app.py

import os
import streamlit as st
import re
from datetime import datetime
from groq import Groq
from langchain_community.vectorstores import FAISS
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_huggingface import HuggingFaceEmbeddings
import pandas as pd


# Load environment variables
try:
    from dotenv import load_dotenv
    load_dotenv()
except ImportError:
    pass  # dotenv not installed, likely running on Hugging Face Spaces

# Function to get the API key


def get_api_key():
    api_key = os.environ.get("GROQ_API_KEY")
    if not api_key:
        api_key = st.secrets.get("GROQ_API_KEY")
    if not api_key:
        st.error(
            "GROQ_API_KEY is not set. Please set it in your environment or Streamlit secrets.")
        st.stop()
    return api_key


def parse_transcript(content):

    parsed_segments = []
    current_speaker = ""
    current_company = ""
    current_timestamp = ""

    # Split the content into paragraphs
    paragraphs = re.split(r'\n\s*\n', content)

    for paragraph in paragraphs:
        paragraph = paragraph.strip()
        if not paragraph:
            continue

        # Check if the paragraph contains speaker information
        speaker_match = re.match(r'(.*?),\s*(.*?)\((.*?)\):', paragraph)
        if speaker_match:
            current_speaker, current_company, current_timestamp = speaker_match.groups()
            text = paragraph.split('\n', 1)[1] if '\n' in paragraph else ''

        # Check if the paragraph contains only a timestamp
        elif re.match(r'\((.*?)\):', paragraph):
            current_timestamp = re.match(r'\((.*?)\):', paragraph).group(1)
            text = re.sub(r'^\(.*?\):\s*', '', paragraph).strip()

        # If it's not a speaker line or timestamp line, it's just text
        else:
            text = paragraph

        # Add the segment
        if text:
            add_segment(parsed_segments, current_speaker,
                        current_company, current_timestamp, text)

    return parsed_segments


def add_segment(parsed_segments, speaker, company, timestamp, text):
    segment = {
        "speaker": speaker,
        "company": company,
        "timestamp": timestamp,
        "text": text
    }
    parsed_segments.append(segment)
    print_segment(speaker, company, timestamp, text)


def print_segment(speaker, company, timestamp, text):
    print(f"Speaker: {speaker}")
    print(f"Company: {company}")
    print(f"Timestamp: {timestamp}")
    print(f"Text: {text[:100]}...")  # Print first 100 characters of text
    print("-" * 50)


def create_searchable_segments(parsed_segments):
    searchable_segments = []
    for segment in parsed_segments:
        searchable_text = f"{segment['speaker']},{segment['company']},{segment['timestamp']}:: {
            segment['text']}"
        searchable_segments.append(searchable_text)
    return searchable_segments


# Load and parse the transcript
def load_transcript(content):
    global vectorstore

    # Parse the transcript
    parsed_transcript = parse_transcript(content)
    searchable_segments = create_searchable_segments(parsed_transcript)

    # Create text splitter and split the searchable segments
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=1000, chunk_overlap=200)

    splits = text_splitter.create_documents(searchable_segments)

    # Create vector store with HuggingFaceEmbeddings
    embeddings = HuggingFaceEmbeddings()
    vectorstore = FAISS.from_documents(documents=splits, embedding=embeddings)


def search_transcript(query, k=30):
    # Perform similarity search
    docs = vectorstore.similarity_search(query=query, k=k)

    # Format results
    results = []
    for doc in docs:
        content = doc.page_content
        match = re.match(r'(.*?),(.*?),(.*?)::\s*(.*)', content, re.DOTALL)
        if match:
            speaker, company, timestamp, text = match.groups()
            results.append({
                "speaker": speaker.strip(),
                "company": company.strip(),
                "timestamp": timestamp.strip(),
                "text": text.strip()
            })

    return results


# Groq client setup
client = Groq(api_key=get_api_key())


def generate_response(query, search_results):
    # Prepare the prompt with search results
    prompt = f"""You are a friendly assistant. Your job is to answer the user's question based on the transcript excerpts provided below:

Transcript excerpts:
{search_results}

Question: {query}

Please provide a concise and relevant answer based on the information in the transcript excerpts. If the information is not directly related to the question, say so and provide the most relevant information available."""

    completion = client.chat.completions.create(
        model="llama3-8b-8192",
        messages=[
            {
                "role": "user",
                "content": prompt
            }
        ],
        temperature=0.5,
        max_tokens=3000,
        top_p=1,
        stream=False,
        stop=None,
    )

    return completion.choices[0].message.content


# Streamlit app
def main():
    st.title("Transcript Search and Q&A")

    st.caption("This site takes a TelecomTV video transcript and allows a chat session with it.  If no transcript is provided it defaults to this one: https://www.telecomtv.com/content/dsp-leaders-forum/enabling-the-autonomous-network-with-ai-50536/")

    # File upload
    uploaded_file = st.file_uploader("Upload a transcript file", type="txt")

    if uploaded_file is None:
        file_name = "Enabling the autonomous network with AI"
        with open("example-transcript.txt", 'r') as file:
            content = file.read()
    else:
        content = uploaded_file.getvalue().decode("utf-8")
        file_name = uploaded_file.name

    # Read and process the uploaded file
    load_transcript(content)

    st.subheader(f"Chat with {file_name}")

    # User input
    user_query = st.text_input(
        "Enter your question:", placeholder="e.g.What are people speaking about? or List all people speaking")

    # Add a slider for selecting the number of results
    num_results = st.slider("Number of relevant transcript excerpts to show:",
                            min_value=1, max_value=50, value=30, step=1)

    if user_query:
        search_results = search_transcript(user_query, k=num_results)
        formatted_results = "\n\n".join([f"{result['speaker']} {result['company']} ({result['timestamp']}): {
                                        result['text']}" for result in search_results])

        response = generate_response(user_query, formatted_results)

        st.subheader("Assistant's response:")
        st.write(response)

        st.subheader("Relevant transcript excerpts:")

        # Create a DataFrame from the search results
        df = pd.DataFrame(search_results)

        # Rename columns for better readability
        df.columns = ['Speaker', 'Company', 'Timestamp', 'Quote']

        # Display the DataFrame as a table
        st.table(df)


if __name__ == "__main__":
    main()