edits

app.py

@@ -12,6 +12,7 @@ from flask import Flask, request, jsonify
 from flask_cors import CORS
 import threading, torch, json
 from datasets import Dataset
+import difflib
 
 # ===============================
 # 2️⃣ إعداد المتغيرات
@@ -33,7 +34,7 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 print("🔹 Using device:", device)
 
 # ===============================
-# 4️⃣ تحميل الموديل الأساسي بأمان
+# 4️⃣ تحميل الموديل الأساسي
 # ===============================
 print("🔄 Loading base model...")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
@@ -108,58 +109,129 @@ users_db = {
 @app.route("/generate", methods=["POST"])
 def generate():
     try:
-        data = request.json
-        prompt = data.get("prompt", "").strip()
+        data = request.json or {}
+        prompt = str(data.get("prompt", "")).strip()
         user_id = str(data.get("userId", "")).strip()
+        focus_level = float(data.get("focus_level", 0.5))
+        n_examples = int(data.get("n_examples", 2))
 
         user_info = users_db.get(user_id)
-        if not user_info:
-            return jsonify({"error": "User not found"}), 404
+        user_name = user_info["name"] if user_info else "مستخدم"
 
-        user_name = user_info["name"]
+        if not prompt:
+            return jsonify({"messages": [{"role": "assistant", "content": "⚠️ الرجاء ارسال سؤال أو استفسار صحيح."}]}), 400
 
+        # =====================
+        # حالة: لا يوجد داتا تدريبية
+        # =====================
         if current_dataset is None:
+            base_prompt = (
+                f"أنت مساعد قانوني ذكي اسمه زين 👨‍⚖️.\n"
+                f"مهمة زين: الإجابة على الأسئلة القانونية بأسلوب واضح ومبسط "
+                f"مع الحفاظ على الدقة القانونية.\n\n"
+                f"السؤال:\n{prompt}\n\nالإجابة:"
+            )
+
+            inputs = tokenizer(base_prompt, return_tensors="pt", truncation=True, max_length=1024)
+            inputs = {k: v.to(model.device) for k, v in inputs.items()}
+            input_len = inputs["input_ids"].shape[-1]
+
+            outputs = model.generate(
+                **inputs,
+                max_new_tokens=200,
+                temperature=0.7,
+                top_p=0.9,
+                do_sample=True,
+                pad_token_id=tokenizer.eos_token_id,
+                repetition_penalty=1.1
+            )
+
+            reply = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True).strip()
+
             return jsonify({
                 "messages": [
-                    {"role": "assistant", "content": f"⚖️ أهلاً {user_name}! أنا زين، مساعدك الذكي. برجاء طرح سؤال قانوني."}
-                ]
+                    {"role": "assistant", "content": f"أهلاً {user_name} 👋"},
+                    {"role": "assistant", "content": reply}
+                ],
+                "meta": {"mode": "base"}
             })
 
-        # Search in dataset
+        # =====================
+        # حالة: يوجد داتا تدريبية
+        # =====================
+        candidates = []
         for item in current_dataset:
-            lines = item["text"].split("\n")
-            for line in lines:
-                if line.lower().startswith("user:") and prompt in line:
-                    response_messages = []
-                    for l in lines:
-                        if ":" in l:
-                            role, content = l.split(":", 1)
-                            response_messages.append({
-                                "role": role.strip().lower(),
-                                "content": content.strip()
-                            })
-                    response_messages.insert(0, {"role": "assistant", "content": f"أهلاً {user_name} 👋"})
-                    return jsonify({"messages": response_messages})
-
-        # Fallback: generate
-        inputs = tokenizer(prompt, return_tensors="pt").to(device)
+            text = item.get("text", "")
+            ratio = difflib.SequenceMatcher(None, prompt.lower(), text.lower()).ratio()
+            candidates.append((ratio, text))
+
+        candidates.sort(reverse=True, key=lambda x: x[0])
+        top_texts = [t for r, t in candidates[:n_examples] if r > 0.05]
+
+        context_parts = []
+        for t in top_texts:
+            lines = t.split("\n")
+            u_lines, a_lines = [], []
+            for l in lines:
+                if ":" in l:
+                    role, content = l.split(":", 1)
+                    role = role.strip().lower()
+                    if role.startswith("user"):
+                        u_lines.append(content.strip())
+                    elif role.startswith("assistant") or role.startswith("ai") or role.startswith("system"):
+                        a_lines.append(content.strip())
+            if u_lines or a_lines:
+                context_parts.append("User: " + " ".join(u_lines) + "\nAssistant: " + " ".join(a_lines))
+
+        context_text = "\n\n---\n\n".join(context_parts)
+
+        instruction = (
+            "التعليمات: استعن بالأمثلة التدريبية كمرجع أساسي، "
+            "لكن أجب بأسلوبك الخاص وبكلمات مختلفة. "
+            "حافظ على صحة المعلومة القانونية ولا تنسخ النصوص كما هي."
+        )
+
+        if context_text and focus_level > 0.0:
+            emphasis = "ركز بشكل كبير على الأمثلة" if focus_level > 0.7 else "خذ الأمثلة كمرجع فقط"
+            final_prompt = (
+                f"{instruction}\n\n{emphasis}\n\n"
+                f"أمثلة تدريبية (مستوى التركيز={focus_level}):\n{context_text}\n\n"
+                f"السؤال:\n{prompt}\n\nالإجابة:"
+            )
+        else:
+            final_prompt = f"{instruction}\n\nالسؤال:\n{prompt}\n\nالإجابة:"
+
+        # ==== توليد
+        inputs = tokenizer(final_prompt, return_tensors="pt", truncation=True, max_length=1024)
+        inputs = {k: v.to(model.device) for k, v in inputs.items()}
+        input_len = inputs["input_ids"].shape[-1]
+
         outputs = model.generate(
             **inputs,
-            max_new_tokens=150,
-            temperature=0.7,
-            do_sample=True
+            max_new_tokens=200,
+            temperature=0.6,
+            top_p=0.9,
+            do_sample=True,
+            pad_token_id=tokenizer.eos_token_id,
+            repetition_penalty=1.1
         )
-        reply = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+        reply = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True).strip()
 
         return jsonify({
             "messages": [
                 {"role": "assistant", "content": f"أهلاً {user_name} 👋"},
                 {"role": "assistant", "content": reply}
-            ]
+            ],
+            "meta": {
+                "mode": "trained",
+                "used_examples": len(context_parts),
+                "focus_level": focus_level
+            }
         })
 
     except Exception as e:
-        return jsonify({"messages": [{"role": "assistant", "content": f"خطأ: {str(e)}"}]})
+        return jsonify({"messages": [{"role": "assistant", "content": f"❌ خطأ: {str(e)}"}]})
 
 # ===============================
 # 7b️⃣ تحميل الداتا
@@ -172,7 +244,6 @@ def load_dataset():
                 data = json.load(f)
             except:
                 data = []
-
         if data:
             formatted = []
             for item in data:
@@ -203,7 +274,6 @@ def train():
         epochs = 1
         batch_size = 1
 
-        # 1️⃣ JSON مباشر
         if request.is_json:
             body = request.get_json()
             if isinstance(body, list):
@@ -215,8 +285,6 @@ def train():
 
             epochs = int(body.get("epochs", 1)) if isinstance(body, dict) else 1
             batch_size = int(body.get("batch_size", 1)) if isinstance(body, dict) else 1
-
-        # 2️⃣ form-data ملف
         else:
             file = request.files.get("file")
             if file:
@@ -226,7 +294,6 @@ def train():
             else:
                 return jsonify({"error": "No file uploaded or JSON provided"}), 400
 
-        # دمج وإزالة التكرارات حسب محتوى الرسائل
         old_data = []
         if os.path.exists(DATASET_FILE):
             with open(DATASET_FILE, "r", encoding="utf-8") as f:
@@ -248,11 +315,9 @@ def train():
                 seen_texts.add(text_key)
                 unique_data.append(item)
 
-        # الحفظ
         with open(DATASET_FILE, "w", encoding="utf-8") as f:
             json.dump(unique_data, f, ensure_ascii=False, indent=2)
 
-        # تجهيز Dataset للتدريب
         formatted = []
         for item in unique_data:
             if "messages" in item:
@@ -262,7 +327,6 @@ def train():
                 formatted.append({"text": conversation.strip()})
             elif "text" in item:
                 formatted.append({"text": item["text"]})
-
         current_dataset = Dataset.from_list(formatted)
 
         def tokenize_function(examples):
@@ -282,7 +346,6 @@ def train():
 
         trainer = Trainer(model=model, args=training_args, train_dataset=tokenized_dataset)
 
-        # تشغيل التدريب في Thread منفصل
         def run_training():
             print("🔄 Training started...")
             trainer.train()
@@ -301,7 +364,6 @@ def train():
     except Exception as e:
         return jsonify({"error": str(e)})
 
-
 # ===============================
 # 8️⃣ تشغيل Flask
 # ===============================