import os import json import gradio as gr import zipfile import tempfile import requests import urllib.parse import io from huggingface_hub import HfApi, login from PyPDF2 import PdfReader from langchain_huggingface import HuggingFaceEmbeddings from langchain_community.vectorstores import Chroma from langchain.text_splitter import RecursiveCharacterTextSplitter from langchain_groq import ChatGroq from dotenv import load_dotenv from langchain.docstore.document import Document # 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) PERSIST_DOC_DIRECTORY = config["persist_doc_directory"] PERSIST_CODE_DIRECTORY =config["persist_code_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"] GITLAB_API_URL = config["gitlab_api_url"] HF_SPACE_NAME = config["hf_space_name"] REPOSITORY_DIRECTORY = config["repository_directory"] GROQ_API_KEY = os.environ["GROQ_API_KEY"] HF_TOKEN = os.environ["HF_Token"] login(HF_TOKEN) api = HfApi() def load_project_id(json_file): with open(json_file, 'r') as f: data = json.load(f) return data['project_id'] def download_gitlab_repo(): print("Start the upload_gitRepository function") project_id = load_project_id('repository_ids.json') encoded_project_id = urllib.parse.quote_plus(project_id) # Define the URL to download the repository archive archive_url = f"{GITLAB_API_URL}/projects/{encoded_project_id}/repository/archive.zip" # Download the repository archive response = requests.get(archive_url) archive_bytes = io.BytesIO(response.content) # Retrieve the original file name from the response headers content_disposition = response.headers.get('content-disposition') if content_disposition: filename = content_disposition.split('filename=')[-1].strip('\"') else: filename = 'archive.zip' # Fallback to a default name if not found # Check if the file already exists in the repository existing_files = api.list_repo_files(repo_id=HF_SPACE_NAME, repo_type='space') target_path = f"{REPOSITORY_DIRECTORY}/{filename}" print(f"Target Path: '{target_path}'") print(f"Existing Files: {[repr(file) for file in existing_files]}") if target_path in existing_files: print(f"File '{target_path}' already exists in the repository. Skipping upload...") else: # Upload the ZIP file to the new folder in the Hugging Face space repository print("Uploading File to directory:") print(f"Archive Bytes: {repr(archive_bytes.getvalue())[:100]}") # Show a preview of bytes print(f"Target Path in Repo: '{target_path}'") api.upload_file( path_or_fileobj=archive_bytes, path_in_repo=target_path, repo_id=HF_SPACE_NAME, repo_type='space' ) print("Upload complete") def get_all_files_in_folder(temp_dir, partial_path): all_files = [] print("inner method of get all files in folder") target_dir = os.path.join(temp_dir, partial_path) print(target_dir) for root, dirs, files in os.walk(target_dir): print(f"Files in current directory ({root}): {files}") for file in files: print(f"Processing file: {file}") all_files.append(os.path.join(root, file)) return all_files def get_file(temp_dir, file_path): full_path = os.path.join(temp_dir, file_path) return full_path def process_directory(directory, partial_paths=None, file_paths=None): all_texts = [] file_references = [] zip_files = [file for file in os.listdir(directory) if file.endswith('.zip')] if not zip_files: print("No zip file found in the directory.") return all_texts, file_references if len(zip_files) > 1: print("More than one zip file found.") return all_texts, file_references else: zip_file_path = os.path.join(directory, zip_files[0]) # Create a temporary directory for the zip file with tempfile.TemporaryDirectory() as tmpdirname: # Unzip the file into the temporary directory with zipfile.ZipFile(zip_file_path, 'r') as zip_ref: zip_ref.extractall(tmpdirname) print(f"Extracted {zip_file_path} to {tmpdirname}") files = [] unzipped_root = os.listdir(tmpdirname) if len(unzipped_root) == 1 and os.path.isdir(os.path.join(tmpdirname, unzipped_root[0])): tmpsubdirpath= os.path.join(tmpdirname, unzipped_root[0]) else: tmpsubdirpath = tmpdirname if not partial_paths and not file_paths: for root, _, files_list in os.walk(tmpdirname): for file in files_list: files.append(os.path.join(root, file)) else: if partial_paths: for partial_path in partial_paths: files += get_all_files_in_folder(tmpsubdirpath, partial_path) if file_paths: files += [get_file(tmpsubdirpath, file_path) for file_path in file_paths] print(f"Total number of files: {len(files)}") for file_path in files: #print(f"Paths of files: {iles}") file_ext = os.path.splitext(file_path)[1] if os.path.getsize(file_path) == 0: print(f"Skipping an empty file: {file_path}") continue with open(file_path, 'rb') as f: if file_ext in ['.rst', '.md', '.txt', '.html', '.json', '.yaml', '.py']: text = f.read().decode('utf-8') elif file_ext in ['.svg']: text = f"SVG file content from {file_path}" elif file_ext in ['.png', '.ico']: text = f"Image metadata from {file_path}" else: continue all_texts.append(text) file_references.append(file_path) return all_texts, file_references import ast def get_source_segment(source_lines, node): start_line, start_col = node.lineno - 1, node.col_offset end_line = node.end_lineno - 1 if hasattr(node, 'end_lineno') else node.lineno - 1 end_col = node.end_col_offset if hasattr(node, 'end_col_offset') else len(source_lines[end_line]) lines = source_lines[start_line:end_line + 1] lines[0] = lines[0][start_col:] lines[-1] = lines[-1][:end_col] return ''.join(lines) from langchain.schema import Document def chunk_python_file_content(content, char_limit=1572): source_lines = content.splitlines(keepends=True) # Parse the content into an abstract syntax tree (AST) tree = ast.parse(content) chunks = [] current_chunk = "" current_chunk_size = 0 # Find all class definitions and top-level functions in the AST class_nodes = [node for node in ast.walk(tree) if isinstance(node, ast.ClassDef)] for class_node in class_nodes: method_nodes = [node for node in class_node.body if isinstance(node, ast.FunctionDef)] if method_nodes: first_method_start_line = method_nodes[0].lineno - 1 class_def_lines = source_lines[class_node.lineno - 1:first_method_start_line] else: class_def_lines = source_lines[class_node.lineno - 1:class_node.end_lineno] class_def = ''.join(class_def_lines) class_def_size = len(class_def) # Add class definition to the current chunk if it fits if current_chunk_size + class_def_size <= char_limit: current_chunk += f"{class_def.strip()}\n" current_chunk_size += class_def_size else: # Start a new chunk if the class definition exceeds the limit if current_chunk: chunks.append(current_chunk.strip()) current_chunk = "" current_chunk_size = 0 current_chunk += f"{class_def.strip()}\n" current_chunk_size = class_def_size for method_node in method_nodes: method_def = get_source_segment(source_lines, method_node) method_def_size = len(method_def) # Add method definition to the current chunk if it fits if current_chunk_size + method_def_size <= char_limit: current_chunk += f"# This is a class method of class: {class_node.name}\n{method_def.strip()}\n" current_chunk_size += method_def_size else: # Start a new chunk if the method definition exceeds the limit if current_chunk: chunks.append(current_chunk.strip()) current_chunk = "" current_chunk_size = 0 current_chunk += f"# This is a class method of class: {class_node.name}\n{method_def.strip()}\n" current_chunk_size = method_def_size if current_chunk: chunks.append(current_chunk.strip()) return chunks # Split python code into chunks def split_pythoncode_into_chunks(texts, references, chunk_size, chunk_overlap): chunks = [] for text, reference in zip(texts, references): file_chunks = chunk_python_file_content(text, char_limit=chunk_size) for chunk in file_chunks: document = Document(page_content=chunk, metadata={"source": reference}) chunks.append(document) print(f"Total number of chunks: {len(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}) for chunk in text_splitter.split_text(text)]) print(f"Total number of chunks: {len(chunks)}") return chunks # Setup Vectorstore def setup_vectorstore(chunks, model_name, persist_directory): embedding_model = HuggingFaceEmbeddings(model_name=model_name) vectorstore = Chroma.from_documents(chunks, embedding=embedding_model, persist_directory=persist_directory) return vectorstore # Setup LLM def setup_llm(model_name, temperature, api_key): llm = ChatGroq(model=model_name, temperature=temperature, api_key=api_key) return llm def retrieve_from_vectorstore(vectorstore, query, k): results = vectorstore.similarity_search(query, k=k) chunks_with_references = [(result.page_content, result.metadata["source"]) for result in results] # Print the chosen chunks and their sources to the console print("\nChosen chunks and their sources for the query:") for chunk, source in chunks_with_references: print(f"Source: {source}\nChunk: {chunk}\n") print("-" * 50) return chunks_with_references def rag_workflow(query): retrieved_doc_chunks = retrieve_from_vectorstore(docstore, query, k=5) retrieved_code_chunks = retrieve_from_vectorstore(codestore, query, k=5) doc_context = "\n\n".join([doc_chunk for doc_chunk, _ in retrieved_doc_chunks]) code_context = "\n\n".join([code_chunk for code_chunk, _ in retrieved_code_chunks]) doc_references = "\n".join([f"[{i+1}] {ref}" for i, (_, ref) in enumerate(retrieved_doc_chunks)]) code_references = "\n".join([f"[{i+1}] {ref}" for i, (_, ref) in enumerate(retrieved_code_chunks)]) print("Document Chunks:\n") print("\n\n".join(["="*80 + "\n" + doc_chunk for doc_chunk, _ in retrieved_doc_chunks])) print("\nDocument References:\n") print(doc_references) print("\n" + "="*80 + "\n") # Separator between doc and code print("Code Chunks:\n") print("\n\n".join(["="*80 + "\n" + code_chunk for code_chunk, _ in retrieved_code_chunks])) print("\nCode References:\n") print(code_references) # print(f"Context for the query:\n{doc_context}\n") # print(f"References for the query:\n{references}\n") prompt = f"""You are an expert python developer. You are assisting in generating code for users who wants to make use of "kadi-apy", an API library. "Doc-context:" provides you with information how to use this API library by givnig code examples and code documentation. "Code-context:" provides you information of API methods and classes from the "kadi-apy" library. Based on the retrieved contexts and the guidelines answer the query. General Guidelines: - If no related information is found from the contexts to answer the query, reply that you do not know. Guidelines when generating code: - First display the full code and then follow with a well structured explanation of the generated code. Doc-context: {doc_context} Code-context: {code_context} Query: {query} """ response = llm.invoke(prompt) return response.content def initialize(): global docstore, codestore, chunks, llm code_partial_paths = ['kadi_apy/lib/'] code_file_path = [] doc_partial_paths = ['docs/source/setup/'] doc_file_paths = ['docs/source/usage/lib.rst'] code_files, code_file_references = process_directory(REPOSITORY_DIRECTORY, code_partial_paths, code_file_path) doc_files, doc_file_references = process_directory(REPOSITORY_DIRECTORY, doc_partial_paths, doc_file_paths) code_chunks = split_pythoncode_into_chunks(code_files, code_file_references, 1500, 0) doc_chunks = split_into_chunks(doc_files, doc_file_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)}") docstore = setup_vectorstore(doc_chunks, EMBEDDING_MODEL_NAME, PERSIST_DOC_DIRECTORY) codestore = setup_vectorstore(code_chunks, EMBEDDING_MODEL_NAME, PERSIST_CODE_DIRECTORY) llm = setup_llm(LLM_MODEL_NAME, LLM_TEMPERATURE, GROQ_API_KEY) 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("## Kadi4Mat - AI Chat-Bot") gr.Markdown("AI assistant for Kadi4Mat based on RAG architecture powered by LLM") with gr.Tab("Kadi4Mat - 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]) #user_txt.submit(check_input_text, user_txt, None).success(add_text, [chatbot, user_txt], [chatbot, user_txt]).then(bot_kadi, [chatbot], [chatbot, doc_citation]) #submit_btn.click(check_input_text, user_txt, None).success(add_text, [chatbot, user_txt], [chatbot, user_txt]).then(bot_kadi, [chatbot], [chatbot, doc_citation]) clear_btn.click(lambda: None, None, chatbot, queue=False) demo.launch() if __name__ == "__main__": main()