import os import json import gradio as gr from huggingface_hub import HfApi, login from PyPDF2 import PdfReader from langchain_huggingface import HuggingFaceEmbeddings from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain_groq import ChatGroq from dotenv import load_dotenv from langchain.docstore.document import Document from langchain.schema import Document from chunk_python_code import chunk_python_code_with_metadata from vectorstore import get_chroma_vectorstore from download_repo_to_huggingface import download_and_upload_kadiAPY_repo_to_huggingfacespace from process_repo_zipfile import extract_files_and_filepath_from_dir from ragchain import RAGChain from llm import get_groq_llm # Load environment variables from .env file load_dotenv() # Load configuration from JSON file with open('config.json') as config_file: config = json.load(config_file) with open("config2.json", "r") as file: config2 = json.load(file) VECTORSTORE_DIRECTORY = config["vectorstore_directory"] CHUNK_SIZE = config["chunk_size"] CHUNK_OVERLAP = config["chunk_overlap"] EMBEDDING_MODEL_NAME = config["embedding_model"] LLM_MODEL_NAME = config["llm_model"] LLM_TEMPERATURE = config["llm_temperature"] HF_SPACE_NAME = config["hf_space_name"] DATA_DIR = config["data_dir"] GROQ_API_KEY = os.environ["GROQ_API_KEY"] HF_TOKEN = os.environ["HF_Token"] login(HF_TOKEN) api = HfApi() def split_python_code_into_chunks(texts, file_paths): chunks = [] for text, file_path in zip(texts, file_paths): """ Custom made python code splitter, algorithm iterates through child nodes of ast-tree(max child depth = 2) aims to have full body of methods along signature (+ can handle decorators) in a chunk and adds method specific metadata e.g visbility: public, _internal type: "class", "methods", "command"(CLI commands) source: with the intend to use a filter when retrieving potentaion useful snippets. """ document_chunks = chunk_python_code_with_metadata(text, file_path) chunks.extend(document_chunks) return chunks # Split text into chunks def split_into_chunks(texts, references, chunk_size, chunk_overlap): text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap) chunks = [] for text, reference in zip(texts, references): chunks.extend([ Document( page_content=chunk, metadata={ "source": reference, "usage": "doc" } ) for chunk in text_splitter.split_text(text) ]) return chunks # Setup Vectorstore def embed_documents_into_vectorstore(chunks, model_name, persist_directory): print("Start setup_vectorstore_function") embedding_model = HuggingFaceEmbeddings(model_name=model_name) vectorstore = get_chroma_vectorstore(embedding_model, persist_directory) vectorstore.add_documents(chunks) return vectorstore def rag_workflow(query): """ RAGChain class to perform the complete RAG workflow. """ # Assume 'llm' and 'vector_store' are already initialized instances rag_chain = RAGChain(llm, vector_store) """ Pre-Retrieval-Stage """ # predict which python library to search in: (standard) kadiAPY-library or kadiAPY-cli-library code_library_usage_prediction = rag_chain.predict_library_usage(query) print(f"Predicted library usage: {code_library_usage_prediction}") rewritten_query = rag_chain.rewrite_query(query) print(f"\n\n Rewritten query: {rewritten_query}\n\n") """ Retrieval-Stage """ kadiAPY_doc_documents = rag_chain.retrieve_contexts(query, k=5, filter={"usage": "doc"}) kadiAPY_code_documents = rag_chain.retrieve_contexts(str(rewritten_query.content), k=3, filter={"usage": code_library_usage_prediction}) print("Retrieved Document Contexts:", kadiAPY_doc_documents) print("Retrieved Code Contexts:", kadiAPY_code_documents) """ Pre-Generation-Stage Adding each doc's metadata to the retrieved content (docs & code snippets) """ formatted_doc_snippets = rag_chain.format_documents(kadiAPY_doc_documents) formatted_code_snippets = rag_chain.format_documents(kadiAPY_code_documents) #print("FORMATTED Retrieved Document Contexts:", formatted_doc_snippets) #print("FORMATTED Retrieved Code Contexts:" , formatted_code_snippets) """ Generation-Stage """ response = rag_chain.generate_response(query, formatted_doc_snippets, formatted_code_snippets) print("Generated Response:", response) return response def initialize(): global vector_store, chunks, llm download_and_upload_kadiAPY_repo_to_huggingfacespace( api_url=config2["gitlab"]["api_url"], project_id=config2["gitlab"]["project"]["id"], version=config2["gitlab"]["project"]["version"] ) code_texts, code_references = extract_files_and_filepath_from_dir(DATA_DIR, ['kadi_apy'], []) #doc_texts, doc_references = extract_files_and_filepath_from_dir(DATA_DIR, ['docs/source/'], []) doc_texts, doc_references = extract_files_and_filepath_from_dir(DATA_DIR, [], []) print("LEEEEEEEEEEEENGTH of code_texts: ", len(code_texts)) print("LEEEEEEEEEEEENGTH of doc_files: ", len(doc_texts)) code_chunks = split_python_code_into_chunks(code_texts, code_references) doc_chunks = split_into_chunks(doc_texts, doc_references, CHUNK_SIZE, CHUNK_OVERLAP) print(f"Total number of code_chunks: {len(code_chunks)}") print(f"Total number of doc_chunks: {len(doc_chunks)}") vector_store = embed_documents_into_vectorstore(doc_chunks + code_chunks, EMBEDDING_MODEL_NAME, VECTORSTORE_DIRECTORY) llm = get_groq_llm(LLM_MODEL_NAME, LLM_TEMPERATURE, GROQ_API_KEY) from langchain_community.document_loaders import TextLoader initialize() # Gradio utils def check_input_text(text): if not text: gr.Warning("Please input a question.") raise TypeError return True def add_text(history, text): history = history + [(text, None)] yield history, "" import gradio as gr def bot_kadi(history): user_query = history[-1][0] response = rag_workflow(user_query) history[-1] = (user_query, response) yield history def main(): with gr.Blocks() as demo: gr.Markdown("## KadiAPY - AI Coding-Assistant") gr.Markdown("AI assistant for KadiAPY based on RAG architecture powered by LLM") with gr.Tab("KadiAPY - AI Assistant"): with gr.Row(): with gr.Column(scale=10): chatbot = gr.Chatbot([], elem_id="chatbot", label="Kadi Bot", bubble_full_width=False, show_copy_button=True, height=600) user_txt = gr.Textbox(label="Question", placeholder="Type in your question and press Enter or click Submit") with gr.Row(): with gr.Column(scale=1): submit_btn = gr.Button("Submit", variant="primary") with gr.Column(scale=1): clear_btn = gr.Button("Clear", variant="stop") gr.Examples( examples=[ "Who is working on Kadi4Mat?", "How do i install the Kadi-Apy library?", "How do i install the Kadi-Apy library for development?", "I need a method to upload a file to a record", ], inputs=user_txt, outputs=chatbot, fn=add_text, label="Try asking...", cache_examples=False, examples_per_page=3, ) user_txt.submit(check_input_text, user_txt, None).success(add_text, [chatbot, user_txt], [chatbot, user_txt]).then(bot_kadi, [chatbot], [chatbot]) submit_btn.click(check_input_text, user_txt, None).success(add_text, [chatbot, user_txt], [chatbot, user_txt]).then(bot_kadi, [chatbot], [chatbot]) clear_btn.click(lambda: None, None, chatbot, queue=False) demo.launch() if __name__ == "__main__": main()