app.py

import os
import json
import redis
import openai
import numpy as np
import gradio as gr
from dotenv import load_dotenv
from pinecone import Pinecone, ServerlessSpec
from langchain_openai import OpenAIEmbeddings
from langchain_voyageai import VoyageAIEmbeddings
from langchain_pinecone import PineconeVectorStore
from langchain_openai import ChatOpenAI
from langchain_core.documents import Document
from langchain_core.output_parsers import StrOutputParser
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema import HumanMessage
from sentence_transformers import CrossEncoder

# Load environment variables
load_dotenv()
openai.api_key = os.getenv("OPENAI_API_KEY")
pinecone_api_key = os.getenv("PINECONE_API_KEY")
pinecone_environment = os.getenv("PINECONE_ENV")
voyage_api_key = os.getenv("VOYAGE_API_KEY")

# Initialize Pinecone
pc = Pinecone(api_key=pinecone_api_key)

# Redis caching for reranking
# redis_client = redis.Redis(host='localhost', port=6379, db=0)

# Initialize embeddings
embeddings = VoyageAIEmbeddings(voyage_api_key=voyage_api_key, model="voyage-law-2")

# Load Cross-Encoder model for reranking
reranker_model = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-12-v2")


# **1. Optimized Pinecone Index Initialization**
def initialize_pinecone_index(index_name):
    """
    Ensures the Pinecone index is optimized for fast ANN-based search.
    """
    if index_name not in pc.list_indexes():
        pc.create_index(
            name=index_name,
            dimension=1024,
            metric="cosine",
            spec=ServerlessSpec(cloud="aws", region="us-west-2"),
            hnsw_config={"ef_construction": 200, "M": 16}  # Fast ANN search
        )
    
    return PineconeVectorStore(index_name=index_name, embedding=embeddings)


# **2. Query Expansion**
QUERY_EXPANSIONS = {
    "docs": "Find all legal documents related to case law.",
    "contract": "Find contracts and legal agreements relevant to the query.",
    "policy": "Retrieve company policies and regulatory guidelines."
}

def expand_query(query):
    """
    Expands the query efficiently using predefined mappings and LLM if needed.
    """
    query = query.strip().lower()
    
    if query in QUERY_EXPANSIONS:
        return QUERY_EXPANSIONS[query]
    
    if len(query.split()) < 3:
        llm = ChatOpenAI(model="gpt-4", openai_api_key=openai.api_key, temperature=0.3)
        prompt = f"Rewrite the following vague search query into a more specific one:\nQuery: {query}.\nSpecific Query:"
        refined_query = llm([HumanMessage(content=prompt)]).content.strip()
        return refined_query if refined_query else query
    
    return query


# **3. Hybrid Search (Dense + Sparse Fusion)**
def search_documents(query, user_groups, index_name="briefmeta"):
    """
    Hybrid search combining semantic and sparse (keyword) retrieval.
    """
    try:
        pc = initialize_pinecone_index(index_name)
        vector_store = PineconeVectorStore(index_name=index_name, embedding=embeddings)

        # Dense search (Semantic embeddings)
        dense_results = vector_store.similarity_search_with_relevance_scores(
            query=query, k=10, sparse_weight=0.3,  
            filter={"groups": {"$in": user_groups}}
        )

        # Sparse search (BM25-style keyword search)
        sparse_results = vector_store.sparse_search(query=query, k=10)

        # Fusion of results
        hybrid_results = {}
        for doc, score in dense_results:
            hybrid_results[doc.metadata["id"]] = {"doc": doc, "score": score * 0.7}

        for doc, score in sparse_results:
            if doc.metadata["id"] in hybrid_results:
                hybrid_results[doc.metadata["id"]]["score"] += score * 0.3
            else:
                hybrid_results[doc.metadata["id"]] = {"doc": doc, "score": score * 0.3}

        # Sort by final score
        final_results = sorted(hybrid_results.values(), key=lambda x: x["score"], reverse=True)

        # Format output
        search_output = [
            {
                "doc_id": item["doc"].metadata.get("doc_id", "N/A"),
                "title": item["doc"].metadata.get("source", "N/A"),
                "text": item["doc"].page_content,
                "score": round(item["score"], 3)
            }
            for item in final_results
        ]

        return search_output

    except Exception as e:
        return [], f"Error in hybrid search: {str(e)}"


# **4. Reranking with Cross-Encoder (Cached)**
def rerank_results(query, search_results):
    """
    Uses a Cross-Encoder for reranking search results.
    """
    if not search_results:
        return search_results  

    cache_key = f"rerank:{query}"
    # cached_result = redis_client.get(cache_key)
    # if cached_result:
    #     return json.loads(cached_result)

    # Prepare input pairs for reranking
    pairs = [(query, doc["text"]) for doc in search_results]
    scores = reranker_model.predict(pairs)

    # Attach scores and sort
    for i, score in enumerate(scores):
        search_results[i]["rerank_score"] = round(float(score), 3)

    sorted_results = sorted(search_results, key=lambda x: x["rerank_score"], reverse=True)

    # redis_client.setex(cache_key, 600, json.dumps(sorted_results))  # Cache for 10 min
    return sorted_results


# **5. Intelligent Search Summary**
def generate_search_summary(search_results, query):
    """
    Generates an intelligent search summary.
    """
    if not search_results:
        return "No relevant documents were found for your search."

    top_docs = search_results[:3]
    doc_titles = [doc["title"] for doc in top_docs]

    summary_prompt = f"""
    Generate a **concise** 2-3 sentence summary of the search results.
    - User Query: "{query}"
    - Matching Documents: {len(search_results)} found
    - Titles: {", ".join(doc_titles)}

    **Summarize in user-friendly language.**
    """

    llm = ChatOpenAI(model="gpt-3.5-turbo", openai_api_key=openai.api_key, temperature=0.5)
    summary = llm([HumanMessage(content=summary_prompt)]).content.strip()

    return summary if summary else "No intelligent summary available."


# **6. Full RAG Workflow**
def complete_workflow(query, user_groups, index_name="briefmeta"):
    """
    Full RAG workflow: Hybrid Search -> Reranking -> Intelligent Summary
    """
    try:
        query = expand_query(query)
        raw_results = search_documents(query, user_groups, index_name)
        reranked_results = rerank_results(query, raw_results)

        document_titles = list({doc["title"] for doc in reranked_results})
        formatted_titles = " " + "\n".join(document_titles)

        intelligent_search_summary = generate_search_summary(reranked_results, query)

        results = {
            "results": reranked_results[:5],
            "total_results": len(reranked_results)
        }

        return results, formatted_titles, intelligent_search_summary
    except Exception as e:
        return {"results": [], "total_results": 0}, f"Error in workflow: {str(e)}"


# **7. Gradio UI**
def gradio_app():
    with gr.Blocks() as app:
        gr.Markdown("## 🔍 AI-Powered Document Search")

        user_query = gr.Textbox(label="Enter Your Search Query")
        user_groups = gr.Textbox(label="Enter User Groups", interactive=True)
        search_btn = gr.Button("Search")
        results_output = gr.JSON(label="Search Results")
        search_summary = gr.Textbox(label="Intelligent Search Summary")

        search_btn.click(complete_workflow, inputs=[user_query, user_groups], outputs=[results_output, search_summary])

    return app

gradio_app().launch()