import os import numpy as np import openai import faiss from transformers import BertTokenizer, BertModel import torch import json import time import warnings import copy import pickle import random import torch.nn.functional as F seed_value = 42 random.seed(seed_value) np.random.seed(seed_value) torch.manual_seed(seed_value) warnings.filterwarnings("ignore") os.environ['KMP_DUPLICATE_LIB_OK'] = 'True' KEY = os.environ['API_KEY'] openai.api_base = 'https://api.together.xyz' llm_model = "mistralai/Mixtral-8x7B-Instruct-v0.1" tokenizer = BertTokenizer.from_pretrained('facebook/contriever') model = BertModel.from_pretrained('facebook/contriever').to(torch.device("cpu")) import datetime import json import arxiv def summarize_research_direction(personal_info): prompt_qa = ( "Based on the list of the researcher's first person persona from different times, please write a comprehensive first person persona. Focus more on more rescent personas. Be concise and clear (around 300 words)." "Here are the personas from different times: {peronalinfo}" ) openai.api_key = KEY input = {} input['peronalinfo'] = personal_info prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed = 42, top_p=0) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed = 42, top_p=0) content = completion.choices[0].message["content"] return content def get_authors(authors, first_author = False): output = str() if first_author == False: output = ", ".join(str(author) for author in authors) else: output = authors[0] return output def sort_papers(papers): output = dict() keys = list(papers.keys()) keys.sort(reverse=True) for key in keys: output[key] = papers[key] return output def get_daily_papers(topic,query="slam", max_results=300): """ @param topic: str @param query: str @return paper_with_code: dict """ # output content = dict() Info = dict() search_engine = arxiv.Search( query = query, max_results = max_results, sort_by = arxiv.SortCriterion.SubmittedDate ) newest_day = None # cnt = 0 for result in search_engine.results(): # paper_id = result.get_short_id() paper_title = result.title paper_url = result.entry_id # paper_abstract = result.summary paper_abstract = result.summary.replace("\n"," ") publish_time = result.published.date() if newest_day is not None and not(newest_day == publish_time): break elif newest_day is None: newest_day = publish_time if publish_time in content: content[publish_time]['abstract'].append(paper_title+ ": "+paper_abstract) content[publish_time]['info'].append(paper_title+": "+paper_url) # Info[publish_time].append(paper_title+": "+paper_url) else: content[publish_time] = {} content[publish_time]['abstract'] = [paper_title+ ": "+paper_abstract] content[publish_time]['info'] = [paper_title+": "+paper_url] # cnt = cnt + 1 # content[publish_time] = [paper_abstract] # Info[publish_time] = # print(publish_time) # content[paper_key] = f"|**{publish_time}**|**{paper_title}**|{paper_first_author} et.al.|[{paper_id}]({paper_url})|\n" data = content # print(cnt) return data, newest_day def papertitleAndLink(dataset): formatted_papers = [] i = 0 # import pdb # pdb.set_trace() for title in dataset: # import pdb # pdb.set_trace() i = i +1 formatted_papers.append("[%d] "%i + title) # i = 0 # formatted_papers = [f"{"[%d]"%i + papers}" i = i + 1 for k in dataset.keys() for papers in dataset[k]['info']] return ';\n'.join(formatted_papers) def paperinfo(dataset): # for k in dataset.keys(): formatted_papers = [f"{paper}" for k in dataset.keys() for paper in dataset[k]['abstract']] return '; '.join(formatted_papers) def generate_ideas (trend): # prompt_qa = ( # "Now you are a researcher with this background {profile}, and here is a high-level summarized trend of a research field {trend}." # "How do you view this field? Do you have any novel ideas or insights?" # ) prompt_qa = ( "Here is a high-level summarized trend of a research field: {trend}." "How do you view this field? Do you have any novel ideas or insights?" "Please give me 3 to 5 novel ideas and insights in bullet points. Each bullet points should be concise, containing 2 or 3 sentences." ) openai.api_key = KEY content_l = [] input = {} # input['profile'] = profile input['trend'] = trend prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed = 42, top_p=0) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed = 42, top_p=0) content = completion.choices[0].message["content"] content_l.append(content) return content_l def summarize_research_field(profile, keywords, dataset,data_embedding): # papers = paperinfo(dataset) query_input = {} input = {} if profile is None: prompt_qa = ( "Given some recent paper titles and abstracts. Could you summarize no more than 10 top keywords of high level research backgounds and trends." # "Here are the keywords: {keywords}" "Here are the retrieved paper abstracts: {papers}" ) query_format = ( "Given the keywords, retrieve some recent paper titles and abstracts can represent research trends in this field." "Here are the keywords: {keywords}" ) # input['keywords'] = keywords query_input['keywords'] = keywords else: prompt_qa = ( "Given some recent paper titles and abstracts. Could you summarize no more than 10 top keywords of high level research backgounds and trends." # "Here is my profile: {profile}" # "Here are the keywords: {keywords}" "Here are the retrieved paper abstracts: {papers}" ) query_format = ( "Given the profile of me, retrieve some recent paper titles and abstracts can represent research trends related to my profile." "Here is my profile: {profile}" # "Here are the keywords: {keywords}" ) query_input['profile'] = profile # import pdb # pdb.set_trace() openai.api_key = KEY content_l = [] query = query_format.format_map(query_input) query_embedding=get_bert_embedding([query]) # text_chunk_l = dataset text_chunk_l = [] data_embedding_l=[] # with open(dataset_path, 'r', encoding='utf-8') as file: # dataset = json.load(file) title_chunk = [] for k in dataset.keys(): # import pdb # pdb.set_trace() title_chunk.extend(dataset[k]['info']) text_chunk_l.extend(dataset[k]['abstract']) data_embedding_l.extend(data_embedding[k]) # import pdb # pdb.set_trace() # print(dataset[k]['info']) # [p if 'graph' in p else "" for p in dataset[k]['info']] chunks_embedding_text_all = data_embedding_l ch_text_chunk=copy.copy(text_chunk_l) ch_text_chunk_embed=copy.copy(chunks_embedding_text_all) num_chunk = 10 # print("raw_chunk_length: ", raw_chunk_length) neib_all = neiborhood_search(ch_text_chunk_embed, query_embedding, num_chunk) neib_all=neib_all.reshape(-1) context = [] retrieve_paper = [] for i in neib_all: context.append(ch_text_chunk[i]) # if i not in retrieve_paper: retrieve_paper.append(title_chunk[i]) # import pdb # pdb.set_trace() input['papers'] = '; '.join(context) prompt = prompt_qa.format_map(input) # import pdb # pdb.set_trace() # import pdb # pdb.set_trace() try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], max_tokens=512) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], max_tokens= 512) content = completion.choices[0].message["content"] content_l.append(content) return content_l, retrieve_paper def update_json_file(filename,data_all): with open(filename,"r") as f: content = f.read() if not content: m = {} else: m = json.loads(content) json_data = m.copy() # update papers in each keywords for data in data_all: for time in data.keys(): papers = data[time] # print(papers.published) json_data[time.strftime("%m/%d/%Y")] = papers for time in json_data.keys(): papers = json_data[time] papers['ch_abs']=copy.deepcopy(papers['abstract']) # print(papers.published) json_data[time] = papers with open(filename,"w") as f_: json.dump(json_data,f_) return json_data def update_pickle_file(filename, data_all): # if os.path.exists(filename): # with open(filename,"rb") as f: # m = pickle.loads(f) # with open(filename,"rb") as f: # content = f.read() # if not content: # m = {} # else: # m = json.load(content) with open(filename, "rb") as file: m = pickle.load(file) # json_data = m.copy() # else: # with open(filename, mode='wb', encoding='utf-8') as ff: # m = {} # with open(filename, "rb") as file: # m = pickle.load(file) pickle_data = m.copy() for time in data_all.keys(): embeddings = data_all[time] pickle_data[time] =embeddings with open(filename, "wb") as f: pickle.dump(pickle_data, f) return pickle_data def json_to_md(filename): """ @param filename: str @return None """ DateNow = datetime.date.today() DateNow = str(DateNow) DateNow = DateNow.replace('-','.') with open(filename,"r") as f: content = f.read() if not content: data = {} else: data = json.loads(content) md_filename = "README.md" # clean README.md if daily already exist else create it with open(md_filename,"w+") as f: pass # write data into README.md with open(md_filename,"a+") as f: f.write("## Updated on " + DateNow + "\n\n") for keyword in data.keys(): day_content = data[keyword] if not day_content: continue # the head of each part f.write(f"## {keyword}\n\n") f.write("|Publish Date|Title|Authors|PDF|\n" + "|---|---|---|---|\n") # sort papers by date day_content = sort_papers(day_content) for _,v in day_content.items(): if v is not None: f.write(v) f.write(f"\n") print("finished") def neiborhood_search(corpus_data, query_data, num=8): d = 768 # dimension neiborhood_num = num xq = torch.cat(query_data, 0).cpu().numpy() xb = torch.cat(corpus_data, 0).cpu().numpy() index = faiss.IndexFlatIP(d) xq = xq.astype('float32') xb = xb.astype('float32') faiss.normalize_L2(xq) faiss.normalize_L2(xb) index.add(xb) # add vectors to the index D, I = index.search(xq, neiborhood_num) return I def get_passage_conclusion_through_LLM(text, question): # prompt_qa = ("Given text:{context},given question:{question},based on this text and question, summarize the above text into a passage so that it can best answer this question.") prompt_qa = ( "Given text:{context},based on this text, summarize the above text into a passage that cannot change its original meaning.") openai.api_key = KEY input = {} input['context'] = text input['question'] = question prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed = 42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed =42) content = completion.choices[0].message["content"] # print(content) return content def retain_useful_info(text, question): prompt_qa = ( "Given text:{context},given question:{question},based on this text and question, summarize the text into a sentence that is most useful in answering this question.") openai.api_key = KEY input = {} input['context'] = text input['question'] = question prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}]) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}]) content = completion.choices[0].message["content"] # print(content) return content def llm_summary(text_l): # prompt_qa = ("Given text:{context},given question:{question},based on this text and question, summarize the above text into a passage so that it can best answer this question.") text = '' for inter in text_l: text += inter prompt_qa = ( "Given text:{context},based on this text, summarize the above text into a fluent passage that cannot change its original meaning.") openai.api_key = KEY input = {} input['context'] = text prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed =42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed=42) content = completion.choices[0].message["content"] # print(content) return content def get_multi_query_through_LLM(question_data, generated_answers=None, support_material=None): PROMPT_DICT = { "without_answer": ( "The input will be a paragraph of text." "Your task is to generate five as diverse, informative, and relevant, as possible versions of supporting materials, perspectives, fact. Provide these alternative materials, perspectives, fact. Each of them occupies a line." "Original text: {question}" "Answer:,Please output a list to split these five answers."), "with_answer": ( "The input will be a paragraph of original text, a previously generated support material and a response for the text based on reviously generated support material by a naive agent, who may make mistakes." "Your task is to generate five as diverse, informative, and relevant, as possible versions of supporting materials,perspectives, fact based on the the above information. Each of them occupies a line." "Provide these alternative materials, perspectives, fact." "Original text:{question}. " "Previously generated support material (the text below are naive, and could be wrong, use with caution): {support_material} " "Response:{answer}." "Answer:,Please output a list to split these five answers."), } prompt_q, prompt_qa = PROMPT_DICT["without_answer"], PROMPT_DICT["with_answer"] openai.api_key = KEY ### question_data inter = {} inter['question'] = question_data if generated_answers != None: inter['answer'] = generated_answers inter['support_material'] = support_material prompt = [prompt_qa.format_map(example) for example in [inter]] else: prompt = [prompt_q.format_map(example) for example in [inter]] try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt[0]}], temperature=0.6, seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt[0]}], temperature=0.6,seed =42) content = completion.choices[0].message["content"] for inter_ in content: inter_ = inter_.strip('1.').strip('2.').strip('3.').strip('4.').strip('5.') # print(content) return content def get_question_through_LLM(question, context): prompt_s = question[0] for i in range(len(context)): prompt_s += "Documents %d: " % (i + 1) + context[i] + '\n' prompt_qa = (prompt_s) openai.api_key = KEY content_l = [] # import pdb # pdb.set_trace() # for inter1 in range(len(context)): # question_i = question[0] # context_i=context[inter1] # input={} # input['question']=question_i # input['context']=context_i prompt = prompt_qa try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6, seed=42) content = completion.choices[0].message["content"] content_l.append(content) # print(content) return content_l def get_response_through_LLM(question, context): prompt_qa = ("Given text: {context}, based on this text, answer the question: {question}") openai.api_key = KEY content_l = [] # print(len(context)) # import pdb # pdb.set_trace() # print() for inter1 in range(len(question)): question_i = question[inter1] context_i = context[inter1] input = {} input['question'] = question_i input['context'] = context_i prompt = prompt_qa.format_map(input) # print(prompt) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) content = completion.choices[0].message["content"] content_l.append(content) # print("Answer for Pre Queston ", inter1, ": ") # print(content,"\n") return content_l def get_response_through_LLM_answer(question, context, profile): # import pdb # pdb.set_trace() if profile is None: prompt_qa = ( "Answer the: {question}, based on materials: {context}" ) else: prompt_qa = ( "Answer the: {question}, based on materials: {context} and my profile: {profile}" ) openai.api_key = KEY content_l = [] # print(len(context)) # import pdb # pdb.set_trace() # print() # print("Length of the question: ", len(question)) # print("Length of the context: ", len(context)) for inter1 in range(len(question)): question_i = question[inter1] context_i = context[inter1] input = {} input['question'] = question_i input['context'] = context_i if profile is not None: profile_i = profile input['profile'] = profile_i # import pdb # pdb.set_trace() prompt = prompt_qa.format_map(input) # print(prompt) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) content = completion.choices[0].message["content"] content_l.append(content) # print(content) return content_l def get_response_through_LLM_cross(question, context): prompt_s = context + '\n' prompt_s += "Based on the above documents, answer the question: {question} in short." prompt_qa = (prompt_s) openai.api_key = KEY content_l = [] for inter1 in range(len(question)): question_i = question[inter1] input = {} input['question'] = question_i prompt = prompt_qa.format_map(input) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) content = completion.choices[0].message["content"] content_l.append(content) # print(content) return content_l def get_bert_embedding(instructions): # encoded_input_all = [tokenizer(text['instruction']+text['input'], return_tensors='pt').to(torch.device("cuda")) for text in instructions] encoded_input_all = [tokenizer(text, return_tensors='pt', truncation=True, max_length=512).to(torch.device("cpu")) for text in instructions] with torch.no_grad(): emb_list = [] for inter in encoded_input_all: emb = model(**inter) emb_list.append(emb['last_hidden_state'].mean(1)) return emb_list def calculate_similarity(tensor_list, input_tensor): flattened_list = [t.flatten() for t in tensor_list] flattened_tensor = input_tensor.flatten() cosine_similarities = [F.cosine_similarity(flattened_tensor.unsqueeze(0), t.unsqueeze(0)) for t in flattened_list] return cosine_similarities def response_verify(question, context, verify = False): if verify: prompt_qa = ( "Input: Given question:{question}, given answer:{context}. Based on the provided question and its corresponding answer, perform the following steps:" "Step 1: Determine if the answer is an actual answer or if it merely indicates that the question cannot be answered due to insufficient information. If the latter is true, just output 'idk' without any extra words " "Step 2: If it is a valid answer, succinctly summarize both the question and answer into a coherent knowledge point, forming a fluent passage." ) else: prompt_qa = ( "Given question:{question},given answer:{context},based on the given question and corresponding answer, " "summarize them into a knowledge point like a fluent passage.") openai.api_key = KEY content_l = [] for inter1 in range(len(question)): question_i = question[inter1] context_i = context[inter1] input = {} input['question'] = question_i input['context'] = context_i prompt = prompt_qa.format_map(input) # print(prompt) try: completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) except: time.sleep(20) completion = openai.ChatCompletion.create( model=llm_model, messages=[ {"role": "user", "content": prompt}], temperature=0.6,seed=42) content = completion.choices[0].message["content"] content_l.append(content) # print(content) return content_l