app.py

import os
import logging
import asyncio
import gradio as gr
from huggingface_hub import InferenceClient
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS
from langchain.schema import Document
from duckduckgo_search import DDGS

# Environment variables and configurations
huggingface_token = os.environ.get("HUGGINGFACE_TOKEN")

MODELS = [
    "mistralai/Mistral-7B-Instruct-v0.3",
    "mistralai/Mixtral-8x7B-Instruct-v0.1",
    "mistralai/Mistral-Nemo-Instruct-2407",
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    "meta-llama/Meta-Llama-3.1-70B-Instruct"
]

def get_embeddings():
    return HuggingFaceEmbeddings(model_name="sentence-transformers/stsb-roberta-large")

def duckduckgo_search(query):
    with DDGS() as ddgs:
        results = ddgs.text(query, max_results=10)
    return results

async def rephrase_query(query, history, model):
    system_message = """You are an AI assistant tasked with analyzing and rephrasing user queries. Your goal is to determine if a query is unique or related to the previous conversation, and then rephrase it appropriately for web search. Keep the rephrased query succinct and in a web search query format.

    If the query is unique, rephrase it to be more specific and searchable.
    If the query is related to the previous conversation, incorporate relevant context from the previous response.

    Provide your analysis in the following format:
    <analysis>Your reasoning about whether the query is unique or related</analysis>
    <rephrased_query>The rephrased query</rephrased_query>"""

    user_message = f"""Current query: {query}

    Previous conversation history:
    {history}

    Analyze the query and provide a rephrased version suitable for web search."""

    client = InferenceClient(model, token=huggingface_token)
    
    try:
        response = await asyncio.to_thread(
            client.text_generation,
            prompt=f"{system_message}\n\n{user_message}",
            max_new_tokens=150,
            temperature=0.2,
        )
        
        # Extract the rephrased query from the response
        analysis_start = response.find("<analysis>")
        analysis_end = response.find("</analysis>")
        rephrased_start = response.find("<rephrased_query>")
        rephrased_end = response.find("</rephrased_query>")
        
        if analysis_start != -1 and analysis_end != -1 and rephrased_start != -1 and rephrased_end != -1:
            analysis = response[analysis_start + 10:analysis_end].strip()
            rephrased_query = response[rephrased_start + 17:rephrased_end].strip()
            return analysis, rephrased_query
        else:
            logging.error("Failed to parse the rephrase response")
            return None, query
    except Exception as e:
        logging.error(f"Error in rephrase_query: {str(e)}")
        return None, query

def create_web_search_vectors(search_results):
    embed = get_embeddings()
    documents = []
    for result in search_results:
        if 'body' in result:
            content = f"{result['title']}\n{result['body']}\nSource: {result['href']}"
            documents.append(Document(page_content=content, metadata={"source": result['href']}))
    return FAISS.from_documents(documents, embed)

async def get_response_with_search(query, model, use_embeddings, num_calls=3, temperature=0.2):
    search_results = duckduckgo_search(query)
    
    if not search_results:
        yield "No web search results available. Please try again.", ""
        return

    if use_embeddings:
        web_search_database = create_web_search_vectors(search_results)
        retriever = web_search_database.as_retriever(search_kwargs={"k": 5})
        relevant_docs = retriever.get_relevant_documents(query)
        context = "\n".join([doc.page_content for doc in relevant_docs])
    else:
        context = "\n".join([f"{result['title']}\n{result['body']}\nSource: {result['href']}" for result in search_results])

    system_message = """ You are a world-class AI system, capable of complex reasoning and reflection. 
    Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags.
    Providing comprehensive and accurate information based on web search results is essential. 
    Your goal is to synthesize the given context into a coherent and detailed response that directly addresses the user's query. 
    Please ensure that your response is well-structured, factual, and cites sources where appropriate.
    If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags."""

    user_message = f"""Using the following context from web search results:
{context}

Write a detailed and complete research document that fulfills the following user request: '{query}'
After writing the document, please provide a list of sources used in your response."""

    # Use Hugging Face API
    client = InferenceClient(model, token=huggingface_token)
    full_response = ""
    
    try:
        for _ in range(num_calls):
            for response in client.chat_completion(
                messages=[
                    {"role": "system", "content": system_message},
                    {"role": "user", "content": user_message}
                ],
                max_tokens=6000,
                temperature=temperature,
                stream=True,
                top_p=0.8,
            ):
                if isinstance(response, dict) and "choices" in response:
                    for choice in response["choices"]:
                        if "delta" in choice and "content" in choice["delta"]:
                            chunk = choice["delta"]["content"]
                            full_response += chunk
                            yield full_response, ""
                else:
                    logging.error("Unexpected response format or missing attributes in the response object.")
                    break
    except Exception as e:
        logging.error(f"Error in get_response_with_search: {str(e)}")
        yield f"An error occurred while processing your request: {str(e)}", ""

    if not full_response:
        logging.warning("No response generated from the model")
        yield "No response generated from the model.", ""

async def respond(message, history, model, temperature, num_calls, use_embeddings):
    logging.info(f"User Query: {message}")
    logging.info(f"Model Used: {model}")
    logging.info(f"Temperature: {temperature}")
    logging.info(f"Number of API Calls: {num_calls}")
    logging.info(f"Use Embeddings: {use_embeddings}")

    try:
        # Rephrase the query
        analysis, rephrased_query = await rephrase_query(message, history, model)
        
        if analysis:
            yield f"Query Analysis: {analysis}\n\nRephrased Query: {rephrased_query}\n\nSearching the web...\n\n"
        
        async for main_content, sources in get_response_with_search(rephrased_query, model, use_embeddings, num_calls=num_calls, temperature=temperature):
            response = f"{main_content}\n\n{sources}"
            yield response
    except asyncio.CancelledError:
        yield "The operation was cancelled. Please try again."
    except Exception as e:
        logging.error(f"Error in respond function: {str(e)}")
        yield f"An error occurred: {str(e)}"

css = """
/* Fine-tune chatbox size */
.chatbot-container {
    height: 600px !important;
    width: 100% !important;
}
.chatbot-container > div {
    height: 100%;
    width: 100%;
}
"""

# Gradio interface setup
def create_gradio_interface():
    custom_placeholder = "Enter your question here for web search."

    demo = gr.ChatInterface(
        respond,
        additional_inputs=[
            gr.Dropdown(choices=MODELS, label="Select Model", value=MODELS[3]),
            gr.Slider(minimum=0.1, maximum=1.0, value=0.2, step=0.1, label="Temperature"),
            gr.Slider(minimum=1, maximum=5, value=1, step=1, label="Number of API Calls"),
            gr.Checkbox(label="Use Embeddings", value=False),
        ],
        title="AI-powered Web Search Assistant",
        description="Use web search to answer questions or generate summaries.",
        theme=gr.Theme.from_hub("allenai/gradio-theme"),
        css=css,
        examples=[
            ["What are the latest developments in artificial intelligence?"],
            ["Explain the concept of quantum computing."],
            ["What are the environmental impacts of renewable energy?"]
        ],
        cache_examples=False,
        analytics_enabled=False,
        textbox=gr.Textbox(placeholder=custom_placeholder, container=False, scale=7),
        chatbot=gr.Chatbot(
            show_copy_button=True,
            likeable=True,
            layout="bubble",
            height=400,
        )
    )

    with demo:
        gr.Markdown("""
        ## How to use
        1. Enter your question in the chat interface.
        2. Select the model you want to use from the dropdown.
        3. Adjust the Temperature to control the randomness of the response.
        4. Set the Number of API Calls to determine how many times the model will be queried.
        5. Check or uncheck the "Use Embeddings" box to toggle between using embeddings or direct text summarization.
        6. Press Enter or click the submit button to get your answer.
        7. Use the provided examples or ask your own questions.
        """)

    return demo

if __name__ == "__main__":
    demo = create_gradio_interface()
    demo.launch(share=True)