Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -1,58 +1,55 @@
|
|
1 |
-
import dspy
|
|
|
2 |
import chromadb
|
3 |
import fitz # PyMuPDF
|
4 |
from sentence_transformers import SentenceTransformer
|
5 |
import json
|
6 |
from dspy import Example, MIPROv2, Evaluate, evaluate
|
7 |
|
8 |
-
# إعداد نموذج
|
9 |
dspy.settings.configure(lm=dspy.LM("mistralai/Mistral-7B-Instruct-v0.2"))
|
10 |
|
11 |
-
# إعداد Chroma
|
12 |
client = chromadb.PersistentClient(path="./chroma_db")
|
13 |
col = client.get_or_create_collection(name="arabic_docs")
|
14 |
|
15 |
-
# نموذج
|
16 |
embedder = SentenceTransformer("sentence-transformers/LaBSE")
|
17 |
|
18 |
# تقطيع النصوص من PDF
|
19 |
def process_pdf(pdf_bytes):
|
20 |
doc = fitz.open(stream=pdf_bytes, filetype="pdf")
|
21 |
texts = []
|
22 |
-
for
|
23 |
-
text =
|
24 |
for chunk in text.split("\n\n"):
|
25 |
if len(chunk.strip()) > 50:
|
26 |
texts.append(chunk.strip())
|
27 |
return texts
|
28 |
|
29 |
-
# إدخال
|
30 |
def ingest(pdf_file):
|
31 |
-
|
32 |
-
pdf_bytes = pdf_file.read()
|
33 |
-
else:
|
34 |
-
with open(pdf_file.name, "rb") as f:
|
35 |
-
pdf_bytes = f.read()
|
36 |
texts = process_pdf(pdf_bytes)
|
37 |
embeddings = embedder.encode(texts, show_progress_bar=True)
|
38 |
for i, (chunk, emb) in enumerate(zip(texts, embeddings)):
|
39 |
col.add(ids=[f"chunk_{i}"], embeddings=[emb.tolist()], metadatas=[{"text": chunk}])
|
40 |
return f"✅ تمت إضافة {len(texts)} مقطعاً."
|
41 |
|
42 |
-
# استرجاع السياق
|
43 |
def retrieve_context(question):
|
44 |
-
|
45 |
-
results = col.query(query_embeddings=[
|
46 |
-
context_list = [m["text"] for m in results["metadatas"][0]]
|
47 |
return "\n\n".join(context_list)
|
48 |
|
49 |
-
# توقيع RAG
|
50 |
class RagSig(dspy.Signature):
|
51 |
question: str = dspy.InputField()
|
52 |
context: str = dspy.InputField()
|
53 |
answer: str = dspy.OutputField()
|
54 |
|
55 |
-
# وحدة
|
56 |
class RagMod(dspy.Module):
|
57 |
def __init__(self):
|
58 |
super().__init__()
|
@@ -64,17 +61,17 @@ class RagMod(dspy.Module):
|
|
64 |
|
65 |
model = RagMod()
|
66 |
|
67 |
-
# توليد
|
68 |
def answer(question):
|
69 |
out = model(question)
|
70 |
return out.answer
|
71 |
|
72 |
-
# تحميل بيانات
|
73 |
def load_dataset(path):
|
74 |
with open(path, "r", encoding="utf-8") as f:
|
75 |
return [Example(**json.loads(l)).with_inputs("question") for l in f]
|
76 |
|
77 |
-
# تحسين النموذج
|
78 |
def optimize(train_file, val_file):
|
79 |
global model
|
80 |
trainset = load_dataset(train_file.name)
|
@@ -86,10 +83,10 @@ def optimize(train_file, val_file):
|
|
86 |
|
87 |
# واجهة Gradio
|
88 |
with gr.Blocks() as demo:
|
89 |
-
gr.Markdown("## 🧠 نظام RAG عربي باستخدام DSPy +
|
90 |
|
91 |
with gr.Tab("📥 تحميل وتخزين"):
|
92 |
-
pdf_input = gr.File(label="ارفع ملف PDF", type="
|
93 |
ingest_btn = gr.Button("إضافة إلى قاعدة البيانات")
|
94 |
ingest_out = gr.Textbox(label="نتيجة الإضافة")
|
95 |
ingest_btn.click(ingest, inputs=pdf_input, outputs=ingest_out)
|
|
|
1 |
+
import dspy
|
2 |
+
import gradio as gr
|
3 |
import chromadb
|
4 |
import fitz # PyMuPDF
|
5 |
from sentence_transformers import SentenceTransformer
|
6 |
import json
|
7 |
from dspy import Example, MIPROv2, Evaluate, evaluate
|
8 |
|
9 |
+
# إعداد نموذج اللغة
|
10 |
dspy.settings.configure(lm=dspy.LM("mistralai/Mistral-7B-Instruct-v0.2"))
|
11 |
|
12 |
+
# إعداد قاعدة Chroma
|
13 |
client = chromadb.PersistentClient(path="./chroma_db")
|
14 |
col = client.get_or_create_collection(name="arabic_docs")
|
15 |
|
16 |
+
# نموذج embeddings يدعم العربية
|
17 |
embedder = SentenceTransformer("sentence-transformers/LaBSE")
|
18 |
|
19 |
# تقطيع النصوص من PDF
|
20 |
def process_pdf(pdf_bytes):
|
21 |
doc = fitz.open(stream=pdf_bytes, filetype="pdf")
|
22 |
texts = []
|
23 |
+
for page in doc:
|
24 |
+
text = page.get_text()
|
25 |
for chunk in text.split("\n\n"):
|
26 |
if len(chunk.strip()) > 50:
|
27 |
texts.append(chunk.strip())
|
28 |
return texts
|
29 |
|
30 |
+
# إدخال البيانات في Chroma
|
31 |
def ingest(pdf_file):
|
32 |
+
pdf_bytes = pdf_file # لأننا استخدمنا type='binary'
|
|
|
|
|
|
|
|
|
33 |
texts = process_pdf(pdf_bytes)
|
34 |
embeddings = embedder.encode(texts, show_progress_bar=True)
|
35 |
for i, (chunk, emb) in enumerate(zip(texts, embeddings)):
|
36 |
col.add(ids=[f"chunk_{i}"], embeddings=[emb.tolist()], metadatas=[{"text": chunk}])
|
37 |
return f"✅ تمت إضافة {len(texts)} مقطعاً."
|
38 |
|
39 |
+
# استرجاع السياق
|
40 |
def retrieve_context(question):
|
41 |
+
embedding = embedder.encode([question])[0]
|
42 |
+
results = col.query(query_embeddings=[embedding.tolist()], n_results=3)
|
43 |
+
context_list = [m["text"] for m in results["metadatas"][0]]
|
44 |
return "\n\n".join(context_list)
|
45 |
|
46 |
+
# توقيع وحدة RAG
|
47 |
class RagSig(dspy.Signature):
|
48 |
question: str = dspy.InputField()
|
49 |
context: str = dspy.InputField()
|
50 |
answer: str = dspy.OutputField()
|
51 |
|
52 |
+
# وحدة RAG
|
53 |
class RagMod(dspy.Module):
|
54 |
def __init__(self):
|
55 |
super().__init__()
|
|
|
61 |
|
62 |
model = RagMod()
|
63 |
|
64 |
+
# توليد الإجابة
|
65 |
def answer(question):
|
66 |
out = model(question)
|
67 |
return out.answer
|
68 |
|
69 |
+
# تحميل مجموعة بيانات التدريب/التحقق
|
70 |
def load_dataset(path):
|
71 |
with open(path, "r", encoding="utf-8") as f:
|
72 |
return [Example(**json.loads(l)).with_inputs("question") for l in f]
|
73 |
|
74 |
+
# تحسين النموذج
|
75 |
def optimize(train_file, val_file):
|
76 |
global model
|
77 |
trainset = load_dataset(train_file.name)
|
|
|
83 |
|
84 |
# واجهة Gradio
|
85 |
with gr.Blocks() as demo:
|
86 |
+
gr.Markdown("## 🧠 نظام RAG عربي باستخدام DSPy + ChromaDB + Mistral")
|
87 |
|
88 |
with gr.Tab("📥 تحميل وتخزين"):
|
89 |
+
pdf_input = gr.File(label="ارفع ملف PDF", type="binary") # ← هنا التعديل
|
90 |
ingest_btn = gr.Button("إضافة إلى قاعدة البيانات")
|
91 |
ingest_out = gr.Textbox(label="نتيجة الإضافة")
|
92 |
ingest_btn.click(ingest, inputs=pdf_input, outputs=ingest_out)
|