Systex/RAG_app.py

from fastapi import FastAPI, Request, HTTPException, status, Body
# from fastapi.templating import Jinja2Templates
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse
from fastapi import Depends
from contextlib import asynccontextmanager
from pydantic import BaseModel
from typing import List, Optional
import uvicorn

from typing import List, Optional
import sqlparse
from sqlalchemy import create_engine
import pandas as pd
#from retrying import retry
import datetime
import json
from json import loads
import pandas as pd
import time
from langchain.callbacks import get_openai_callback

from langchain_community.vectorstores import Chroma
from langchain_openai import OpenAIEmbeddings
from RAG_strategy import multi_query, naive_rag, naive_rag_for_qapairs
from Indexing_Split import create_retriever as split_retriever
from Indexing_Split import gen_doc_from_database, gen_doc_from_history

from dotenv import load_dotenv
import os
from langchain_community.vectorstores import SupabaseVectorStore
from langchain_openai import OpenAIEmbeddings
from supabase.client import Client, create_client


from add_vectordb import GetVectorStore

load_dotenv()
URI = os.getenv("SUPABASE_URI")

global_retriever = None

@asynccontextmanager
async def lifespan(app: FastAPI):
    global global_retriever
    global vector_store
    
    start = time.time()
    # global_retriever = split_retriever(path='./Documents', extension="docx")
    # global_retriever = raptor_retriever(path='../Documents', extension="txt")
    # global_retriever = unstructured_retriever(path='../Documents')

    supabase_url = os.environ.get("SUPABASE_URL")
    supabase_key = os.environ.get("SUPABASE_KEY")
    document_table = "documents"
    supabase: Client = create_client(supabase_url, supabase_key)

    embeddings = OpenAIEmbeddings()
    vector_store = GetVectorStore(embeddings, supabase, document_table)
    global_retriever = vector_store.as_retriever(search_kwargs={"k": 4})

    print(time.time() - start)
    yield

def get_retriever():
    return global_retriever


def get_vector_store():
    return vector_store

app = FastAPI(lifespan=lifespan)

# templates = Jinja2Templates(directory="temp")
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

@app.get("/answer2")
def multi_query_answer(question, retriever=Depends(get_retriever)):
    start = time.time()

    with get_openai_callback() as cb:
        # qa_doc = gen_doc_from_database()
        # qa_history_doc = gen_doc_from_history()
        # qa_doc.extend(qa_history_doc)
        # vectorstore = Chroma.from_documents(documents=qa_doc, embedding=OpenAIEmbeddings(), collection_name="qa_pairs")
        # retriever_qa = vectorstore.as_retriever(search_kwargs={"k": 3})
        # final_answer, reference_docs = naive_rag_for_qapairs(question, retriever_qa)
        final_answer = 'False'
        if final_answer == 'False':
            final_answer, reference_docs = multi_query(question, retriever, chat_history=[])

    # print(CHAT_HISTORY)
    
    # with get_openai_callback() as cb:
    #     final_answer, reference_docs = multi_query(question, retriever)
    processing_time = time.time() - start
    print(processing_time)
    save_history(question, final_answer, reference_docs, cb, processing_time)

    return {"Answer": final_answer}

class ChatHistoryItem(BaseModel):
    q: str
    a: str

@app.post("/answer_with_history")
def multi_query_answer(question: Optional[str] = '', chat_history: List[ChatHistoryItem] = Body(...), retriever=Depends(get_retriever)):
    start = time.time()
    
    chat_history = [(item.q, item.a) for item in chat_history if item.a != ""]
    print(chat_history)

    # TODO: similarity search
    
    with get_openai_callback() as cb:
        final_answer, reference_docs = multi_query(question, retriever, chat_history)
    processing_time = time.time() - start
    print(processing_time)
    save_history(question, final_answer, reference_docs, cb, processing_time)

    return {"Answer": final_answer}


@app.post("/answer_with_history2")
def multi_query_answer(question: Optional[str] = '', extension: Optional[str] = 'pdf', chat_history: List[ChatHistoryItem] = Body(...), retriever=Depends(get_retriever)):
    start = time.time()

    retriever = vector_store.as_retriever(search_kwargs={"k": 4,
                                                         'filter': {'extension':extension}})
    
    chat_history = [(item.q, item.a) for item in chat_history if item.a != ""]
    print(chat_history)

    # TODO: similarity search
    
    with get_openai_callback() as cb:
        final_answer, reference_docs = multi_query(question, retriever, chat_history)
    processing_time = time.time() - start
    print(processing_time)
    save_history(question, final_answer, reference_docs, cb, processing_time)

    return {"Answer": final_answer}

def save_history(question, answer, reference, 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],
        'Contexts': [str(reference)]
    }
    df = pd.DataFrame(record)
    engine = create_engine(URI)
    df.to_sql(name='systex_records', con=engine, index=False, if_exists='append')

class history_output(BaseModel):
    Question: str
    Answer: str
    Contexts: str
    Total_Tokens: int
    Total_Cost: float
    Processing_time: float
    Time: datetime.datetime
    
@app.get('/history', response_model=List[history_output])
async def get_history():
    engine = create_engine(URI, echo=True)

    df = pd.read_sql_table("systex_records", engine.connect())  
    df.fillna('', inplace=True)
    result = df.to_json(orient='index', force_ascii=False)
    result = loads(result)
    return result.values()

if __name__ == "__main__":
    uvicorn.run("RAG_app:app", host='cmm.ai', port=8081, reload=True, ssl_keyfile="/etc/letsencrypt/live/cmm.ai/privkey.pem", 
                ssl_certfile="/etc/letsencrypt/live/cmm.ai/fullchain.pem")