main_easy.py

import os
import time
from dotenv import load_dotenv
from config_loader import cfg 

# Optimized imports - only what we need for Retrieval and Generation
from data.vector_db import get_index_by_name, load_chunks_from_pinecone # Using the new helper
from retriever.retriever import HybridRetriever
from retriever.generator import RAGGenerator
from retriever.processor import ChunkProcessor
from retriever.evaluator import RAGEvaluator

# Model Fleet
from models.llama_3_8b import Llama3_8B
from models.mistral_7b import Mistral_7b
from models.qwen_2_5 import Qwen2_5
from models.deepseek_v3 import DeepSeek_V3
from models.tiny_aya import TinyAya

MODEL_MAP = {
    "Llama-3-8B": Llama3_8B,
    "Mistral-7B": Mistral_7b,
    "Qwen-2.5": Qwen2_5,
    "DeepSeek-V3": DeepSeek_V3,
    "TinyAya": TinyAya
}

load_dotenv()

def main():
    hf_token = os.getenv("HF_TOKEN")
    pinecone_key = os.getenv("PINECONE_API_KEY")
    query = "How do transformers handle long sequences?"

    # 1. Connect to Existing Index (No creation, no uploading)
    # We use the slugified name directly or via config
    index_name = f"{cfg.db['base_index_name']}-{cfg.processing['technique']}"
    index = get_index_by_name(pinecone_key, index_name)

    # 2. Setup Processor (Required for the Encoder/Embedding model)
    proc = ChunkProcessor(model_name=cfg.processing['embedding_model'])

    # 3. Load BM25 Corpus (The "Source of Truth")
    # This replaces the entire data_loader/chunking block
    # Note: On first run, this hits Pinecone. Use a pickle cache here for 0s delay.
    print("🔄 Loading BM25 context from Pinecone metadata...")
    final_chunks = load_chunks_from_pinecone(index) 

    # 4. Retrieval Setup
    retriever = HybridRetriever(final_chunks, proc.encoder)
    
    print(f"🔎 Searching via {cfg.retrieval['mode']} mode...")
    context_chunks = retriever.search(
        query, index,
        mode=cfg.retrieval['mode'],
        rerank_strategy=cfg.retrieval['rerank_strategy'],
        use_mmr=cfg.retrieval['use_mmr'],
        top_k=cfg.retrieval['top_k'],
        final_k=cfg.retrieval['final_k']
    )

    # 5. Initialization of Contestants
    rag_engine = RAGGenerator()
    models = {name: MODEL_MAP[name](token=hf_token) for name in cfg.model_list}
    
    evaluator = RAGEvaluator(
        judge_model=cfg.gen['judge_model'],
        embedding_model=proc.encoder,
        api_key=os.getenv("GROQ_API_KEY")
    )

    tournament_results = {}

    # 6. Tournament Loop
    for name, model_inst in models.items():
        print(f"\n🏆 Tournament: {name} is generating...")
        try:
            # Generation
            answer = rag_engine.get_answer(
                model_inst, query, context_chunks, 
                temperature=cfg.gen['temperature']
            )
            
            # Faithfulness Evaluation
            faith = evaluator.evaluate_faithfulness(answer, context_chunks)
            # Relevancy Evaluation
            rel = evaluator.evaluate_relevancy(query, answer)

            tournament_results[name] = {
                "Answer": answer[:100] + "...", # Preview
                "Faithfulness": faith['score'],
                "Relevancy": rel['score']
            }
            print(f"✅ {name} Score - Faith: {faith['score']} | Rel: {rel['score']}")
            
        except Exception as e:
            print(f"❌ Error evaluating {name}: {e}")

    print("\n--- Final Tournament Standings ---")
    for name, scores in tournament_results.items():
        print(f"{name}: F={scores['Faithfulness']}, R={scores['Relevancy']}")

if __name__ == "__main__":
    main()