
import gradio as gr
|
| 7 |
+
from sentence_transformers import SentenceTransformer
|
| 8 |
+
from huggingface_hub import InferenceClient
|
| 9 |
+
from datasets import load_dataset
|
| 10 |
+
import json
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
DATASET_REPO = "LCA/HACKATHON_PARTS"
|
| 14 |
+
|
| 15 |
+
dataset = load_dataset(DATASET_REPO, split="train")
|
| 16 |
+
df = dataset.to_pandas()
|
| 17 |
+
|
| 18 |
+
descriptions = df['DESIGNATION'].tolist()
|
| 19 |
+
codes = df["CODE"].astype(str).tolist()
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
# --- Embedding model ---
|
| 23 |
+
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
|
| 24 |
+
|
| 25 |
+
#--- Load or compute embeddings + FAISS index ---
|
| 26 |
+
#For start, test perf without caching this
|
| 27 |
+
if os.path.exists("embeddings.npy") and os.path.exists("faiss.index"):
|
| 28 |
+
embeddings = np.load("embeddings.npy")
|
| 29 |
+
index = faiss.read_index("faiss.index")
|
| 30 |
+
else:
|
| 31 |
+
embeddings = embedding_model.encode(descriptions, convert_to_numpy=True)
|
| 32 |
+
faiss.normalize_L2(embeddings)
|
| 33 |
+
index = faiss.IndexFlatIP(embeddings.shape[1])
|
| 34 |
+
index.add(embeddings)
|
| 35 |
+
# Save embeddings and index for future use
|
| 36 |
+
np.save("embeddings.npy", embeddings)
|
| 37 |
+
faiss.write_index(index, "faiss.index")
|
| 38 |
+
|
| 39 |
+
# --- Inference API client ---
|
| 40 |
+
# client = InferenceClient("HuggingFaceH4/zephyr-7b-beta", token=os.getenv("HF_TOKEN"))
|
| 41 |
+
|
| 42 |
+
def rechercher_article(articleSource):
|
| 43 |
+
|
| 44 |
+
source = articleSource["designation"]
|
| 45 |
+
query_embedding = embedding_model.encode([source], convert_to_numpy=True)
|
| 46 |
+
faiss.normalize_L2(query_embedding)
|
| 47 |
+
# Recherche du/des voisin(s) le(s) plus proche(s)
|
| 48 |
+
similarity_scores, indices = index.search(query_embedding, k=1)
|
| 49 |
+
# Gérer la qualité du retour avec un seuil de similarité
|
| 50 |
+
threshold = 0.7 # à ajuster selon vos tests
|
| 51 |
+
if similarity_scores[0][0] < threshold:
|
| 52 |
+
print(f"Score de similarité trop faible ({similarity_scores[0][0]:.2f}) pour '{source}'")
|
| 53 |
+
return "UNKNOWN"
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
article = {}
|
| 57 |
+
article["code"] = codes[indices[0][0]]
|
| 58 |
+
article["designation"] = descriptions[indices[0][0]]
|
| 59 |
+
article["source"] = source
|
| 60 |
+
article["quantite"] = articleSource.get("quantite", None)
|
| 61 |
+
print(f"Code trouvé pour '{source}': {article['code']} / {article['designation']}")
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
return article
|
| 65 |
+
|
| 66 |
+
def respond(message):
|
| 67 |
+
|
| 68 |
+
# Prompt par défaut
|
| 69 |
+
custom_prompt = """Tu es un analyseur de messages expert.
|
| 70 |
+
Ta mission est de déterminer dans le messages fourni quels sont les articles qui sont demandés et pour quelle quantité.
|
| 71 |
+
La réponse est au format json et donne 2 informations par article identifié : la désignation et le nombre
|
| 72 |
+
La désignation est formé du type d'article et des caractéristiques comme la matière ou les dimensions
|
| 73 |
+
|
| 74 |
+
Ne retourne que le JSON.
|
| 75 |
+
|
| 76 |
+
"""
|
| 77 |
+
# query_embedding = embedding_model.encode([message], convert_to_numpy=True)
|
| 78 |
+
# faiss.normalize_L2(query_embedding)
|
| 79 |
+
# _, indices = index.search(query_embedding, k=5)
|
| 80 |
+
# context = "\n".join([f"{codes[i]}: {descriptions[i]}" for i in indices[0]])
|
| 81 |
+
|
| 82 |
+
# Utilise le prompt personnalisé
|
| 83 |
+
|
| 84 |
+
|
| 85 |
+
# message = custom_prompt.format(message=message)
|
| 86 |
+
|
| 87 |
+
messages = [{"role": "system", "content": custom_prompt}]
|
| 88 |
+
messages += [{"role": "user", "content": message}]
|
| 89 |
+
|
| 90 |
+
# full_response = client.text_generation(message)
|
| 91 |
+
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta", token=os.getenv("HF_TOKEN"))
|
| 92 |
+
# client = InferenceClient(
|
| 93 |
+
# "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
|
| 94 |
+
# token=os.getenv("HF_TOKEN"),
|
| 95 |
+
# #provider="auto" # or choose a supported provider from the error message
|
| 96 |
+
# )
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
full_response = ""
|
| 100 |
+
for chunk in client.chat_completion(
|
| 101 |
+
messages,
|
| 102 |
+
max_tokens=512,
|
| 103 |
+
stream=True,
|
| 104 |
+
temperature=0.1,
|
| 105 |
+
top_p=0.8,
|
| 106 |
+
):
|
| 107 |
+
token = chunk.choices[0].delta.content
|
| 108 |
+
if token:
|
| 109 |
+
full_response += token
|
| 110 |
+
# yield full_response.replace("\n", "\n\n")
|
| 111 |
+
|
| 112 |
+
# If you expect a JSON response, you can try to parse it here
|
| 113 |
+
# import json
|
| 114 |
+
# try:
|
| 115 |
+
order = {}
|
| 116 |
+
try:
|
| 117 |
+
data = json.loads(full_response)
|
| 118 |
+
articles = []
|
| 119 |
+
for article in data.get("articles", []):
|
| 120 |
+
found_article = rechercher_article(article)
|
| 121 |
+
if found_article != "UNKNOWN":
|
| 122 |
+
articles.append(found_article)
|
| 123 |
+
order["articles"] = articles
|
| 124 |
+
# Ajouter les champs destinataire et delai avec des valeurs figées
|
| 125 |
+
order["destinataire"] = {
|
| 126 |
+
"societe": "Société Exemple",
|
| 127 |
+
"nom": "Dupont",
|
| 128 |
+
"prenom": "Jean",
|
| 129 |
+
"email": "jean.dupont@exemple.com"
|
| 130 |
+
}
|
| 131 |
+
order["delai"] = "2024-07-15"
|
| 132 |
+
except Exception as e:
|
| 133 |
+
print("Could not parse articles:", e)
|
| 134 |
+
order = {}
|
| 135 |
+
|
| 136 |
+
return order
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
with gr.Blocks() as demo:
|
| 140 |
+
gr.Markdown("# Part identification Assistant")
|
| 141 |
+
#prompt_box = gr.Textbox(label="Prompt système", value=DEFAULT_PROMPT, lines=8)
|
| 142 |
+
#temperature_slider = gr.Slider(label="Température", minimum=0.0, maximum=1.0, value=0.1, step=0.01)
|
| 143 |
+
#top_p_slider = gr.Slider(label="Top-p", minimum=0.0, maximum=1.0, value=0.8, step=0.01)
|
| 144 |
+
message_box = gr.Textbox(label="Votre question")
|
| 145 |
+
response_box = gr.Textbox(label="Réponse de l'assistant", interactive=False, lines=30)
|
| 146 |
+
send_btn = gr.Button("Envoyer")
|
| 147 |
+
|
| 148 |
+
def chat(message):
|
| 149 |
+
history = [] # ou récupère l'historique si tu veux le gérer
|
| 150 |
+
gen = respond(message)
|
| 151 |
+
# full_response = ""
|
| 152 |
+
# for response in gen:
|
| 153 |
+
# full_response = full_response + response
|
| 154 |
+
# On renvoie la dernière réponse et le contexte utilisé
|
| 155 |
+
# Il faut recalculer le contexte ici pour l'afficher
|
| 156 |
+
# query_embedding = embedding_model.encode([message], convert_to_numpy=True)
|
| 157 |
+
# faiss.normalize_L2(query_embedding)
|
| 158 |
+
# _, indices = index.search(query_embedding, k=5)
|
| 159 |
+
# context = "\n".join([f"{codes[i]}: {descriptions[i]}" for i in indices[0]])
|
| 160 |
+
return json.dumps(gen, indent=2, ensure_ascii=False)
|
| 161 |
+
|
| 162 |
+
send_btn.click(
|
| 163 |
+
chat,
|
| 164 |
+
inputs=[message_box],
|
| 165 |
+
outputs=[response_box]
|
| 166 |
+
)
|
| 167 |
+
|
| 168 |
+
if __name__ == "__main__":
|
| 169 |
+
demo.launch(share=True)
|
| 170 |
+
|