import requests import json import gradio as gr # from concurrent.futures import ThreadPoolExecutor import pdfplumber import pandas as pd import time from cnocr import CnOcr from sentence_transformers import SentenceTransformer, models, util word_embedding_model = models.Transformer('uer/sbert-base-chinese-nli', do_lower_case=True) pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension(), pooling_mode='cls') embedder = SentenceTransformer(modules=[word_embedding_model, pooling_model]) ocr = CnOcr() # chat_url = 'https://souljoy-my-api.hf.space/sale' chat_url = 'https://souljoy-my-api.hf.space/chatpdf' headers = { 'Content-Type': 'application/json', } # thread_pool_executor = ThreadPoolExecutor(max_workers=4) history_max_len = 500 all_max_len = 3000 def get_emb(text): emb_url = 'https://souljoy-my-api.hf.space/embeddings' data = {"content": text} try: result = requests.post(url=emb_url, data=json.dumps(data), headers=headers ) return result.json()['data'][0]['embedding'] except Exception as e: print('data', data, 'result json', result.json()) def doc_emb(doc: str): texts = doc.split('\n') # futures = [] emb_list = embedder.encode(texts) # for text in texts: # futures.append(thread_pool_executor.submit(get_emb, text)) # for f in futures: # emb_list.append(f.result()) print('\n'.join(texts)) return texts, emb_list, gr.Textbox.update(visible=True), gr.Button.update(visible=True), gr.Markdown.update( value="""操作说明 step 3：PDF解析提交成功！ 🙋 可以开始对话啦~"""), gr.Chatbot.update(visible=True) def get_response(msg, bot, doc_text_list, doc_embeddings): # future = thread_pool_executor.submit(get_emb, msg) now_len = len(msg) req_json = {'question': msg} his_bg = -1 for i in range(len(bot) - 1, -1, -1): if now_len + len(bot[i][0]) + len(bot[i][1]) > history_max_len: break now_len += len(bot[i][0]) + len(bot[i][1]) his_bg = i req_json['history'] = [] if his_bg == -1 else bot[his_bg:] # query_embedding = future.result() query_embedding = embedder.encode([msg]) cos_scores = util.cos_sim(query_embedding, doc_embeddings)[0] score_index = [[score, index] for score, index in zip(cos_scores, [i for i in range(len(cos_scores))])] score_index.sort(key=lambda x: x[0], reverse=True) print('score_index:\n', score_index) index_set, sub_doc_list = set(), [] for s_i in score_index: doc = doc_text_list[s_i[1]] if now_len + len(doc) > all_max_len: break index_set.add(s_i[1]) now_len += len(doc) # 可能段落截断错误，所以把上下段也加入进来 if s_i[1] > 0 and s_i[1] -1 not in index_set: doc = doc_text_list[s_i[1]-1] if now_len + len(doc) > all_max_len: break index_set.add(s_i[1]-1) now_len += len(doc) if s_i[1] + 1 < len(doc_text_list) and s_i[1] + 1 not in index_set: doc = doc_text_list[s_i[1]+1] if now_len + len(doc) > all_max_len: break index_set.add(s_i[1]+1) now_len += len(doc) index_list = list(index_set) index_list.sort() for i in index_list: sub_doc_list.append(doc_text_list[i]) req_json['doc'] = '' if len(sub_doc_list) == 0 else '\n'.join(sub_doc_list) data = {"content": json.dumps(req_json)} print('data:\n', req_json) result = requests.post(url=chat_url, data=json.dumps(data), headers=headers ) res = result.json()['content'] bot.append([msg, res]) return bot[max(0, len(bot) - 3):] def up_file(files): doc_text_list = [] for idx, file in enumerate(files): print(file.name) with pdfplumber.open(file.name) as pdf: for i in range(len(pdf.pages)): # 读取PDF文档第i+1页 page = pdf.pages[i] res_list = page.extract_text().split('\n')[:-1] for j in range(len(page.images)): # 获取图片的二进制流 img = page.images[j] file_name = '{}-{}-{}.png'.format(str(time.time()), str(i), str(j)) with open(file_name, mode='wb') as f: f.write(img['stream'].get_data()) try: res = ocr.ocr(file_name) except Exception as e: res = [] if len(res) > 0: res_list.append(' '.join([re['text'] for re in res])) tables = page.extract_tables() for table in tables: # 第一列当成表头： df = pd.DataFrame(table[1:], columns=table[0]) try: records = json.loads(df.to_json(orient="records", force_ascii=False)) for rec in records: res_list.append(json.dumps(rec, ensure_ascii=False)) except Exception as e: res_list.append(str(df)) doc_text_list += res_list doc_text_list = [str(text).strip() for text in doc_text_list if len(str(text).strip()) > 0] print(doc_text_list) return gr.Textbox.update(value='\n'.join(doc_text_list), visible=True), gr.Button.update( visible=True), gr.Markdown.update( value="操作说明 step 2：确认PDF解析结果（可修正），点击“提交解析结果”，随后进行对话") with gr.Blocks() as demo: with gr.Row(): with gr.Column(): file = gr.File(file_types=['.pdf'], label='点击上传PDF，进行解析(支持多文档、表格、OCR)', file_count='multiple') doc_bu = gr.Button(value='提交解析结果', visible=False) txt = gr.Textbox(label='PDF解析结果', visible=False) doc_text_state = gr.State([]) doc_emb_state = gr.State([]) with gr.Column(): md = gr.Markdown("""操作说明 step 1：点击左侧区域，上传PDF，进行解析""") chat_bot = gr.Chatbot(visible=False) msg_txt = gr.Textbox(label='消息框', placeholder='输入消息，点击发送', visible=False) chat_bu = gr.Button(value='发送', visible=False) file.change(up_file, [file], [txt, doc_bu, md]) doc_bu.click(doc_emb, [txt], [doc_text_state, doc_emb_state, msg_txt, chat_bu, md, chat_bot]) chat_bu.click(get_response, [msg_txt, chat_bot, doc_text_state, doc_emb_state], [chat_bot]) if __name__ == "__main__": demo.queue().launch() # demo.queue().launch(share=False, server_name='172.22.2.54', server_port=9191)