rag_pipeline.py

from sentence_transformers import CrossEncoder, SentenceTransformer
from transformers import AutoTokenizer, AutoModelForCausalLM
import faiss
import numpy as np
from typing import List, Dict

class ArabicRAGSystem:
    def __init__(self):
        """Initialize with fallback models for Arabic support"""
        # Solution 1: Use reliable Arabic embedding model
        self.embedding_model = SentenceTransformer("UBC-NLP/AraBERT")
        
        # Solution 2: Fallback cross-encoder options
        try:
            self.cross_encoder = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2")  # Multilingual fallback
        except:
            self.cross_encoder = None  # System will work without it
            
        # Solution 3: Main Arabic LLM with error handling
        try:
            self.tokenizer = AutoTokenizer.from_pretrained("inception-mbzuai/jais-13b-chat")
            self.llm = AutoModelForCausalLM.from_pretrained("inception-mbzuai/jais-13b-chat")
        except:
            # Fallback to smaller Arabic model
            self.tokenizer = AutoTokenizer.from_pretrained("aubmindlab/aragpt2-base")
            self.llm = AutoModelForCausalLM.from_pretrained("aubmindlab/aragpt2-base")
            
        self.index = faiss.IndexFlatL2(768)  # AraBERT uses 768-dim embeddings

    def generate_answer(self, question: str, documents: List[Dict], 
                      top_k: int = 5, temperature: float = 0.7) -> tuple:
        """Enhanced with fallback retrieval methods"""
        # Index documents
        texts = [doc["text"] for doc in documents]
        self.index.add(np.array(self.embedding_model.encode(texts)))
        
        # Two-phase retrieval with fallback
        query_embedding = self.embedding_model.encode([question])
        distances, indices = self.index.search(query_embedding, top_k*2)
        
        # Solution 4: Cross-encoder fallback logic
        if self.cross_encoder:
            pairs = [[question, documents[idx]["text"]] for idx in indices[0]]
            scores = self.cross_encoder.predict(pairs)
            top_indices = [indices[0][i] for i in np.argsort(scores)[-top_k:][::-1]]
        else:
            top_indices = indices[0][:top_k]
        
        # Prepare context with metadata
        context = "\n\n".join([
            f"المرجع: {documents[idx]['source']}\n"
            f"الصفحة: {documents[idx].get('page', 'N/A')}\n"
            f"النص: {documents[idx]['text']}\n"
            for idx in top_indices
        ])
        
        # Generation with error handling
        prompt = f"""
        نظام التحليل الديني العربي:
        السياق:
        {context}
        
        السؤال: {question}
        
        التعليمات:
        - أجب باللغة العربية الفصحى فقط
        - استخدم المعلومات من السياق فقط
        - أشر إلى المصادر باستخدام [المرجع: اسم الملف، الصفحة]
        - إذا لم تجد إجابة واضحة قل "لا تتوفر معلومات كافية"
        
        الإجابة:
        """.strip()
        
        try:
            inputs = self.tokenizer(prompt, return_tensors="pt")
            outputs = self.llm.generate(
                inputs.input_ids,
                max_new_tokens=512,
                temperature=temperature,
                do_sample=True,
                pad_token_id=self.tokenizer.eos_token_id
            )
            answer = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
            answer = answer.split("الإجابة:")[-1].strip()
        except:
            answer = "عذراً، حدث خطأ في معالجة السؤال. يرجى المحاولة مرة أخرى."
        
        # Prepare sources
        sources = [{
            "text": documents[idx]["text"],
            "source": documents[idx]["source"],
            "page": documents[idx].get("page", "N/A"),
            "score": float(1 - distances[0][i]) if self.cross_encoder else 0.0
        } for i, idx in enumerate(top_indices)]
        
        return answer, sources