import asyncio import logging import nltk import string import warnings import io import traceback import zipfile import json import os import requests import subprocess import pandas as pd import re from lxml import etree from typing import Literal from dotenv import load_dotenv from nltk.tokenize import word_tokenize from bs4 import BeautifulSoup from nltk.corpus import stopwords from nltk.stem import WordNetLemmatizer from fastapi import FastAPI, BackgroundTasks, HTTPException, Request from fastapi.staticfiles import StaticFiles from schemas import * from fastapi.middleware.cors import CORSMiddleware from fastapi.responses import FileResponse, StreamingResponse from litellm.router import Router from aiolimiter import AsyncLimiter load_dotenv() logging.basicConfig( level=logging.DEBUG, format='[%(asctime)s][%(levelname)s][%(filename)s:%(lineno)d]: %(message)s', datefmt='%Y-%m-%d %H:%M:%S' ) # Download required packages for NLTK nltk.download('stopwords') nltk.download('punkt_tab') nltk.download('wordnet') warnings.filterwarnings("ignore") app = FastAPI(title="Requirements Extractor") app.mount("/static", StaticFiles(directory="static"), name="static") app.add_middleware(CORSMiddleware, allow_credentials=True, allow_headers=[ "*"], allow_methods=["*"], allow_origins=["*"]) llm_router = Router(model_list=[ { "model_name": "gemini-v1", "litellm_params": { "model": "gemini/gemini-2.0-flash", "api_key": os.environ.get("GEMINI"), "max_retries": 10, "rpm": 15, "allowed_fails": 1, "cooldown": 30, } }, { "model_name": "gemini-v2", "litellm_params": { "model": "gemini/gemini-2.5-flash", "api_key": os.environ.get("GEMINI"), "max_retries": 10, "rpm": 10, "allowed_fails": 1, "cooldown": 30, } }], fallbacks=[{"gemini-v2": ["gemini-v1"]}], num_retries=10, retry_after=30) limiter_mapping = { model["model_name"]: AsyncLimiter(model["litellm_params"]["rpm"], 60) for model in llm_router.model_list } lemmatizer = WordNetLemmatizer() NSMAP = { 'w': 'http://schemas.openxmlformats.org/wordprocessingml/2006/main', 'v': 'urn:schemas-microsoft-com:vml' } def lemma(text: str): stop_words = set(stopwords.words('english')) txt = text.translate(str.maketrans('', '', string.punctuation)).strip() tokens = [token for token in word_tokenize( txt.lower()) if token not in stop_words] return [lemmatizer.lemmatize(token) for token in tokens] def get_docx_archive(url: str) -> zipfile.ZipFile: """Récupère le docx depuis l'URL et le retourne comme objet ZipFile""" if not url.endswith("zip"): raise ValueError("URL doit pointer vers un fichier ZIP") doc_id = os.path.splitext(os.path.basename(url))[0] resp = requests.get(url, verify=False, headers={ "User-Agent": 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36' }) resp.raise_for_status() with zipfile.ZipFile(io.BytesIO(resp.content)) as zf: for file_name in zf.namelist(): if file_name.endswith(".docx"): docx_bytes = zf.read(file_name) return zipfile.ZipFile(io.BytesIO(docx_bytes)) elif file_name.endswith(".doc"): input_path = f"/tmp/{doc_id}.doc" output_path = f"/tmp/{doc_id}.docx" docx_bytes = zf.read(file_name) with open(input_path, "wb") as f: f.write(docx_bytes) subprocess.run([ "libreoffice", "--headless", "--convert-to", "docx", "--outdir", "/tmp", input_path ], check=True) with open(output_path, "rb") as f: docx_bytes = f.read() os.remove(input_path) os.remove(output_path) return zipfile.ZipFile(io.BytesIO(docx_bytes)) raise ValueError("Aucun fichier docx/doc trouvé dans l'archive") def parse_document_xml(docx_zip: zipfile.ZipFile) -> etree._ElementTree: """Parse le document.xml principal""" xml_bytes = docx_zip.read('word/document.xml') parser = etree.XMLParser(remove_blank_text=True) return etree.fromstring(xml_bytes, parser=parser) def clean_document_xml(root: etree._Element) -> None: """Nettoie le XML en modifiant l'arbre directement""" # Suppression des balises et leur contenu for del_elem in root.xpath('//w:del', namespaces=NSMAP): parent = del_elem.getparent() if parent is not None: parent.remove(del_elem) # Désencapsulation des balises for ins_elem in root.xpath('//w:ins', namespaces=NSMAP): parent = ins_elem.getparent() index = parent.index(ins_elem) for child in ins_elem.iterchildren(): parent.insert(index, child) index += 1 parent.remove(ins_elem) # Nettoyage des commentaires for tag in ['w:commentRangeStart', 'w:commentRangeEnd', 'w:commentReference']: for elem in root.xpath(f'//{tag}', namespaces=NSMAP): parent = elem.getparent() if parent is not None: parent.remove(elem) def create_modified_docx(original_zip: zipfile.ZipFile, modified_root: etree._Element) -> bytes: """Crée un nouveau docx avec le XML modifié""" output = io.BytesIO() with zipfile.ZipFile(output, 'w', compression=zipfile.ZIP_DEFLATED) as new_zip: # Copier tous les fichiers non modifiés for file in original_zip.infolist(): if file.filename != 'word/document.xml': new_zip.writestr(file, original_zip.read(file.filename)) # Ajouter le document.xml modifié xml_str = etree.tostring( modified_root, xml_declaration=True, encoding='UTF-8', pretty_print=True ) new_zip.writestr('word/document.xml', xml_str) output.seek(0) return output.getvalue() def docx_to_txt(doc_id: str, url: str): docx_zip = get_docx_archive(url) root = parse_document_xml(docx_zip) clean_document_xml(root) modified_bytes = create_modified_docx(docx_zip, root) input_path = f"/tmp/{doc_id}_cleaned.docx" output_path = f"/tmp/{doc_id}_cleaned.txt" with open(input_path, "wb") as f: f.write(modified_bytes) subprocess.run([ "libreoffice", "--headless", "--convert-to", "txt", "--outdir", "/tmp", input_path ], check=True) with open(output_path, "r", encoding="utf-8") as f: txt_data = [line.strip() for line in f if line.strip()] os.remove(input_path) os.remove(output_path) return txt_data @app.get("/") def render_page(): return FileResponse("index.html") @app.post("/get_meetings", response_model=MeetingsResponse) def get_meetings(req: MeetingsRequest): working_group = req.working_group tsg = re.sub(r"\d+", "", working_group) wg_number = re.search(r"\d", working_group).group(0) logging.debug(tsg, wg_number) url = "https://www.3gpp.org/ftp/tsg_" + tsg logging.debug(url) resp = requests.get(url, verify=False) soup = BeautifulSoup(resp.text, "html.parser") meeting_folders = [] all_meetings = [] wg_folders = [item.get_text() for item in soup.select("tr td a")] selected_folder = None for folder in wg_folders: if "wg" + str(wg_number) in folder.lower(): selected_folder = folder break url += "/" + selected_folder logging.debug(url) if selected_folder: resp = requests.get(url, verify=False) soup = BeautifulSoup(resp.text, "html.parser") meeting_folders = [item.get_text() for item in soup.select("tr td a") if item.get_text( ).startswith("TSG") or (item.get_text().startswith("CT") and "-" in item.get_text())] all_meetings = [working_group + "#" + meeting.split("_", 1)[1].replace("_", " ").replace( "-", " ") if meeting.startswith('TSG') else meeting.replace("-", "#") for meeting in meeting_folders] return MeetingsResponse(meetings=dict(zip(all_meetings, meeting_folders))) # ============================================================================================================================================ @app.post("/get_dataframe", response_model=DataResponse) def get_change_request_dataframe(req: DataRequest): working_group = req.working_group tsg = re.sub(r"\d+", "", working_group) wg_number = re.search(r"\d", working_group).group(0) url = "https://www.3gpp.org/ftp/tsg_" + tsg logging.info("Fetching TDocs dataframe") resp = requests.get(url, verify=False) soup = BeautifulSoup(resp.text, "html.parser") wg_folders = [item.get_text() for item in soup.select("tr td a")] selected_folder = None for folder in wg_folders: if "wg" + str(wg_number) in folder.lower(): selected_folder = folder break url += "/" + selected_folder + "/" + req.meeting + "/docs" resp = requests.get(url, verify=False) soup = BeautifulSoup(resp.text, "html.parser") files = [item.get_text() for item in soup.select("tr td a") if item.get_text().endswith(".xlsx")] def gen_url(tdoc: str): return f"{url}/{tdoc}.zip" df = pd.read_excel(str(url + "/" + files[0]).replace("#", "%23")) filtered_df = df[(((df["Type"] == "CR") & ((df["CR category"] == "B") | (df["CR category"] == "C"))) | (df["Type"] == "pCR")) & ~( df["Uploaded"].isna())][["TDoc", "Title", "CR category", "Source", "Type", "Agenda item", "Agenda item description", "TDoc Status"]] filtered_df["URL"] = filtered_df["TDoc"].apply(gen_url) df = filtered_df.fillna("") return DataResponse(data=df[["TDoc", "Title", "Type", "TDoc Status", "Agenda item description", "URL"]].to_dict(orient="records")) # ================================================================================================================================== @app.post("/download_tdocs") def download_tdocs(req: DownloadRequest): """Download the specified TDocs and zips them in a single archive""" documents = req.documents logging.info(f"Downloading TDocs: {documents}") def process_document(doc: str): doc_id = doc url = requests.post( 'https://organizedprogrammers-3gppdocfinder.hf.space/find', headers={"Content-Type": "application/json"}, data=json.dumps({"doc_id": doc_id}), verify=False ) logging.info( f"Retrieving URL for doc {doc_id} returned http status {url.status_code}") url = url.json()['url'] logging.debug(f"Doc URL for {doc_id} is {url}") try: txt = "\n".join(docx_to_txt(doc_id, url)) except Exception as e: txt = f"Document {doc_id} text extraction failed: {e}" return doc_id, txt.encode("utf-8") # PERF: use asyncio? def process_batch(batch): results = {} for doc in batch: try: doc_id, file_bytes = process_document(doc) results[doc_id] = file_bytes except Exception as e: traceback.print_exception(e) results[doc] = b"Erreur" return results documents_bytes = process_batch(documents) zip_buffer = io.BytesIO() with zipfile.ZipFile(zip_buffer, mode='w', compression=zipfile.ZIP_DEFLATED) as zip_file: for doc_id, txt_data in documents_bytes.items(): zip_file.writestr(f'{doc_id}.txt', txt_data) zip_buffer.seek(0) return StreamingResponse( zip_buffer, media_type="application/zip" ) # ======================================================================================================================== @app.post("/generate_requirements", response_model=RequirementsResponse) async def gen_reqs(req: RequirementsRequest, background_tasks: BackgroundTasks): """Extract requirements from the specified TDocs using a LLM""" documents = req.documents n_docs = len(documents) logging.info("Generating requirements for documents: {}".format( [doc.document for doc in documents])) def prompt(doc_id, full): return f"Here's the document whose ID is {doc_id} : {full}\n\nExtract all requirements and group them by context, returning a list of objects where each object includes a document ID, a concise description of the context where the requirements apply (not a chapter title or copied text), and a list of associated requirements; always return the result as a list, even if only one context is found. Remove the errors" async def process_document(doc): doc_id = doc.document url = doc.url try: full = "\n".join(docx_to_txt(doc_id, url)) except Exception as e: traceback.print_exception(e) return RequirementsResponse(requirements=[DocRequirements(document=doc_id, context="Error LLM", requirements=[])]).requirements try: model_used = "gemini-v2" # À adapter si fallback activé async with limiter_mapping[model_used]: resp_ai = await llm_router.acompletion( model=model_used, messages=[ {"role": "user", "content": prompt(doc_id, full)}], response_format=RequirementsResponse ) return RequirementsResponse.model_validate_json(resp_ai.choices[0].message.content).requirements except Exception as e: if "rate limit" in str(e).lower(): try: model_used = "gemini-v2" # À adapter si fallback activé async with limiter_mapping[model_used]: resp_ai = await llm_router.acompletion( model=model_used, messages=[ {"role": "user", "content": prompt(doc_id, full)}], response_format=RequirementsResponse ) return RequirementsResponse.model_validate_json(resp_ai.choices[0].message.content).requirements except Exception as fallback_e: traceback.print_exception(fallback_e) return RequirementsResponse(requirements=[DocRequirements(document=doc_id, context="Error LLM", requirements=[])]).requirements else: traceback.print_exception(e) return RequirementsResponse(requirements=[DocRequirements(document=doc_id, context="Error LLM", requirements=[])]).requirements async def process_batch(batch): results = await asyncio.gather(*(process_document(doc) for doc in batch)) return [item for sublist in results for item in sublist] all_requirements = [] if n_docs <= 30: batch_results = await process_batch(documents) all_requirements.extend(batch_results) else: batch_size = 30 batches = [documents[i:i + batch_size] for i in range(0, n_docs, batch_size)] for i, batch in enumerate(batches): batch_results = await process_batch(batch) all_requirements.extend(batch_results) if i < len(batches) - 1: background_tasks.add_task(asyncio.sleep, 60) return RequirementsResponse(requirements=all_requirements) # ====================================================================================================================================================================================== class ProgressUpdate(BaseModel): """Defines the structure of a single SSE message.""" status: Literal["progress", "complete"] data: dict total_docs: int processed_docs: int @app.post("/generate_requirements/sse") async def gen_reqs(req: RequirementsRequest, con: Request): """Extract requirements from the specified TDocs using a LLM and returns SSE events about the progress of ongoing operations""" documents = req.documents n_docs = len(documents) logging.info("Generating requirements for documents: {}".format( [doc.document for doc in documents])) def prompt(doc_id, full): return f"Here's the document whose ID is {doc_id} : {full}\n\nExtract all requirements and group them by context, returning a list of objects where each object includes a document ID, a concise description of the context where the requirements apply (not a chapter title or copied text), and a list of associated requirements; always return the result as a list, even if only one context is found. Remove the errors" async def _process_document(doc) -> list[DocRequirements]: doc_id = doc.document url = doc.url # convert the docx to txt for use try: full = "\n".join(docx_to_txt(doc_id, url)) except Exception as e: traceback.print_exception(e) return [DocRequirements(document=doc_id, context="Error LLM", requirements=[])] try: model_used = "gemini-v2" resp_ai = await llm_router.acompletion( model=model_used, messages=[ {"role": "user", "content": prompt(doc_id, full)}], response_format=RequirementsResponse ) return RequirementsResponse.model_validate_json(resp_ai.choices[0].message.content).requirements except Exception as e: return [DocRequirements(document=doc_id, context="Error LLM", requirements=[])] # futures for all processed documents process_futures = [_process_document(doc) for doc in documents] # lambda to print progress def progress_update(x): return f"data: {x.model_dump_json()}\n\n" # async generator that generates the SSE events for progress async def _stream_generator(docs: list[asyncio.Future]): items = [] n_processed = 0 yield progress_update(ProgressUpdate(status="progress", data={}, total_docs=n_docs, processed_docs=0)) for doc in asyncio.as_completed(docs): result = await doc items.extend(result) n_processed += 1 yield progress_update(ProgressUpdate(status="progress", data={}, total_docs=n_docs, processed_docs=n_processed)) final_response = RequirementsResponse(requirements=items) yield progress_update(ProgressUpdate(status="complete", data=final_response.model_dump(), total_docs=n_docs, processed_docs=n_processed)) return StreamingResponse(_stream_generator(process_futures), media_type="text/event-stream") # ======================================================================================================================================================================================= @app.post("/get_reqs_from_query", response_model=ReqSearchResponse) def find_requirements_from_problem_description(req: ReqSearchRequest): requirements = req.requirements query = req.query requirements_text = "\n".join( [f"[Selection ID: {r.req_id} | Document: {r.document} | Context: {r.context} | Requirement: {r.requirement}]" for r in requirements]) print("Called the LLM") resp_ai = llm_router.completion( model="gemini-v2", messages=[{"role": "user", "content": f"Given all the requirements : \n {requirements_text} \n and the problem description \"{query}\", return a list of 'Selection ID' for the most relevant corresponding requirements that reference or best cover the problem. If none of the requirements covers the problem, simply return an empty list"}], response_format=ReqSearchLLMResponse ) print("Answered") print(resp_ai.choices[0].message.content) out_llm = ReqSearchLLMResponse.model_validate_json( resp_ai.choices[0].message.content).selected if max(out_llm) > len(requirements) - 1: raise HTTPException( status_code=500, detail="LLM error : Generated a wrong index, please try again.") return ReqSearchResponse(requirements=[requirements[i] for i in out_llm])