import os import re import requests import uuid import datetime import zipfile import tempfile import shutil import secrets import time import json from typing import List, Tuple, Any, Dict, Set, Generator from urllib.parse import urljoin, urlparse # Third-party libraries import gradio as gr from huggingface_hub import InferenceClient, HfApi, hf_hub_download from huggingface_hub.utils import HfHubHTTPError from pypdf import PdfReader from bs4 import BeautifulSoup import nltk # --- CONFIGURATION --- class Config: """Centralized configuration for the Maestro application.""" HF_MODEL = os.getenv("HF_MODEL", "mistralai/Mixtral-8x7B-Instruct-v0.1") HF_TOKEN = os.getenv("HF_TOKEN") HF_DATASET_REPO = "Omnibus/tmp" # As specified in the user's script MEMORY_MAIN_PATH = "mem-test2/main.json" MEMORY_INDEX_PATH = "mem-test2/index.json" MEMORY_DATA_PATH = "mem-test2" VERBOSE = os.getenv("VERBOSE", "True").lower() == "true" MAX_TOKENS_SYNTHESIS = 4096 MAX_TOKENS_REPORT = 8192 MAX_TOKENS_CHAT = 2048 MAX_DATA_CHUNK = 20000 # For processing large text bodies REQUESTS_TIMEOUT = 20 USER_AGENT = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/125.0.0.0 Safari/537.36' # --- PROMPT LIBRARY (Integrated for simplicity) --- class PromptLibrary: """A centralized library of meticulously crafted prompt templates.""" AGENT_PREFIX = """ You are Maestro, an Expert Information Retrieval and Synthesis Agent. Your operation is governed by these directives: 1. Ethical Safeguard [v2.4]: Refuse to process harmful, illegal, or unethical requests. 2. Temporal Awareness: Use the timestamp {dynamic_timestamp_utc} to evaluate data relevance. 3. Contextual Prioritization: Analyze the user's purpose '{user_purpose}' to weigh data relevance. """ COMPRESS_JSON = """ Task: {task} Based on the AGENT_PREFIX context and the following data, generate a structured and concise JSON summary. Input Data Chunk: --- {history} --- Existing Knowledge (for context): --- {knowledge} --- Instructions: Compile and categorize the data above into a JSON dictionary string. Extract key information, group related entities, and ensure the output is a single, valid JSON object. """ COMPRESS_REPORT = """ Task: {task} Based on the AGENT_PREFIX context and the summarized knowledge you have, compile a detailed, exhaustive report (~8000 words). Summarized Knowledge: --- {knowledge} --- Last Chunk of Raw Data (for final context): --- {history} --- Instructions: Synthesize all provided information into a single, comprehensive narrative. Be thorough, detailed, and structure the report with clear headings and sections. """ SAVE_MEMORY = """ Task: {task} Data: --- {history} --- Instructions: Compile and categorize the data above into a JSON dictionary string. Include ALL text, datapoints, titles, descriptions, and source urls indexed into an easy to search JSON format. Required keys: "keywords", "title", "description", "content", "url". The "keywords" list should be comprehensive. """ RECALL_MEMORY = """ The user will give you a query and a list of keywords from a database index. Your duty is to choose the words from the list that are most closely related to the search query. If no keywords are relevant, return an empty list: []. Respond only with a single, valid JSON list of strings. USER QUERY: {prompt} KEYWORD LIST: {keywords} """ # --- UTILITIES --- def log(message: str) -> None: if Config.VERBOSE: print(f"[{datetime.datetime.now(datetime.timezone.utc).isoformat()}] {message}") # --- CORE APPLICATION ENGINE --- class MaestroEngine: """Handles all data processing, memory management, and LLM interaction.""" def __init__(self): if not Config.HF_TOKEN: raise ValueError("HF_TOKEN environment variable not set!") self.client = InferenceClient(model=Config.HF_MODEL, token=Config.HF_TOKEN) self.api = HfApi(token=Config.HF_TOKEN) try: nltk.data.find("tokenizers/punkt") except LookupError: log("Downloading NLTK 'punkt' tokenizer...") nltk.download('punkt', quiet=True) log("MaestroEngine initialized.") # --- Data Ingestion --- def _read_pdf_from_path(self, path: str) -> str: try: return "\n".join(page.extract_text() or "" for page in PdfReader(path).pages) except Exception as e: return f"Error reading PDF {os.path.basename(path)}: {e}" def _read_pdf_from_url(self, url: str) -> str: try: response = requests.get(url, stream=True, timeout=Config.REQUESTS_TIMEOUT) response.raise_for_status() with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp_file: tmp_file.write(response.content) return self._read_pdf_from_path(tmp_file.name) except Exception as e: return f"Failed to download or read PDF from {url}: {e}" finally: if 'tmp_file' in locals() and os.path.exists(tmp_file.name): os.remove(tmp_file.name) def _get_web_text(self, url: str) -> str: try: response = requests.get(url, headers={'User-Agent': Config.USER_AGENT}, timeout=Config.REQUESTS_TIMEOUT) response.raise_for_status() return BeautifulSoup(response.content, 'lxml').get_text(separator="\n", strip=True) except Exception as e: return f"Failed to fetch URL {url}: {e}" def process_data_sources(self, text: str, files: List[str], url: str, pdf_url: str, pdf_batch: str) -> Tuple[str, List[str]]: """Orchestrates data ingestion from all provided sources.""" all_content, errors = [], [] if text: all_content.append(text) if url: all_content.append(self._get_web_text(url)) if pdf_url: all_content.append(self._read_pdf_from_url(pdf_url)) if pdf_batch: urls = [u.strip() for u in pdf_batch.split(',') if u.strip()] for u in urls: content = self._read_pdf_from_url(u) if content.startswith("Error"): errors.append(content) else: all_content.append(content) if files: for path in files: if not path: continue filename, ext = os.path.basename(path), os.path.splitext(path)[1].lower() if ext == '.pdf': all_content.append(self._read_pdf_from_path(path)) elif ext == '.txt': with open(path, 'r', encoding='utf-8', errors='ignore') as f: all_content.append(f.read()) else: errors.append(f"Unsupported file type: {filename}") return "\n\n---\n\n".join(all_content), errors # --- LLM Interaction --- def _run_gpt(self, prompt_template: str, max_tokens: int, **kwargs) -> str: """Core LLM call function.""" system_prompt = PromptLibrary.AGENT_PREFIX.format( dynamic_timestamp_utc=datetime.datetime.now(datetime.timezone.utc).isoformat(), user_purpose=kwargs.get('task', 'completing a system task.') ) full_prompt = f"[INST] {system_prompt}\n\n{prompt_template.format(**kwargs)} [/INST]" log(f"Running GPT. Template: {prompt_template[:50]}...") try: return self.client.text_generation(full_prompt, max_new_tokens=max_tokens, temperature=0.8, top_p=0.95).strip() except Exception as e: log(f"LLM Error: {e}") return f'{{"error": "LLM call failed", "details": "{e}"}}' def _chunk_and_process(self, text: str, prompt_template: str, task: str, max_tokens: int) -> List[str]: """Chunks large text and processes each chunk with an LLM.""" text_len = len(text) if text_len == 0: return [] num_chunks = (text_len + Config.MAX_DATA_CHUNK - 1) // Config.MAX_DATA_CHUNK chunk_size = (text_len + num_chunks - 1) // num_chunks results, knowledge = [], "" for i in range(num_chunks): chunk = text[i*chunk_size : (i+1)*chunk_size] log(f"Processing chunk {i+1}/{num_chunks}...") resp = self._run_gpt(prompt_template, max_tokens, task=task, knowledge=knowledge, history=chunk) knowledge = resp if len(resp) < 2000 else resp[:2000] # Use response as context for next chunk results.append(resp) return results # --- Synthesis & Reporting Workflow --- def synthesis_workflow(self, text: str, task: str, do_summarize: bool, do_report: bool) -> Tuple[str, List[Dict]]: """Handles the multi-stage summarization and reporting process.""" if not text: return "No data to process.", [] json_summary_objects, final_report = [], "" if do_summarize or do_report: # Summarization is a prerequisite for reporting log("Starting summarization stage...") summaries = self._chunk_and_process(text, PromptLibrary.COMPRESS_JSON, task, Config.MAX_TOKENS_SYNTHESIS) for s in summaries: try: json_summary_objects.append(json.loads(s)) except json.JSONDecodeError: json_summary_objects.append({"error": "Failed to parse summary JSON", "raw": s}) log("Summarization stage complete.") if do_report: log("Starting report generation stage...") # Use the collected JSON summaries as knowledge for the final report knowledge_for_report = json.dumps(json_summary_objects, indent=2) final_report = self._run_gpt(PromptLibrary.COMPRESS_REPORT, Config.MAX_TOKENS_REPORT, task=task, knowledge=knowledge_for_report, history="All data chunks have been summarized.") log("Report generation complete.") return final_report, json_summary_objects return "Summarization complete.", json_summary_objects # --- Persistent Memory System --- def _hf_download_json(self, repo_path: str, default: Any = []) -> Any: try: path = hf_hub_download(repo_id=Config.HF_DATASET_REPO, filename=repo_path, repo_type="dataset", token=Config.HF_TOKEN) with open(path, 'r') as f: return json.load(f) except HfHubHTTPError: return default # File doesn't exist, return default except (json.JSONDecodeError, IOError): return default def _hf_upload_json(self, data: Any, repo_path: str): with tempfile.NamedTemporaryFile(mode='w+', delete=False, suffix=".json") as tmp_file: json.dump(data, tmp_file, indent=4) tmp_path = tmp_file.name self.api.upload_file(path_or_fileobj=tmp_path, path_in_repo=repo_path, repo_id=Config.HF_DATASET_REPO, repo_type="dataset") os.remove(tmp_path) def save_to_memory(self, text: str, task: str) -> List[Dict]: """Saves processed text to the Hugging Face Dataset repo.""" log("Starting memory save process...") json_chunks = self._chunk_and_process(text, PromptLibrary.SAVE_MEMORY, task, Config.MAX_TOKENS_SYNTHESIS) parsed_chunks, main_file = [], self._hf_download_json(Config.MEMORY_MAIN_PATH) for i, chunk_str in enumerate(json_chunks): try: data = json.loads(chunk_str) ts = datetime.datetime.now(datetime.timezone.utc) filename = f"{ts.strftime('%Y-%m-%d-%H-%M-%S')}-{uuid.uuid4().hex[:8]}.json" self._hf_upload_json(data, f"{Config.MEMORY_DATA_PATH}/{filename}") main_file.append({"file_name": filename, "keywords": data.get("keywords", []), "description": data.get("description", "")}) parsed_chunks.append(data) except json.JSONDecodeError: log(f"Could not parse memory chunk {i} into JSON.") self._hf_upload_json(main_file, Config.MEMORY_MAIN_PATH) self.update_keyword_index(main_file) log("Memory save complete.") return parsed_chunks def update_keyword_index(self, main_file_content: List[Dict]): log("Updating keyword index...") keyword_index = {} for entry in main_file_content: for keyword in entry.get("keywords", []): k = keyword.strip().lower() if k not in keyword_index: keyword_index[k] = [] if entry["file_name"] not in keyword_index[k]: keyword_index[k].append(entry["file_name"]) self._hf_upload_json(keyword_index, Config.MEMORY_INDEX_PATH) log("Keyword index updated.") def recall_from_memory(self, query: str) -> str: log("Recalling from memory...") index = self._hf_download_json(Config.MEMORY_INDEX_PATH, default={}) if not index: return "Memory index is empty or could not be loaded." relevant_keywords_str = self._run_gpt(PromptLibrary.RECALL_MEMORY, 256, prompt=query, keywords=list(index.keys())) try: relevant_keywords = json.loads(relevant_keywords_str) except json.JSONDecodeError: return "Could not determine relevant keywords from memory." if not relevant_keywords: return "Found no relevant information in memory for that query." # Fetch data from relevant files matched_files, fetched_data = set(), [] for k in relevant_keywords: for fname in index.get(k.lower().strip(), []): matched_files.add(fname) for fname in list(matched_files)[:5]: # Limit fetches data = self._hf_download_json(f"{Config.MEMORY_DATA_PATH}/{fname}", default={}) fetched_data.append(data) return f"Recalled {len(fetched_data)} entries from memory:\n\n{json.dumps(fetched_data, indent=2)}" # --- GRADIO APPLICATION --- class GradioApp: def __init__(self, engine: MaestroEngine): self.engine = engine self.app = self._build_ui() def _build_ui(self) -> gr.Blocks: with gr.Blocks(theme=gr.themes.Soft(primary_hue="sky"), title="Maestro AI Engine") as app: session_id = gr.State(lambda: secrets.token_hex(16)) gr.Markdown("# 🧠 Maestro: AI Data Engine & Synthesis Platform") with gr.Tabs(): with gr.TabItem("⚙️ Ingestion & Synthesis"): with gr.Row(): with gr.Column(scale=3): task_instructions = gr.Textbox(label="Primary Task / Instructions", placeholder="e.g., 'Summarize the key findings regarding renewable energy adoption'") with gr.Tabs(): with gr.TabItem("Text Input"): text_input = gr.Textbox(lines=10) with gr.TabItem("File Upload"): file_upload = gr.File(label="Upload Files (.pdf, .txt)", file_count="multiple", type="filepath") with gr.TabItem("Web URL"): url_input = gr.Textbox(label="URL") with gr.TabItem("PDF URL"): pdf_url_input = gr.Textbox(label="Single PDF URL") with gr.TabItem("Batch PDF URLs"): pdf_batch_input = gr.Textbox(label="Comma-separated PDF URLs", lines=3) with gr.Column(scale=1): gr.Markdown("### Processing Options") summarize_check = gr.Checkbox(label="Create JSON Summary", value=True) report_check = gr.Checkbox(label="Generate Full Report (requires summary)", value=False) memory_check = gr.Checkbox(label="Save to Persistent Memory", value=False) process_button = gr.Button("🚀 Process & Synthesize", variant="primary", scale=2) gr.Markdown("### Results") with gr.Row(): final_report_output = gr.Markdown(label="Final Report") json_summary_output = gr.JSON(label="JSON Summaries") with gr.TabItem("🔎 Memory Recall"): memory_query = gr.Textbox(label="Query Persistent Memory", placeholder="e.g., 'What do we know about market trends in 2024?'") recall_button = gr.Button("Recall", variant="primary") memory_output = gr.Textbox(label="Recalled Information", lines=20, interactive=False) # --- CORRECTION PART 1: The event listener now expects 4 outputs --- # The output components match the error: [state, textbox, textbox, button] # In our code, these are: session_id, final_report_output, json_summary_output, process_button process_button.click( self._synthesis_workflow, [session_id, task_instructions, text_input, file_upload, url_input, pdf_url_input, pdf_batch_input, summarize_check, report_check, memory_check], [session_id, final_report_output, json_summary_output, process_button] ) recall_button.click(self.engine.recall_from_memory, [memory_query], [memory_output]) return app # --- CORRECTION PART 2: The handler function is now a generator that yields updates --- def _synthesis_workflow(self, session, task, text, files, url, pdf_url, pdf_batch, do_sum, do_rep, do_mem): log(f"Starting synthesis workflow for session: {session}") # 1. First yield: Immediately update the UI to show a "processing" state. # This provides a value for all 4 output components. yield { session_id: session, # The state component doesn't need to be changed final_report_output: "⚙️ Processing... Please wait.", json_summary_output: None, process_button: gr.update(value="Processing...", interactive=False) } # 2. Perform the actual work ingested_text, errors = self.engine.process_data_sources(text, files, url, pdf_url, pdf_batch) if errors: log(f"Ingestion errors: {errors}") if not ingested_text: # Final yield (or return) in case of error yield { session_id: session, final_report_output: "## Error\nNo data was successfully ingested. Please check your inputs.", json_summary_output: {"errors": errors}, process_button: gr.update(value="🚀 Process & Synthesize", interactive=True) } return # Stop execution here if do_mem: self.engine.save_to_memory(ingested_text, task) report_result, summaries_result = "Processing was not requested.", None if do_sum or do_rep: report_result, summaries_result = self.engine.synthesis_workflow(ingested_text, task, do_sum, do_rep) # 3. Final yield: Return the final results and re-enable the button. # This also provides a value for all 4 output components. yield { session_id: session, final_report_output: report_result, json_summary_output: summaries_result, process_button: gr.update(value="🚀 Process & Synthesize", interactive=True) } def launch(self): self.app.launch(debug=Config.VERBOSE, share=False) if __name__ == "__main__ ": if not Config.HF_TOKEN: print("FATAL: HF_TOKEN environment variable not set.") else: log("Instantiating Maestro Engine...") engine = MaestroEngine() app = GradioApp(engine) log("Launching Gradio App...") app.launch()