#%% from tiktoken import get_encoding, encoding_for_model from weaviate_interface import WeaviateClient, WhereFilter from sentence_transformers import SentenceTransformer from prompt_templates import question_answering_prompt_series, question_answering_system from openai_interface import GPT_Turbo from app_features import (convert_seconds, generate_prompt_series, search_result, validate_token_threshold, load_content_cache, load_data, expand_content) from retrieval_evaluation import execute_evaluation, calc_hit_rate_scores from llama_index.finetuning import EmbeddingQAFinetuneDataset from openai import BadRequestError from reranker import ReRanker from loguru import logger import streamlit as st from streamlit_option_menu import option_menu import hydralit_components as hc import sys import json import os, time, requests, re from datetime import timedelta import pathlib import gdown import tempfile import base64 import shutil def get_base64_of_bin_file(bin_file): with open(bin_file, 'rb') as file: data = file.read() return base64.b64encode(data).decode() from dotenv import load_dotenv, find_dotenv load_dotenv(find_dotenv('env'), override=True) # I use a key that I increment each time I want to change a text_input if 'key' not in st.session_state: st.session_state.key = 0 # key = st.session_state['key'] if not pathlib.Path('models').exists(): os.mkdir('models') # TODO cache these things but no time left # I put a file local.txt in my desktop models folder to find out if it's running online we_are_online = not pathlib.Path("models/local.txt").exists() we_are_not_online = not we_are_online golden_dataset = EmbeddingQAFinetuneDataset.from_json("data/golden_100.json") ## PAGE CONFIGURATION st.set_page_config(page_title="Ask Impact Theory", page_icon="assets/impact-theory-logo-only.png", layout="wide", initial_sidebar_state="collapsed", menu_items={'Report a bug': "https://www.extremelycoolapp.com/bug"}) image = "https://is2-ssl.mzstatic.com/image/thumb/Music122/v4/bd/34/82/bd348260-314c-5898-26c0-bef2e0388ebe/source/1200x1200bb.png" def add_bg_from_local(image_file): bin_str = get_base64_of_bin_file(image_file) page_bg_img = f''' ''' st.markdown(page_bg_img, unsafe_allow_html=True) ## RERANKER reranker = ReRanker('cross-encoder/ms-marco-MiniLM-L-6-v2') ## ENCODING --> tiktoken library model_ids = ['gpt-3.5-turbo-16k', 'gpt-3.5-turbo-0613'] model_nameGPT = model_ids[1] encoding = encoding_for_model(model_nameGPT) ## DATA data_path = './data/impact_theory_data.json' cache_path = 'data/impact_theory_cache.parquet' data = load_data(data_path) cache = None # load_content_cache(cache_path) guest_list = sorted(list(set([d['guest'] for d in data]))) if 'secrets' in st.secrets: # st.write("Loading secrets from [secrets] section") # for streamlit online or local, which uses a [secrets] section Wapi_key = st.secrets['secrets']['WEAVIATE_API_KEY'] url = st.secrets['secrets']['WEAVIATE_ENDPOINT'] openai_api_key = st.secrets['secrets']['OPENAI_API_KEY'] # hf_token = st.secrets['secrets']['LLAMA2_ENDPOINT_HF_TOKEN'] # hf_endpoint = st.secrets['secrets']['LLAMA2_ENDPOINT'] else : # st.write("Loading secrets for Huggingface") # for Huggingface (no [secrets] section) Wapi_key = st.secrets['WEAVIATE_API_KEY'] url = st.secrets['WEAVIATE_ENDPOINT'] openai_api_key = st.secrets['OPENAI_API_KEY'] # hf_token = st.secrets['LLAMA2_ENDPOINT_HF_TOKEN'] # hf_endpoint = st.secrets['LLAMA2_ENDPOINT'] #%% # model_default = 'sentence-transformers/all-mpnet-base-v2' model_default = 'models/finetuned-all-mpnet-base-v2-300' if we_are_not_online \ else 'sentence-transformers/all-mpnet-base-v2' available_models = ['sentence-transformers/all-mpnet-base-v2', 'sentence-transformers/all-MiniLM-L6-v2', 'models/finetuned-all-mpnet-base-v2-300', 'sentence-transformers/all-MiniLM-L12-v2'] #%% models_urls = {'models/finetuned-all-mpnet-base-v2-300': "https://drive.google.com/drive/folders/1asJ37-AUv5nytLtH6hp6_bVV3_cZOXfj"} def download_model_from_Gdrive(model_name_or_path, model_local_path): st.write("Downloading model from Google Drive") assert model_name_or_path in models_urls, f"Model {model_name_or_path} not found in models_urls" url = models_urls[model_name_or_path] gdown.download_folder(url, output=model_local_path, quiet=False, use_cookies=False) print(f"Model downloaded from Gdrive and saved to {model_local_path} folder") # st.write("Model downloaded") def download_model(model_name_or_path, model_local_path): if model_name_or_path.startswith("models/"): download_model_from_Gdrive(model_name_or_path, model_local_path) elif model_name_or_path.startswith("sentence-transformers/"): st.sidebar.write(f"Downloading {model_name_or_path}") model = SentenceTransformer(model_name_or_path) st.sidebar.write(f"Model {model_name_or_path} downloaded") models_urls[model_name_or_path] = model_local_path model.save(model_local_path) # st.sidebar.write(f"Model {model_name_or_path} saved to {model_new_path}") #%% # for streamlit online, we must download the model from google drive # because github LFS doesn't work on forked repos def check_model(model_name_or_path): model_name = model_name_or_path.split('/')[-1] # remove 'sentence-transformers' model_local_path = str(pathlib.Path("models") / model_name) # this creates a models folder inside /models if pathlib.Path(model_local_path).exists(): # let's use the model that's already there print(f"Model {model_local_path} already exists") else: # let's download the model, HF is not limited in space like Streamlit.io download_model(model_name_or_path, model_local_path) #%% instantiate Weaviate client def get_weaviate_client(api_key, url, model_name_or_path, openai_api_key): try: client = WeaviateClient(api_key, url, model_name_or_path=model_name_or_path, openai_api_key=openai_api_key) except Exception: # client not available, wrong key, expired free sandbox etc return None, None try: client.display_properties.append('summary') # available_classes = sorted(client.show_classes()) # doesn't work anymore # print(available_classes) available_classes = sorted([c['class'] for c in client.schema.get()['classes']]) # print(available_classes) # st.write(f"Available classes: {available_classes}") # st.write(f"Available classes type: {type(available_classes)}") logger.info(available_classes) return client, available_classes except Exception: return client, [] ############## def main(): with st.sidebar: _, center, _ = st.columns([3, 5, 3]) with center: st.text("Search Lab") _, center, _ = st.columns([2, 5, 3]) with center: if we_are_online: st.text("Running ONLINE") # st.text("(UNSTABLE)") else: st.text("Running OFFLINE") st.write("----------") hybrid_search = st.toggle('Hybrid Search', True) if hybrid_search: alpha_input = st.slider(label='Alpha',min_value=0.00, max_value=1.00, value=0.40, step=0.05, key=1) retrieval_limit = st.slider(label='Hybrid Search Results', min_value=10, max_value=300, value=10, step=10) hybrid_filter = st.toggle('Filter Search using Guest name', True) # i.e. look only at guests' data rerank = st.toggle('Rerank', True) if rerank: reranker_topk = st.slider(label='Reranker Top K',min_value=1, max_value=5, value=3, step=1) else: # needed to not fill the LLM with too many responses (> context size) # we could make it dependent on the model reranker_topk = 3 rag_it = st.toggle(f"RAG it with '{model_nameGPT}'", True) if rag_it: # st.write(f"Using LLM '{model_nameGPT}'") llm_temperature = st.slider(label='LLM T˚', min_value=0.0, max_value=2.0, value=0.01, step=0.10 ) model_name_or_path = st.selectbox(label='Model Name:', options=available_models, index=available_models.index(model_default), placeholder='Select Model') delete_models = st.button('Delete models') if delete_models: # model_path = os.path.join("models", model_name_or_path.split('/')[-1]) # if os.path.isdir(model_path): # shutil.rmtree(model_path) for model in os.listdir("models"): model_path = os.path.join("models", model) if os.path.isdir(model_path) and 'finetuned-all-mpnet-base-v2-300' not in model_path: shutil.rmtree(model_path) st.write("Models deleted") if we_are_not_online: st.write("Experimental and time limited 2'") c1,c2 = st.columns([8,1]) with c1: finetune_model = st.button('Finetune on Modal A100 GPU') if finetune_model: from finetune_backend import finetune if 'finetuned' in model_name_or_path: st.write("Model already finetuned") elif "models/" in model_name_or_path: st.write("sentence-transformers models only!") else: try: if 'finetuned' in model_name_or_path: st.write("Model already finetuned") else: with c2: with st.spinner(''): model_path = finetune(model_name_or_path, savemodel=True, outpath='models') with c1: if model_path is not None: if model_name_or_path.split('/')[-1] not in model_path: st.sidebar.write(model_path) # a warning from finetuning in this case # TODO: add model to Weaviate and to model list except Exception: st.write("Model not found on HF or error") else: st.write("Finetuning not available on Streamlit online because of space limitations") check_model(model_name_or_path) client, available_classes = get_weaviate_client(Wapi_key, url, model_name_or_path, openai_api_key) print("Available classes:", available_classes) if client is None: # maybe the free sandbox has expired, or the api key is wrong st.sidebar.write(f"Weaviate sandbox not accessible or expired") # st.stop() elif available_classes: start_class = 'Impact_theory_all_mpnet_base_v2_finetuned' class_name = st.selectbox( label='Class Name:', options=available_classes, index=available_classes.index(start_class), placeholder='Select Class Name' ) st.write("----------") if we_are_not_online: c1,c2 = st.columns([8,1]) with c1: show_metrics = st.button('Show Metrics on Golden set') if show_metrics: # we must add it because the hybrid search toggle could hide it alpha_input2 = st.slider(label='Alpha',min_value=0.00, max_value=1.00, value=0.40, step=0.05, key=2) # _, center, _ = st.columns([3, 5, 3]) # with center: # st.text("Metrics") with c2: with st.spinner(''): metrics = execute_evaluation(golden_dataset, class_name, client, alpha=alpha_input2) with c1: kw_hit_rate = metrics['kw_hit_rate'] kw_mrr = metrics['kw_mrr'] hybrid_hit_rate = metrics['hybrid_hit_rate'] vector_hit_rate = metrics['vector_hit_rate'] vector_mrr = metrics['vector_mrr'] total_misses = metrics['total_misses'] st.text(f"KW hit rate: {kw_hit_rate}") st.text(f"Vector hit rate: {vector_hit_rate}") st.text(f"Hybrid hit rate: {hybrid_hit_rate}") st.text(f"Hybrid MRR: {vector_mrr}") st.text(f"Total misses: {total_misses}") st.write("----------") else: # Weaviate doesn't know this model, maybe we're just finetuning a model st.sidebar.write(f"Model Unknown to Weaviate") st.title("Chat with the Impact Theory podcasts!") # st.image('./assets/impact-theory-logo.png', width=400) st.image('assets/it_tom_bilyeu.png', use_column_width=True) # st.subheader(f"Chat with the Impact Theory podcast: ") st.write('\n') # st.stop() st.write("\u21D0 Open the sidebar to change Search settings \n ") # https://home.unicode.org also 21E0, 21B0 B2 D0 if client is None: st.write("Weaviate sandbox not accessible or expired!!! Stopping execution!") st.stop() elif not available_classes: # we have to stop here, to exit the 'with st.sidebar' block and display the banner at least st.stop() if not hybrid_search: st.stop() col1, _ = st.columns([3,7]) with col1: guest = st.selectbox('Select A Guest', options=guest_list, index=None, placeholder='Select Guest') col1, col2 = st.columns([7,3]) with col1: if guest is None: msg = f'Select a guest before asking your question:' else: msg = f'Enter your question about {guest}:' textbox = st.empty() # best solution I found to be able to change the text inside a text_input box afterwards, using a key query = textbox.text_input(msg, value="", placeholder="You can refer to the guest with PRONOUNS", key=st.session_state.key) # st.write(f"Guest = {guest}") # st.write(f"key = {st.session_state.key}") st.write('\n\n\n\n\n') reworded_query = {'changed': False, 'status': 'error'} # at start, the query is empty valid_response = [] # at start, the query is empty, so prevent the search if query: if guest is None: st.session_state.key += 1 query = textbox.text_input(msg, value="", placeholder="YOU MUST SELECT A GUEST BEFORE ASKING A QUESTION", key=st.session_state.key) # st.write(f"key = {st.session_state.key}") st.stop() else: # st.write(f'It looks like you selected {guest} as a filter (It is ignored for now).') with col2: # let's add a nice pulse bar while generating the response with hc.HyLoader('', hc.Loaders.pulse_bars, primary_color= 'red', height=50): #"#0e404d" for image green with col1: if st.toggle('Rewrite query with LLM', True): # let's use Llama2, and fall back on GPT3.5 if it fails reworded_query = reword_query(query, guest, model_name='gpt-3.5-turbo-0125') new_query = reworded_query['rewritten_question'] if reworded_query['status'] != 'error': # or reworded_query['changed']: guest_lastname = guest.split(' ')[1] if guest_lastname not in new_query: # if the guest name is not in the rewritten question, we add it new_query = f"About {guest}, " + new_query query = new_query st.write(f"New query: {query}") # we can arrive here only if a guest was selected where_filter = WhereFilter(path=['guest'], operator='Equal', valueText=guest).todict() \ if hybrid_filter else None hybrid_response = client.hybrid_search(query, class_name, # properties=['content'], #['title', 'summary', 'content'], alpha=alpha_input, display_properties=client.display_properties, where_filter=where_filter, limit=retrieval_limit) response = hybrid_response if rerank: # rerank results with cross encoder ranked_response = reranker.rerank(response, query, apply_sigmoid=True, # score between 0 and 1 top_k=reranker_topk) logger.info(ranked_response) expanded_response = expand_content(ranked_response, cache, content_key='doc_id', create_new_list=True) response = expanded_response # make sure token count < threshold token_threshold = 8000 if model_nameGPT == model_ids[0] else 3500 valid_response = validate_token_threshold(response, question_answering_prompt_series, query=query, tokenizer= encoding,# variable from ENCODING, token_threshold=token_threshold, verbose=True) # st.write(f"Number of results: {len(valid_response)}") # I jumped out of col1 to get all page width, so need to retest query if query: # creates container for LLM response to position it above search results chat_container, response_box = [], st.empty() # # RAG time !! execute chat call to LLM if rag_it: # st.subheader("Response from Impact Theory (context)") # will appear under the answer, moved it into the response box # generate LLM prompt prompt = generate_prompt_series(query=query, results=valid_response) GPTllm = GPT_Turbo(model=model_nameGPT, api_key=openai_api_key) try: # inserts chat stream from LLM for resp in GPTllm.get_chat_completion(prompt=prompt, temperature=llm_temperature, max_tokens=350, show_response=True, stream=True): with response_box: content = resp.choices[0].delta.content if content: chat_container.append(content) result = "".join(chat_container).strip() response_box.markdown(f"### Response from Impact Theory (RAG):\n\n{result}") except BadRequestError as e: logger.info('Making request with smaller context') valid_response = validate_token_threshold(response, question_answering_prompt_series, query=query, tokenizer=encoding, token_threshold=3500, verbose=True) # if reranker is off, we may receive a LOT of responses # so we must reduce the context size manually if not rerank: valid_response = valid_response[:reranker_topk] prompt = generate_prompt_series(query=query, results=valid_response) for resp in GPTllm.get_chat_completion(prompt=prompt, temperature=llm_temperature, max_tokens=350, # expand for more verbose answers show_response=True, stream=True): try: # inserts chat stream from LLM with response_box: content = resp.choice[0].delta.content if content: chat_container.append(content) result = "".join(chat_container).strip() response_box.markdown(f"### Response from Impact Theory (RAG):\n\n{result}") except Exception as e: print(e) st.markdown("----") st.subheader("Search Results") for i, hit in enumerate(valid_response): col1, col2 = st.columns([7, 3], gap='large') image = hit['thumbnail_url'] # get thumbnail_url episode_url = hit['episode_url'] # get episode_url title = hit["title"] # get title show_length = hit["length"] # get length time_string = str(timedelta(seconds=show_length)) # convert show_length to readable time string with col1: st.write(search_result(i=i, url=episode_url, guest=hit['guest'], title=title, content='', length=time_string), unsafe_allow_html=True) st.write('\n\n') with col2: #st.write(f"", # unsafe_allow_html=True) #st.markdown(f"[![{title}]({image})]({episode_url})") # st.markdown(f'' # f'' # f'', # unsafe_allow_html=True) st.image(image, caption=title.split('|')[0], width=200, use_column_width=False) # let's use all width for the content st.write(hit['content']) def get_answer(query, valid_response, GPTllm): # generate LLM prompt prompt = generate_prompt_series(query=query, results=valid_response) return GPTllm.get_chat_completion(prompt=prompt, system_message='answer this question based on the podcast material', temperature=0, max_tokens=500, stream=False, show_response=False) def reword_query(query, guest, model_name='llama2-13b-chat', response_processing=True): """ Asks LLM to rewrite the query when the guest name is missing. Args: query (str): user query guest (str): guest name model_name (str, optional): name of a LLM model to be used """ # tags = {'llama2-13b-chat': {'start': '', 'end': '', 'instruction': '[INST]', 'system': '[SYS]'}, # 'gpt-3.5-turbo-0613': {'start': '<|startoftext|>', 'end': '', 'instruction': "```", 'system': ```}} prompt_fields = { "you_are":f"You are an expert in linguistics and semantics, analyzing the question asked by a user to a vector search system, \ and making sure that the question is well formulated and understandable by any average reader.", "your_task":f"Your task is to detect if the name of the guest ({guest}) is mentioned in the question '{query}', \ If that is not the case, rewrite the question using the guest name, \ without changing the meaning of the question. \ Most of the time, the user will have used a pronoun to designate the guest, in which case, \ simply replace the pronoun with the guest name. \ If the guest name is already present in the question, return the original question as is.", "final_instruction":f"Only regenerate the requested rewritten question or the original, WITHOUT ANY COMMENT OR REPHRASING. \ Your answer must be as close as possible to the original question, \ and exactly identical, word for word, if the user mentions the guest name, i.e. {guest}.", "question":f"{query}" } # prompt created by chatGPT :-) # and Llama still outputs the original question and precedes the answer with 'rewritten question' prompt_fields2 = { "you_are": ( "You are an expert in linguistics and semantics. Your role is to analyze questions asked to a vector search system." ), "your_task": ( f"Detect if the guest's FULL name, {guest}, is mentioned in the user's question. " "If not, rewrite the question by replacing pronouns or indirect references with the guest's name." \ "If yes, return the original question as is, without any change at all, not even punctuation," "except a question mark that you MUST add if it's missing." ), "question": ( f"Original question: '{query}'. " "Rewrite this question to include the guest's FULL name if it's not already mentioned." "Add a question mark if it's missing, nothing else." ), "final_instruction": ( "Create a rewritten question or keep the original question as is. " "Do not include any labels, titles, or additional text before or after the question." "The Only thing you can and MUST add is a question mark if it's missing." "Return a json object, with the key 'original_question' for the original question, \ and 'rewritten_question' for the rewritten question \ and 'changed' being True if you changed the answer, otherwise False." ), } if model_name == 'llama2-13b-chat': # special tags are used: # `` - start prompt tag # `[INST], [/INST]` - Opening and closing model instruction tags # `<<>>, <>` - Opening and closing system prompt tags llama_prompt = """ [INST] <> {you_are} <> {your_task}\n ``` \n\n Question: {question}\n {final_instruction} [/INST] Answer: """ prompt = llama_prompt.format(**prompt_fields2) headers = {"Authorization": f"Bearer {hf_token}", "Content-Type": "application/json",} json_body = { "inputs": prompt, "parameters": {"max_new_tokens":400, "repetition_penalty": 1.0, "temperature":0.01} } response = requests.request("POST", hf_endpoint, headers=headers, data=json.dumps(json_body)) response = json.loads(response.content.decode("utf-8")) # ^ will not process the badly formatted generated text, so we do it ourselves if isinstance(response, dict) and 'error' in response: print("Found error") print(response) # return {'error': response['error'], 'rewritten_question': query, 'changed': False, 'status': 'error'} # I test this here otherwise it gets in col 2 or 1, which are too # if reworded_query['status'] == 'error': # st.write(f"Error in LLM response: 'error':{reworded_query['error']}") # st.write("The LLM could not connect to the server. Please try again later.") # st.stop() return reword_query(query, guest, model_name='gpt-3.5-turbo-0125') if response_processing: if isinstance(response, list) and isinstance(response[0], dict) and 'generated_text' in response[0]: print("Found generated text") response0 = response[0]['generated_text'] pattern = r'\"(\w+)\":\s*(\".*?\"|\w+)' matches = re.findall(pattern, response0) # let's build a dictionary result = {key: json.loads(value) if value.startswith("\"") else value for key, value in matches} return result | {'status': 'success'} else: print("Found no answer") return reword_query(query, guest, model_name='gpt-3.5-turbo-0125') # return {'original_question': query, 'rewritten_question': query, 'changed': False, 'status': 'no properly formatted answer' } else: return response # return response # assert 'error' not in response, f"Error in LLM response: {response['error']}" # assert 'generated_text' in response[0], f"Error in LLM response: {response}, no 'generated_text' field" # # let's extract the rewritten question # return response[0]['generated_text'] .split("Rewritten question: '")[-1][:-1] else: # we assume / force openai model_ids = ['gpt-3.5-turbo-0125', 'gpt-3.5-turbo-16k', 'gpt-3.5-turbo-0613'] if model_name not in model_ids: model_name = model_ids[0] GPTllm = GPT_Turbo(model=model_name, api_key=openai_api_key) openai_prompt = """ {your_task} \n {final_instruction} /n ``` \n\n Question: {question}\n Answer: """ prompt = openai_prompt.format(**prompt_fields) try: # https://platform.openai.com/docs/guides/text-generation/chat-completions-api resp = GPTllm.get_chat_completion(prompt=prompt, system_message=prompt_fields['you_are'], user_message = None, temperature=0.01, max_tokens=1500, # it's a long question... show_response=True, stream=False) if resp.choices[0].finish_reason == 'stop': if guest in resp.choices[0].message.content: new_question = resp.choices[0].message.content return {'rewritten_question': new_question, 'changed': True, 'status': 'success'} else: raise Exception("LLM did not stop") # to go to the except block except Exception: return {'rewritten_question': query, 'changed': False, 'status': 'not success'} if __name__ == '__main__': main() # streamlit run app.py --server.allowRunOnSave True