import google.generativeai as genai import requests import numpy as np import faiss from sentence_transformers import SentenceTransformer from bs4 import BeautifulSoup import gradio as gr # Configure Gemini API key gemini_api_secret_name = 'AIzaSyA0yLvySmj8xjMd0sedSgklg1fj0wBDyyw' from google.colab import userdata try: GOOGLE_API_KEY = userdata.get(gemini_api_secret_name) genai.configure(api_key=GOOGLE_API_KEY) except userdata.SecretNotFoundError as e: print(f'Secret not found\n\nThis expects you to create a secret named {gemini_api_secret_name} in Colab\n\nVisit https://makersuite.google.com/app/apikey to create an API key\n\nStore that in the secrets section on the left side of the notebook (key icon)\n\nName the secret {gemini_api_secret_name}') raise e except userdata.NotebookAccessError as e: print(f'You need to grant this notebook access to the {gemini_api_secret_name} secret in order for the notebook to access Gemini on your behalf.') raise e except Exception as e: # unknown error print(f"There was an unknown error. Ensure you have a secret {gemini_api_secret_name} stored in Colab and it's a valid key from https://makersuite.google.com/app/apikey") raise e # Fetch lecture notes and model architectures def fetch_lecture_notes(): lecture_urls = [ "https://stanford-cs324.github.io/winter2022/lectures/introduction/", "https://stanford-cs324.github.io/winter2022/lectures/capabilities/", "https://stanford-cs324.github.io/winter2022/lectures/data/", "https://stanford-cs324.github.io/winter2022/lectures/modeling/" ] lecture_texts = [] for url in lecture_urls: response = requests.get(url) if response.status_code == 200: print(f"Fetched content from {url}") lecture_texts.append((extract_text_from_html(response.text), url)) else: print(f"Failed to fetch content from {url}, status code: {response.status_code}") return lecture_texts def fetch_model_architectures(): url = "https://github.com/Hannibal046/Awesome-LLM#milestone-papers" response = requests.get(url) if response.status_code == 200: print(f"Fetched model architectures, status code: {response.status_code}") return extract_text_from_html(response.text), url else: print(f"Failed to fetch model architectures, status code: {response.status_code}") return "", url # Extract text from HTML content def extract_text_from_html(html_content): soup = BeautifulSoup(html_content, 'html.parser') for script in soup(["script", "style"]): script.extract() text = soup.get_text(separator="\n", strip=True) return text # Generate embeddings using SentenceTransformers def create_embeddings(texts, model): texts_only = [text for text, _ in texts] embeddings = model.encode(texts_only) return embeddings # Initialize FAISS index def initialize_faiss_index(embeddings): dimension = embeddings.shape[1] # Assuming all embeddings have the same dimension index = faiss.IndexFlatL2(dimension) index.add(embeddings.astype('float32')) return index # Handle natural language queries conversation_history = [] def handle_query(query, faiss_index, embeddings_texts, model): global conversation_history query_embedding = model.encode([query]).astype('float32') # Search FAISS index _, indices = faiss_index.search(query_embedding, 3) # Retrieve top 3 results relevant_texts = [embeddings_texts[idx] for idx in indices[0]] # Combine relevant texts and truncate if necessary combined_text = "\n".join([text for text, _ in relevant_texts]) max_length = 500 # Adjust as necessary if len(combined_text) > max_length: combined_text = combined_text[:max_length] + "..." # Generate a response using Gemini try: response = genai.generate_text( model="models/text-bison-001", prompt=f"Based on the following context:\n\n{combined_text}\n\nAnswer the following question: {query}", max_output_tokens=200 ) generated_text = response.result except Exception as e: print(f"Error generating text: {e}") generated_text = "An error occurred while generating the response." # Update conversation history conversation_history.append(f"User: {query}") conversation_history.append(f"System: {generated_text}") # Extract sources sources = [url for _, url in relevant_texts] return generated_text, sources def generate_concise_response(prompt, context): try: response = genai.generate_text( model="models/text-bison-001", prompt=f"{prompt}\n\nContext: {context}\n\nAnswer:", max_output_tokens=200 ) return response.result except Exception as e: print(f"Error generating concise response: {e}") return "An error occurred while generating the concise response." # Main function to execute the pipeline def chatbot(message , history): lecture_notes = fetch_lecture_notes() model_architectures = fetch_model_architectures() all_texts = lecture_notes + [model_architectures] # Load the SentenceTransformers model embedding_model = SentenceTransformer('paraphrase-MiniLM-L6-v2') embeddings = create_embeddings(all_texts, embedding_model) # Initialize FAISS index faiss_index = initialize_faiss_index(np.array(embeddings)) response, sources = handle_query(message, faiss_index, all_texts, embedding_model) print("Query:", message) print("Response:", response) total_text = response if sources: print("Sources:", sources) relevant_source = "" for source in sources: relevant_source += source +"\n" total_text += "\n\nSources:\n" + relevant_source else: print("Sources: None of the provided sources were used.") print("----") # Generate a concise and relevant summary using Gemini prompt = "Summarize the user queries so far" user_queries_summary = " ".join(message) concise_response = generate_concise_response(prompt, user_queries_summary) print("Concise Response:") print(concise_response) return total_text iface = gr.ChatInterface( chatbot, title="LLM Research Assistant", description="Ask questions about LLM architectures, datasets, and training techniques.", examples=[ "What are some milestone model architectures in LLMs?", "Explain the transformer architecture.", "Tell me about datasets used to train LLMs.", "How are LLM training datasets cleaned and preprocessed?", "Summarize the user queries so far" ], retry_btn="Regenerate", undo_btn="Undo", clear_btn="Clear", ) if __name__ == "__main__": iface.launch(debug=True)