app.py

import requests
import gradio as gr
from bs4 import BeautifulSoup
import logging
from urllib.parse import urlparse
from requests.adapters import HTTPAdapter
from requests.packages.urllib3.util.retry import Retry
from requests.exceptions import Timeout
from urllib.request import urlopen, Request
import json
from huggingface_hub import InferenceClient
import random
import time
from sentence_transformers import SentenceTransformer, util
import torch
from datetime import datetime
import os
from dotenv import load_dotenv
import certifi
import requests
from newspaper import Article
import PyPDF2
import io
import requests
import random
import datetime
from groq import Groq
import faiss
import numpy as np

# Automatically get the current year
current_year = datetime.datetime.now().year

# Load environment variables from a .env file
load_dotenv()

# Set up logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# SearXNG instance details
SEARXNG_URL = 'https://shreyas094-searxng-local.hf.space/search'
SEARXNG_KEY = 'f9f07f93b37b8483aadb5ba717f556f3a4ac507b281b4ca01e6c6288aa3e3ae5'

# Use the environment variable
HF_TOKEN = os.getenv("HF_TOKEN")
client = InferenceClient(
    "mistralai/Mistral-Nemo-Instruct-2407",
    token=HF_TOKEN,
)

# Default API key for examples (replace with a dummy value or leave empty)
GROQ_API_KEY = os.getenv("GROQ_API_KEY")

# Initialize Groq client
groq_client = Groq(api_key=GROQ_API_KEY)

# Initialize the similarity model
similarity_model = SentenceTransformer('all-MiniLM-L6-v2')

# Global variable to store the FAISS index
faiss_index = None
document_store = []

# Set up a session with retry mechanism
def requests_retry_session(
    retries=0,
    backoff_factor=0.1,
    status_forcelist=(500, 502, 504),
    session=None,
):
    session = session or requests.Session()
    retry = Retry(
        total=retries,
        read=retries,
        connect=retries,
        backoff_factor=backoff_factor,
        status_forcelist=status_forcelist,
    )
    adapter = HTTPAdapter(max_retries=retry)
    session.mount('http://', adapter)
    session.mount('https://', adapter)
    return session

def is_valid_url(url):
    try:
        result = urlparse(url)
        return all([result.scheme, result.netloc])
    except ValueError:
        return False

def scrape_pdf_content(url, max_chars=3000, timeout=5):
    try:
        logger.info(f"Scraping PDF content from: {url}")
        
        # Download the PDF file
        response = requests.get(url, timeout=timeout)
        response.raise_for_status()
        
        # Create a PDF reader object
        pdf_reader = PyPDF2.PdfReader(io.BytesIO(response.content))
        
        # Extract text from all pages
        content = ""
        for page in pdf_reader.pages:
            content += page.extract_text() + "\n"
        
        # Limit the content to max_chars
        return content[:max_chars] if content else ""
    except requests.Timeout:
        logger.error(f"Timeout error while scraping PDF content from {url}")
        return ""
    except Exception as e:
        logger.error(f"Error scraping PDF content from {url}: {e}")
        return ""

def scrape_with_newspaper(url):
    if url.lower().endswith('.pdf'):
        return scrape_pdf_content(url)
    
    logger.info(f"Starting to scrape with Newspaper3k: {url}")
    try:
        article = Article(url)
        article.download()
        article.parse()
        
        # Combine title and text
        content = f"Title: {article.title}\n\n"
        content += article.text
        
        # Add publish date if available
        if article.publish_date:
            content += f"\n\nPublish Date: {article.publish_date}"
        
        # Add authors if available
        if article.authors:
            content += f"\n\nAuthors: {', '.join(article.authors)}"
        
        # Add top image URL if available
        if article.top_image:
            content += f"\n\nTop Image URL: {article.top_image}"
        
        return content
    except Exception as e:
        logger.error(f"Error scraping {url} with Newspaper3k: {e}")
        return ""

def rephrase_query(chat_history, query, temperature=0.2):
    system_prompt = f"""
You are a highly intelligent and context-aware conversational assistant. Your tasks are as follows:

1. Determine if the new query is a continuation of the previous conversation or an entirely new topic.

2. For both continuations and new topics:
   a. **Entity Identification and Quotation**:
      - Analyze the user's query to identify the main entities (e.g., organizations, brands, products, locations).
      - For each identified entity, enclose ONLY the entity itself in double quotes within the query.
      - If no identifiable entities are found, proceed without adding quotes.
   b. **Query Preservation**:
      - Maintain the entire original query, including any parts after commas or other punctuation.
      - Do not remove or truncate any part of the original query.

3. If it's a continuation:
   - Incorporate relevant information from the context to make the query more specific and contextual.
   - Ensure that entities from the previous context are properly quoted if they appear in the rephrased query.

4. For both continuations and new topics:
   - Append "after: {current_year}" to the end of the rephrased query.
   - Ensure there is a space before "after:" for proper formatting.
   - Do not use quotes or the "+" operator when adding the year.

5. **Output**:
   - Return ONLY the rephrased query, ensuring it is concise, clear, and contextually accurate.
   - Do not include any additional commentary or explanation.

### Example Scenarios
**Scenario 1: New Topic**
- **User Query**: "What is the latest news on Golomt Bank?"
- **Rephrased Query**: "What is the latest news on \"Golomt Bank\" after: {current_year}"

**Scenario 2: Continuation**
- **Previous Query**: "What is the latest news on Golomt Bank?"
- **User Query**: "How did the Bank perform in Q2 2024?"
- **Rephrased Query**: "How did \"Golomt Bank\" perform in Q2 2024 after: {current_year}"

**Scenario 3: Query with Multiple Entities and Comma**
- **User Query**: "What is the latest news about Prospect Capital, did the rating change?"
- **Rephrased Query**: "What is the latest news about \"Prospect Capital\", did the rating change after: {current_year}"

**Scenario 4: Query Without Recognizable Entities**
- **User Query**: "How does photosynthesis work?"
- **Rephrased Query**: "How does photosynthesis work? after: {current_year}"
"""
    user_prompt = f"""
Conversation context:
{chat_history}
New query: {query}
Rephrased query:
"""
    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt}
    ]
    try:
        logger.info(f"Sending rephrasing request to LLM with temperature {temperature}")
        response = client.chat_completion(
            messages=messages,
            max_tokens=150,
            temperature=temperature
        )
        logger.info("Received rephrased query from LLM")
        rephrased_question = response.choices[0].message.content.strip()
        # Remove surrounding quotes if present
        if (rephrased_question.startswith('"') and rephrased_question.endswith('"')) or \
           (rephrased_question.startswith("'") and rephrased_question.endswith("'")):
            rephrased_question = rephrased_question[1:-1].strip()
        logger.info(f"Rephrased Query (cleaned): {rephrased_question}")
        return rephrased_question
    except Exception as e:
        logger.error(f"Error rephrasing query with LLM: {e}")
        return query  # Fallback to original query if rephrasing fails

def rerank_documents(query, documents, similarity_threshold=0.95, max_results=5):
    try:
        # Step 1: Encode the query and document summaries
        query_embedding = similarity_model.encode(query, convert_to_tensor=True)
        doc_summaries = [doc['summary'] for doc in documents]
        
        if not doc_summaries:
            logger.warning("No document summaries to rerank.")
            return documents
        
        doc_embeddings = similarity_model.encode(doc_summaries, convert_to_tensor=True)
        
        # Step 2: Compute Cosine Similarity
        cosine_scores = util.cos_sim(query_embedding, doc_embeddings)[0]
        
        # Combine documents and cosine scores
        scored_documents = list(zip(documents, cosine_scores))

        # Step 3: Sort documents by cosine similarity score
        scored_documents.sort(key=lambda x: x[1], reverse=True)
        
        # Step 4: Filter out similar documents
        filtered_docs = []
        for doc, score in scored_documents:
            if score < 0.5:  # If similarity to query is too low, skip
                continue
            
            # Check similarity with already selected documents
            is_similar = False
            for selected_doc in filtered_docs:
                similarity = util.pytorch_cos_sim(
                    similarity_model.encode(doc['summary'], convert_to_tensor=True),
                    similarity_model.encode(selected_doc['summary'], convert_to_tensor=True)
                )
                if similarity > similarity_threshold:
                    is_similar = True
                    break
            
            if not is_similar:
                filtered_docs.append(doc)
            
            if len(filtered_docs) >= max_results:
                break
        
        logger.info(f"Reranked and filtered to {len(filtered_docs)} unique documents.")
        return filtered_docs
    except Exception as e:
        logger.error(f"Error during reranking documents: {e}")
        return documents[:max_results]  # Fallback to first max_results documents if reranking fails

def compute_similarity(text1, text2):
    # Encode the texts
    embedding1 = similarity_model.encode(text1, convert_to_tensor=True)
    embedding2 = similarity_model.encode(text2, convert_to_tensor=True)
    
    # Compute cosine similarity
    cosine_similarity = util.pytorch_cos_sim(embedding1, embedding2)
    
    return cosine_similarity.item()

def is_content_unique(new_content, existing_contents, similarity_threshold=0.8):
    for existing_content in existing_contents:
        similarity = compute_similarity(new_content, existing_content)
        if similarity > similarity_threshold:
            return False
    return True

def assess_relevance_and_summarize(llm_client, query, document, temperature=0.2):
    system_prompt = """You are a world-class AI assistant specializing in financial news analysis. Your task is to assess the relevance of a given document to a user's query and provide a detailed summary if it's relevant."""

    user_prompt = f"""
Query: {query}

Document Title: {document['title']}
Document Content:
{document['content'][:1000]}  # Limit to first 1000 characters for efficiency

Instructions:
1. Assess if the document is relevant to the QUERY made by the user.
2. If relevant, provide a detailed summary that captures the unique aspects of this particular news item. Include:
   - Key facts and figures
   - Dates of events or announcements
   - Names of important entities mentioned
   - Any financial metrics or changes reported
   - The potential impact or significance of the news
3. If not relevant, simply state "Not relevant".

Your response should be in the following format:
Relevant: [Yes/No]
Summary: [Your detailed summary if relevant, or "Not relevant" if not]

Remember to focus on financial aspects and implications in your assessment and summary. Aim to make the summary distinctive, highlighting what makes this particular news item unique compared to similar news.
"""

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt}
    ]

    try:
        response = llm_client.chat_completion(
            messages=messages,
            max_tokens=300,  # Increased to allow for more detailed summaries
            temperature=temperature,
            top_p=0.9,
            frequency_penalty=1.4
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        logger.error(f"Error assessing relevance and summarizing with LLM: {e}")
        return "Error: Unable to assess relevance and summarize"

def scrape_full_content(url, scraper="bs4", max_chars=3000, timeout=5):
    try:
        logger.info(f"Scraping full content from: {url}")
        
        # Check if the URL ends with .pdf
        if url.lower().endswith('.pdf'):
            return scrape_pdf_content(url, max_chars, timeout)
        
        if scraper == "bs4":
            session = requests_retry_session()
            response = session.get(url, timeout=timeout)
            response.raise_for_status()
            soup = BeautifulSoup(response.content, 'html.parser')
            
            # Try to find the main content
            main_content = soup.find('main') or soup.find('article') or soup.find('div', class_='content')
            
            if main_content:
                content = main_content.get_text(strip=True, separator='\n')
            else:
                content = soup.get_text(strip=True, separator='\n')
        elif scraper == "trafilatura":
            content = scrape_with_trafilatura(url, max_chars, timeout, use_beautifulsoup=True)
        elif scraper == "scrapy":
            content = scrape_with_scrapy(url, timeout)
        elif scraper == "newspaper":
            content = scrape_with_newspaper(url)
        else:
            logger.error(f"Unknown scraper: {scraper}")
            return ""
        
        # Limit the content to max_chars
        return content[:max_chars] if content else ""
    except requests.Timeout:
        logger.error(f"Timeout error while scraping full content from {url}")
        return ""
    except Exception as e:
        logger.error(f"Error scraping full content from {url}: {e}")
        return ""

def llm_summarize(json_input, model, temperature=0.2):
    system_prompt = """You are Sentinel, a world-class Financial analysis AI model who is expert at searching the web and answering user's queries. You are also an expert at summarizing web pages or documents and searching for content in them."""

    user_prompt = f"""
Please provide a comprehensive summary based on the following JSON input:
{json_input}

Instructions:
1. Analyze the query and the provided documents.
2. Write a detailed, long, and complete research document that is informative and relevant to the user's query.
3. Use an unbiased and professional tone in your response.
4. Do not repeat text verbatim from the input.
5. Provide the answer in the response itself.
6. You can use markdown to format your response.
7. Use bullet points to list information where appropriate.
8. Cite the answer using [number] notation along with the appropriate source URL embedded in the notation.
9. Place these citations at the end of the relevant sentences.
10. You can cite the same sentence multiple times if it's relevant to different parts of your answer.

Your response should be detailed, informative, accurate, and directly relevant to the user's query."""

    messages = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt}
    ]

    try:
        if model == "groq":
            response = groq_client.chat.completions.create(
                messages=messages,
                model="llama-3.1-8b-instant",
                max_tokens=5500,
                temperature=temperature,
                top_p=0.9,
                presence_penalty=1.2,
                stream=False
            )
            return response.choices[0].message.content.strip()
        else:
            response = client.chat_completion(
                messages=messages,
                max_tokens=10000,
                temperature=temperature,
                frequency_penalty=1.4,
                top_p=0.9
            )
            return response.choices[0].message.content.strip()
    except Exception as e:
        logger.error(f"Error in LLM summarization: {e}")
        return "Error: Unable to generate a summary. Please try again."

def create_or_reset_faiss_index(dimension=384):  # 384 is the dimension for 'all-MiniLM-L6-v2' model
    global faiss_index
    faiss_index = faiss.IndexFlatL2(dimension)
    
def add_documents_to_faiss(documents):
    global faiss_index, document_store
    
    # Clear previous documents
    document_store.clear()
    
    # Create embeddings for the documents
    embeddings = []
    for doc in documents:
        # Combine title and content for embedding
        text_to_embed = f"{doc['title']} {doc['content'][:500]}"  # Limit content to first 500 chars for efficiency
        embedding = embedding_model.encode(text_to_embed)
        embeddings.append(embedding)
        document_store.append(doc)
    
    # Convert to numpy array
    embeddings_array = np.array(embeddings).astype('float32')
    
    # Add to FAISS index
    faiss_index.add(embeddings_array)

def search_similar_documents(query, k=5):
    global faiss_index, document_store
    
    # Create query embedding
    query_embedding = embedding_model.encode(query)
    query_embedding = np.array([query_embedding]).astype('float32')
    
    # Search in FAISS index
    distances, indices = faiss_index.search(query_embedding, k)
    
    # Retrieve similar documents
    similar_docs = [document_store[i] for i in indices[0]]

    return similar_docs

def search_and_scrape(query, chat_history, num_results=5, max_chars=3000, time_range="", language="all", category="", 
                      engines=[], safesearch=2, method="GET", llm_temperature=0.2, timeout=5, model="huggingface"):   
    try:
        # Step 1: Rephrase the Query
        rephrased_query = rephrase_query(chat_history, query, temperature=llm_temperature)
        logger.info(f"Rephrased Query: {rephrased_query}")

        if not rephrased_query or rephrased_query.lower() == "not_needed":
            logger.info("No need to perform search based on the rephrased query.")
            return "No search needed for the provided input."

        # Step 2: Perform search
        # Search query parameters
        params = {
            'q': rephrased_query,
            'format': 'json',
            'time_range': time_range,
            'language': language,
            'category': category,
            'engines': ','.join(engines),
            'safesearch': safesearch
        }

        # Remove empty parameters
        params = {k: v for k, v in params.items() if v != ""}

        # If no engines are specified, set default engines
        if 'engines' not in params:
            params['engines'] = 'google'  # Default to 'google' or any preferred engine
            logger.info("No engines specified. Defaulting to 'google'.")

        # Headers for SearXNG request
        headers = {
            'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36',
            'Accept': 'application/json, text/javascript, */*; q=0.01',
            'Accept-Language': 'en-US,en;q=0.5',
            'Origin': 'https://shreyas094-searxng-local.hf.space',
            'Referer': 'https://shreyas094-searxng-local.hf.space/',
            'DNT': '1',
            'Connection': 'keep-alive',
            'Sec-Fetch-Dest': 'empty',
            'Sec-Fetch-Mode': 'cors',
            'Sec-Fetch-Site': 'same-origin',
        }

        scraped_content = []
        page = 1
        while len(scraped_content) < num_results:
            # Update params with current page
            params['pageno'] = page

            # Send request to SearXNG
            logger.info(f"Sending request to SearXNG for query: {rephrased_query} (Page {page})")
            session = requests_retry_session()

            try:
                if method.upper() == "GET":
                    response = session.get(SEARXNG_URL, params=params, headers=headers, timeout=10, verify=certifi.where())
                else:  # POST
                    response = session.post(SEARXNG_URL, data=params, headers=headers, timeout=10, verify=certifi.where())
                
                response.raise_for_status()
            except requests.exceptions.RequestException as e:
                logger.error(f"Error during SearXNG request: {e}")
                return f"An error occurred during the search request: {e}"

            search_results = response.json()
            logger.debug(f"SearXNG Response: {search_results}")

            results = search_results.get('results', [])
            if not results:
                logger.warning(f"No more results returned from SearXNG on page {page}.")
                break

            for result in results:
                if len(scraped_content) >= num_results:
                    break
        
                url = result.get('url', '')
                title = result.get('title', 'No title')
        
                if not is_valid_url(url):
                    logger.warning(f"Invalid URL: {url}")
                    continue
        
                try:
                    logger.info(f"Processing content from: {url}")
                    
                    content = scrape_full_content(url, max_chars, timeout)
                    
                    if not content:
                        logger.warning(f"Failed to scrape content from {url}")
                        continue
                    
                    scraped_content.append({
                        "title": title,
                        "url": url,
                        "content": content,
                        "scraper": "pdf" if url.lower().endswith('.pdf') else "newspaper"
                    })
                    logger.info(f"Successfully scraped content from {url}. Total scraped: {len(scraped_content)}")
                except requests.exceptions.RequestException as e:
                    logger.error(f"Error scraping {url}: {e}")
                except Exception as e:
                    logger.error(f"Unexpected error while scraping {url}: {e}")

            page += 1

        if not scraped_content:
            logger.warning("No content scraped from search results.")
            return "No content could be scraped from the search results."

        logger.info(f"Successfully scraped {len(scraped_content)} documents.")

        # Step 3: Assess relevance, summarize, and check for uniqueness
        relevant_documents = []
        unique_summaries = []
        for doc in scraped_content:
            assessment = assess_relevance_and_summarize(client, rephrased_query, doc, temperature=llm_temperature)
            relevance, summary = assessment.split('\n', 1)

            if relevance.strip().lower() == "relevant: yes":
                summary_text = summary.replace("Summary: ", "").strip()
                
                if is_content_unique(summary_text, unique_summaries):
                    relevant_documents.append({
                        "title": doc['title'],
                        "url": doc['url'],
                        "summary": summary_text,
                        "scraper": doc['scraper']
                    })
                    unique_summaries.append(summary_text)
                else:
                    logger.info(f"Skipping similar content: {doc['title']}")

        if not relevant_documents:
            logger.warning("No relevant and unique documents found.")
            return "No relevant and unique financial news found for the given query."

        # Step 4: Rerank documents based on similarity to query
        reranked_docs = rerank_documents(rephrased_query, relevant_documents, similarity_threshold=0.95, max_results=num_results)
        
        if not reranked_docs:
            logger.warning("No documents remained after reranking.")
            return "No relevant financial news found after filtering and ranking."
        
        logger.info(f"Reranked and filtered to top {len(reranked_docs)} unique, finance-related documents.")

        # After Step 5: Scrape full content for top documents
        # Create or reset FAISS index
        create_or_reset_faiss_index()
        
        # Add documents to FAISS index
        add_documents_to_faiss(reranked_docs[:num_results])
        
        # Search for similar documents in the vector DB
        similar_docs = search_similar_documents(query, k=num_results)
        
        # Prepare JSON for LLM, now including similar documents from vector DB
        llm_input = {
            "query": query,
            "documents": [
                {
                    "title": doc['title'],
                    "url": doc['url'],
                    "summary": doc['summary'],
                    "full_content": doc['full_content']
                } for doc in reranked_docs[:num_results]
            ],
            "similar_documents": [
                {
                    "title": doc['title'],
                    "url": doc['url'],
                    "content": doc['content'][:500]  # Limit content for brevity
                } for doc in similar_docs
            ]
        }

        # Step 6: LLM Summarization (keep as is)
        llm_summary = llm_summarize(json.dumps(llm_input), model, temperature=llm_temperature)
        
        return llm_summary

    except Exception as e:
        logger.error(f"Unexpected error in search_and_scrape: {e}")
        return f"An unexpected error occurred during the search and scrape process: {e}"

def chat_function(message, history, num_results, max_chars, time_range, language, category, engines, safesearch, method, llm_temperature, model):
    chat_history = "\n".join([f"{role}: {msg}" for role, msg in history])
    
    response = search_and_scrape(
        query=message,
        chat_history=chat_history,
        num_results=num_results,
        max_chars=max_chars,
        time_range=time_range,
        language=language,
        category=category,
        engines=engines,
        safesearch=safesearch,
        method=method,
        llm_temperature=llm_temperature,
        model=model
    )

    yield response

iface = gr.ChatInterface(
    chat_function,
    title="Web Scraper for Financial News",
    description="Enter your query, and I'll search the web for the most recent and relevant financial news, scrape content, and provide summarized results.",
    theme=gr.Theme.from_hub("allenai/gradio-theme"),
    additional_inputs=[
        gr.Slider(5, 20, value=10, step=1, label="Number of initial results"),
        gr.Slider(500, 10000, value=1500, step=100, label="Max characters to retrieve"),
        gr.Dropdown(["", "day", "week", "month", "year"], value="", label="Time Range"),
        gr.Dropdown(["", "all", "en", "fr", "de", "es", "it", "nl", "pt", "pl", "ru", "zh"], value="", label="Language"),
        gr.Dropdown(["", "general", "news", "images", "videos", "music", "files", "it", "science", "social media"], value="", label="Category"),
        gr.Dropdown(
            ["google", "bing", "duckduckgo", "baidu", "yahoo", "qwant", "startpage"],
            multiselect=True,
            value=["google", "duckduckgo", "bing", "qwant"],
            label="Engines"
        ),
        gr.Slider(0, 2, value=2, step=1, label="Safe Search Level"),
        gr.Radio(["GET", "POST"], value="POST", label="HTTP Method"),
        gr.Slider(0, 1, value=0.2, step=0.1, label="LLM Temperature"),
        gr.Dropdown(["huggingface", "groq"], value="huggingface", label="LLM Model"),
    ],
    additional_inputs_accordion=gr.Accordion("⚙️ Advanced Parameters", open=True),
    retry_btn="Retry",
    undo_btn="Undo",
    clear_btn="Clear",
    chatbot=gr.Chatbot(
        show_copy_button=True,
        likeable=True,
        layout="bubble",
        height=500,
    )
)

if __name__ == "__main__":
    logger.info("Starting the SearXNG Scraper for Financial News using ChatInterface with Advanced Parameters")
    iface.launch(share=True)