app.py

from fastapi import FastAPI, File, UploadFile, Form, Request, HTTPException
from fastapi.responses import HTMLResponse
from fastapi.templating import Jinja2Templates
from typing import List, Optional
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_community.document_loaders import PyPDFLoader, UnstructuredCSVLoader, UnstructuredExcelLoader, Docx2txtLoader, UnstructuredPowerPointLoader
from langchain.chains import StuffDocumentsChain
from langchain.chains.llm import LLMChain
from langchain.prompts import PromptTemplate
from langchain.vectorstores import FAISS
from langchain_google_genai import GoogleGenerativeAIEmbeddings
import json
import os
import google.generativeai as genai
import re
import nest_asyncio
import nltk
from langchain.text_splitter import CharacterTextSplitter


app = FastAPI()
templates = Jinja2Templates(directory="templates")

if os.getenv("FASTAPI_ENV") == "development":
    nest_asyncio.apply()


nltk.download('averaged_perceptron_tagger_eng')
from nltk.tokenize import word_tokenize

# Initialize your model and other variables
uploaded_file_path = None
document_analyzed = False
summary = None
question_responses = []
api = None
llm = None

safety_settings = [
    {"category": "HARM_CATEGORY_DANGEROUS", "threshold": "BLOCK_NONE"},
    {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"},
    {"category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "BLOCK_NONE"},
    {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "BLOCK_NONE"},
    {"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "BLOCK_NONE"},
]

def format_text(text: str) -> str:
    text = re.sub(r'\*\*(.*?)\*\*', r'<b>\1</b>', text)
    text = text.replace('*', '<br>')
    return text



# Route for main page
@app.get("/", response_class=HTMLResponse)
async def read_main(request: Request):
    return templates.TemplateResponse("analyze.html", {
        "request": request,
        "summary": summary,
        "show_conversation": document_analyzed,
        "question_responses": question_responses
    })

# Route for analyzing documents
@app.post("/", response_class=HTMLResponse)
async def analyze_document(

    request: Request,

    api_key: str = Form(...),

    iam: str = Form(...),

    context: str = Form(...),

    output: str = Form(...),

    summary_length: str = Form(...),

    file: UploadFile = File(...)

):
    global uploaded_file_path, document_analyzed, summary, question_responses, api, llm
    loader = None

    try:
        # Initialize or update API key and models
        api = api_key
        genai.configure(api_key=api)
        llm = ChatGoogleGenerativeAI(model="gemini-1.5-flash-latest", google_api_key=api)

        # Save the uploaded file
        uploaded_file_path = "uploaded_file" + os.path.splitext(file.filename)[1]
        with open(uploaded_file_path, "wb") as f:
            f.write(file.file.read())

        # Determine the file type and load accordingly
        file_extension = os.path.splitext(uploaded_file_path)[1].lower()
        print(f"File extension: {file_extension}")  # Debugging statement

        if file_extension == ".pdf":
            loader = PyPDFLoader(uploaded_file_path)
        elif file_extension == ".csv":
            loader = UnstructuredCSVLoader(uploaded_file_path, mode="elements", encoding="utf8")
        elif file_extension == ".xlsx":
            loader = UnstructuredExcelLoader(uploaded_file_path, mode="elements")
        elif file_extension == ".docx":
            loader = Docx2txtLoader(uploaded_file_path)
        elif file_extension == ".pptx":
            loader = UnstructuredPowerPointLoader(uploaded_file_path)
        elif file_extension == ".mp3":
            # Process audio files differently
            audio_file = genai.upload_file(path=uploaded_file_path)
            model = genai.GenerativeModel(model_name="gemini-1.5-flash")
            prompt = f"I am an {iam}. This file is about {context}. Answer the question based on this file: {output}. Write a {summary_length} concise summary."
            response = model.generate_content([prompt, audio_file], safety_settings=safety_settings)
            summary = format_text(response.text)
            document_analyzed = True
            outputs = {"summary": summary}
            with open("output_summary.json", "w") as outfile:
                json.dump(outputs, outfile)
            return templates.TemplateResponse("analyze.html", {
                "request": request,
                "summary": summary,
                "show_conversation": document_analyzed,
                "question_responses": question_responses
            })

        # If no loader is set, raise an exception
        if loader is None:
            raise HTTPException(status_code=400, detail=f"Unsupported file type: {file_extension}")

        docs = loader.load()
        prompt_template = PromptTemplate.from_template(
            f"I am an {iam}. This file is about {context}. Answer the question based on this file: {output}. Write a {summary_length} concise summary of the following text: {{text}}"
        )
        llm_chain = LLMChain(llm=llm, prompt=prompt_template)
        stuff_chain = StuffDocumentsChain(llm_chain=llm_chain, document_variable_name="text")
        response = stuff_chain.invoke(docs)
        summary = format_text(response["output_text"])
        document_analyzed = True
        outputs = {"summary": summary}
        with open("output.json", "w") as outfile:
            json.dump(outputs, outfile)
        return templates.TemplateResponse("analyze.html", {
            "request": request,
            "summary": summary,
            "show_conversation": document_analyzed,
            "question_responses": question_responses
        })

    except Exception as e:
        raise HTTPException(status_code=500, detail=f"An error occurred: {e}")

# Route for asking questions
from langchain.text_splitter import CharacterTextSplitter  # Ensure this is imported

@app.post("/ask", response_class=HTMLResponse)
async def ask_question(request: Request, question: str = Form(...)):
    global uploaded_file_path, question_responses, llm, api

    loader = None

    if uploaded_file_path:
        # Determine the file type and load accordingly
        file_extension = os.path.splitext(uploaded_file_path)[1].lower()
        if file_extension == ".pdf":
            loader = PyPDFLoader(uploaded_file_path)
        elif file_extension == ".csv":
            loader = UnstructuredCSVLoader(uploaded_file_path, mode="elements")
        elif file_extension == ".xlsx":
            loader = UnstructuredExcelLoader(uploaded_file_path, mode="elements")
        elif file_extension == ".docx":
            loader = Docx2txtLoader(uploaded_file_path)
        elif file_extension == ".pptx":
            loader = UnstructuredPowerPointLoader(uploaded_file_path)
        elif file_extension == ".mp3":
            audio_file = genai.upload_file(path=uploaded_file_path)
            model = genai.GenerativeModel(model_name="gemini-1.5-flash")
            latest_conversation = request.cookies.get("latest_question_response", "")
            prompt = "Answer the question based on the speech: " + question + (f" Latest conversation: {latest_conversation}" if latest_conversation else "")
            response = model.generate_content([prompt, audio_file], safety_settings=safety_settings)
            current_response = response.text
            current_question = f"You asked: {question}"

            # Save the latest question and response to the session
            question_responses.append((current_question, current_response))

            # Perform vector embedding and search
            text = current_response  # Use the summary generated from the MP3 content
            os.environ["GOOGLE_API_KEY"] = api
            embeddings = GoogleGenerativeAIEmbeddings(model="models/embedding-001")
            summary_embedding = embeddings.embed_query(text)
            document_search = FAISS.from_texts([text], embeddings)

            if document_search:
                query_embedding = embeddings.embed_query(question)
                results = document_search.similarity_search_by_vector(query_embedding, k=1)

                if results:
                    current_response = results[0].page_content
                else:
                    current_response = "No matching document found in the database."
            else:
                current_response = "Vector database not initialized."

            # Append the question and response from FAISS search
            question_responses.append((current_question, current_response))

            # Save all results including FAISS response to output.json
            save_to_json(summary, question_responses)

            # Save the latest question and response to the session
            response = templates.TemplateResponse("analyze.html", {"request": request, "summary": summary, "show_conversation": document_analyzed, "question_responses": question_responses})
            response.set_cookie(key="latest_question_response", value=current_response)
            return response
        
        # If no loader is set, raise an exception
        if loader is None:
            raise HTTPException(status_code=400, detail=f"Unsupported file type: {file_extension}")

        docs = loader.load()
        text = "\n".join([doc.page_content for doc in docs])
        os.environ["GOOGLE_API_KEY"] = api

        # Split the text into chunks
        text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
        chunks = text_splitter.split_text(text)

        # Define the Summarize Chain for the question
        latest_conversation = request.cookies.get("latest_question_response", "")
        template1 = question + """ answer the question based on the following:

                    "{text}" 

                    :""" + (f" Answer the Question with no more than 3 sentences. Latest conversation: {latest_conversation}" if latest_conversation else "")
        
        current_response = ""
        for chunk in chunks:
            prompt1 = PromptTemplate.from_template(template1.format(text=chunk))
            # Initialize the LLMChain with the prompt
            llm_chain1 = LLMChain(llm=llm, prompt=prompt1)
            response1 = llm_chain1.invoke({"text": chunk})
            current_response += response1["text"] + "\n"

        # Generate embeddings for the combined responses
        embeddings = GoogleGenerativeAIEmbeddings(model="models/embedding-001")
        summary_embedding = embeddings.embed_query(current_response)
        document_search = FAISS.from_texts([current_response], embeddings)

        # Perform a search on the FAISS vector database if it's initialized
        if document_search:
            query_embedding = embeddings.embed_query(question)
            results = document_search.similarity_search_by_vector(query_embedding, k=1)

            if results:
                current_response = format_text(results[0].page_content)
            else:
                current_response = "No matching document found in the database."
        else:
            current_response = "Vector database not initialized."

        # Append the question and response from FAISS search
        current_question = f"You asked: {question}"
        question_responses.append((current_question, current_response))

        # Save all results to output.json
        save_to_json(summary, question_responses)

        # Save the latest question and response to the session
        response = templates.TemplateResponse("analyze.html", {"request": request, "summary": summary, "show_conversation": document_analyzed, "question_responses": question_responses})
        response.set_cookie(key="latest_question_response", value=current_response)
        return response
    else:
        raise HTTPException(status_code=400, detail="No file has been uploaded yet.")



def save_to_json(summary, question_responses):
    outputs = {
        "summary": summary,
        "question_responses": question_responses
    }
    with open("output_summary.json", "w") as outfile:
        json.dump(outputs, outfile)

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="127.0.0.1", port=8000)