systex_phase2/faiss_index.py

import faiss
import numpy as np
import json
from time import time
import asyncio
from datasets import Dataset
from typing import List
from dotenv import load_dotenv
import os
import pickle
from supabase.client import Client, create_client
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
import pandas as pd
from langchain_core.documents import Document
from langchain.load import dumps, loads
from langchain_community.chat_models import ChatOllama
from langchain.callbacks import get_openai_callback

# Import from the parent directory
import sys

from RAG_strategy import multi_query_chain
sys.path.append('..')
# from RAG_strategy_Taide import taide_llm, system_prompt, multi_query
system_prompt: str = "你是一個來自台灣的AI助理，你的名字是 TAIDE，樂於以台灣人的立場幫助使用者，會用繁體中文回答問題。"
# llm = ChatOpenAI(model_name="gpt-4o-mini", temperature=0)


from langchain.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from file_loader.add_vectordb import GetVectorStore
# from local_llm import ollama_, hf
# # from local_llm import ollama_, taide_llm, hf
# llm = hf()
# llm = taide_llm


# Import RAGAS metrics
from ragas import evaluate
from ragas.metrics import answer_relevancy, faithfulness, context_recall, context_precision

# Load environment variables
load_dotenv('../../.env')
supabase_url = os.getenv("SUPABASE_URL")
supabase_key = os.getenv("SUPABASE_KEY")
openai_api_key = os.getenv("OPENAI_API_KEY")
document_table = "documents2"

# Initialize Supabase client
supabase: Client = create_client(supabase_url, supabase_key)

# Initialize embeddings and chat model
embeddings = OpenAIEmbeddings(openai_api_key=openai_api_key)

def download_embeddings():
    response = supabase.table(document_table).select("id, embedding, metadata, content").execute()
    # response = supabase.table(document_table).select("id, embedding, metadata, content").eq('metadata ->> source', 'supplement.docx').execute()
    embeddings = []
    ids = []
    metadatas = []
    contents = []
    for item in response.data:
        embedding = json.loads(item['embedding'])
        embeddings.append(embedding)
        ids.append(item['id'])
        metadatas.append(item['metadata'])
        contents.append(item['content'])
    return np.array(embeddings, dtype=np.float32), ids, metadatas, contents

def create_faiss_index(embeddings):
    dimension = embeddings.shape[1]
    index = faiss.IndexFlatIP(dimension)  # Use Inner Product for cosine similarity
    faiss.normalize_L2(embeddings)  # Normalize embeddings for cosine similarity
    index.add(embeddings)
    return index

def save_faiss_index(index, file_path):
    faiss.write_index(index, file_path)
    print(f"FAISS index saved to {file_path}")

def load_faiss_index(file_path):
    if os.path.exists(file_path):
        index = faiss.read_index(file_path)
        print(f"FAISS index loaded from {file_path}")
        return index
    return None

def save_metadata(ids, metadatas, contents, file_path):
    with open(file_path, 'wb') as f:
        pickle.dump((ids, metadatas, contents), f)
    print(f"Metadata saved to {file_path}")

def load_metadata(file_path):
    if os.path.exists(file_path):
        with open(file_path, 'rb') as f:
            ids, metadatas, contents = pickle.load(f)
        print(f"Metadata loaded from {file_path}")
        return ids, metadatas, contents
    return None, None, None

def search_faiss(index, query_vector, k=4):
    query_vector = np.array(query_vector, dtype=np.float32).reshape(1, -1)
    faiss.normalize_L2(query_vector)
    distances, indices = index.search(query_vector, k)
    return distances[0], indices[0]

class FAISSRetriever:
    def __init__(self, index, ids, metadatas, contents, embeddings_model):
        self.index = index
        self.ids = ids
        self.metadatas = metadatas
        self.contents = contents
        self.embeddings_model = embeddings_model

    def get_relevant_documents(self, query: str, k: int = 4) -> List[Document]:
        query_vector = self.embeddings_model.embed_query(query)
        _, indices = search_faiss(self.index, query_vector, k=k)
        return [
            Document(page_content=self.contents[i], metadata=self.metadatas[i])
            for i in indices
        ]
    def map(self, query_list: List[list]) -> List[Document]:
        def get_unique_union(documents: List[list]):
            """ Unique union of retrieved docs """
            # Flatten list of lists, and convert each Document to string
            flattened_docs = [dumps(doc) for sublist in documents for doc in sublist]
            # Get unique documents
            unique_docs = list(set(flattened_docs))
            # Return
            return [loads(doc) for doc in unique_docs]
        
        documents = []
        for query in query_list:
            if query != "":
                docs = self.get_relevant_documents(query)

            documents.extend(docs)

        return get_unique_union(documents)

def load_qa_pairs():
    # df = pd.read_csv("../QA_database_rows.csv")
    response = supabase.table('QA_database').select("Question, Answer").execute()
    df = pd.DataFrame(response.data)

    return df['Question'].tolist(), df['Answer'].tolist()

def faiss_multiquery(question: str, retriever: FAISSRetriever, llm, k: int = 4):
    generate_queries = multi_query_chain(llm)

    questions = generate_queries.invoke(question)
    questions = [item for item in questions if item != ""]
    questions.append(question)
    for q in questions:
        print(q)

    # docs = list(map(retriever.get_relevant_documents, questions))
    docs = list(map(lambda query: retriever.get_relevant_documents(query, k=k), questions))
    docs = [item for sublist in docs for item in sublist]

    return docs

def faiss_query(retriever, question: str, llm, k: int = 4, multi_query: bool = False) -> str:
    if multi_query:
        docs = faiss_multiquery(question, retriever, llm, k)
        # print(docs)
    else:
        docs = retriever.get_relevant_documents(question, k)
        # print(docs)
    context = docs
    
    system_prompt: str = "你是一個來自台灣的AI助理，樂於以台灣人的立場幫助使用者，會用繁體中文回答問題。"
    template = """
    <|begin_of_text|>
    
    <|start_header_id|>system<|end_header_id|>
    你是一個來自台灣的ESG的AI助理，請用繁體中文回答問題 \n
    You should not mention anything about "根據提供的文件內容" or other similar terms.
    請盡可能的詳細回答問題。
    如果你不知道答案請回答："很抱歉，目前我無法回答您的問題，請將您的詢問發送至 test@systex.com 以便獲得更進一步的幫助，謝謝。"
    勿回答無關資訊或任何與某特定公司相關的問題
    <|eot_id|>
    
    <|start_header_id|>user<|end_header_id|>
    Answer the following question based on this context:

    {context}

    Question: {question}
    用繁體中文回答問題，請用一段話詳細的回答。勿回答無關資訊或任何與某特定公司相關的問題。
    如果你不知道答案請回答："很抱歉，目前我無法回答您的問題，請將您的詢問發送至 test@systex.com 以便獲得更進一步的幫助，謝謝。"
    
    <|eot_id|>
    
    <|start_header_id|>assistant<|end_header_id|>
    """
    prompt = ChatPromptTemplate.from_template(
        system_prompt + "\n\n" +
        template
    )

    rag_chain = prompt | llm | StrOutputParser()
    return context, rag_chain.invoke({"context": context, "question": question})


def create_faiss_retriever():
    faiss_index_path = "faiss_index.bin"
    metadata_path = "faiss_metadata.pkl"

    index = load_faiss_index(faiss_index_path)
    ids, metadatas, contents = load_metadata(metadata_path)

    if index is None or ids is None:
        print("FAISS index or metadata not found. Creating new index...")
        print("Downloading embeddings from Supabase...")
        embeddings_array, ids, metadatas, contents = download_embeddings()

        print("Creating FAISS index...")
        index = create_faiss_index(embeddings_array)

        save_faiss_index(index, faiss_index_path)
        save_metadata(ids, metadatas, contents, metadata_path)
    else:
        print("Using existing FAISS index and metadata.")

    print("Creating FAISS retriever...")
    faiss_retriever = FAISSRetriever(index, ids, metadatas, contents, embeddings)

    return faiss_retriever


async def run_evaluation():
    local_llm = "llama3-groq-tool-use:latest"
    llama3 = ChatOllama(model=local_llm, temperature=0)
    openai = ChatOpenAI(model_name="gpt-4o-mini", temperature=0)

    retriever = create_faiss_retriever()

    questions, ground_truths = load_qa_pairs()

    for question, ground_truth in zip(questions[:5], ground_truths[:5]):
        print(f"\nQuestion: {question}")

        start_time = time()
        llama3_docs, llama3_answer = faiss_query(retriever, question, llama3, multi_query=True)
        llama3_time = time() - start_time
        print(f"llama3 Answer: {llama3_answer}")
        print(f"llama3 Time: {llama3_time:.4f} seconds")


        llama3_datasets = {
            "question": [question],
            "answer": [llama3_answer],
            "contexts": [[doc.page_content for doc in llama3_docs]],
            "ground_truth": [ground_truth]
        }
        llama3_evalsets = Dataset.from_dict(llama3_datasets)

        llama3_result = evaluate(
            llama3_evalsets,
            metrics=[
                context_precision,
                faithfulness,
                answer_relevancy,
                context_recall,
            ],
        )

        print("llama3 RAGAS Evaluation:")
        llama3_result['time'] = llama3_time
        df = llama3_result.to_pandas()
        print(df)
        
        df.to_csv("llama.csv", mode='a')
        #############################################################
        start_time = time()
        openai_docs, openai_answer = faiss_query(retriever, question, openai, multi_query=True)
        openai_time = time() - start_time
        print(f"openai Answer: {openai_answer}")
        print(f"openai Time: {openai_time:.4f} seconds")


        openai_datasets = {
            "question": [question],
            "answer": [openai_answer],
            "contexts": [[doc.page_content for doc in openai_docs]],
            "ground_truth": [ground_truth]
        }
        openai_evalsets = Dataset.from_dict(openai_datasets)

        openai_result = evaluate(
            openai_evalsets,
            metrics=[
                context_precision,
                faithfulness,
                answer_relevancy,
                context_recall,
            ],
        )

        print("openai RAGAS Evaluation:")
        openai_result['time'] = llama3_time
        df = openai_result.to_pandas()
        print(df)
        
        df.to_csv("openai.csv", mode='a')


    print("\nPerformance comparison complete.")

def load_kg_q():
    import pandas as pd
    sheet_id ="1eVmO8fIjjtEQBlmqtipi2d0H-8VgJkI-icqhSTfdhHQ"
    gid = "0"
    df = pd.read_csv(f"https://docs.google.com/spreadsheets/d/{sheet_id}/export?format=csv&gid={gid}")
    
    return df['Question'].tolist()
    
async def run_q_batch():
    # local_llm = "llama3-groq-tool-use:latest"
    # llama3 = ChatOllama(model=local_llm, temperature=0)
    openai = ChatOpenAI(model_name="gpt-4o-mini", temperature=0)

    retriever = create_faiss_retriever()

    questions = load_kg_q()
    result_list = []

    for question in questions:
        print(f"\nQuestion: {question}")

        with get_openai_callback() as cb:
            start_time = time()
            # openai_docs, openai_answer = faiss_query(retriever, question, openai, multi_query=True)
            documents, answer = faiss_query(retriever, question, openai, multi_query=True)
            openai_time = time() - start_time
            print(f"Answer: {answer}")
            print(f"Time: {openai_time:.4f} seconds")
            save_history(question, answer, cb, openai_time)
        
        result = {'Question': question, 'Answer': answer}
        result_list.append(result)

    df = pd.DataFrame.from_records(result_list)
    print(df)
        
    df.to_csv("kg_qa.csv", mode='w')

    print("\nQA complete.")
    
def save_history(question, answer, cb, processing_time):
    # reference = [doc.dict() for doc in reference]
    record = {
        'Question': question,
        'Answer': answer,
        'Total_Tokens': cb.total_tokens,
        'Total_Cost': cb.total_cost,
        'Processing_time': processing_time,
    }
    response = (
        supabase.table("agent_records")
        .insert(record)
        .execute()
    )
if __name__ == "__main__":
    # asyncio.run(run_evaluation())
    asyncio.run(run_q_batch())
    # print(load_kg_q())