import gradio as gr from transformers.generation.stopping_criteria import StoppingCriteria, StoppingCriteriaList from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig from peft import PeftModel import torch import pinecone from sentence_transformers import SentenceTransformer from tqdm import tqdm from sentence_transformers.cross_encoder import CrossEncoder import numpy as np from torch import nn import os # Set up semantic search PINECONE_API_KEY = os.environ.get('PINECONE_API_KEY') def get_embedding(text): embed_text = sentencetransformer_model.encode(text) vector_text = embed_text.tolist() return vector_text def query_from_pinecone(query, top_k=3): # get embedding from THE SAME embedder as the documents query_embedding = get_embedding(query) return index.query( vector=query_embedding, top_k=top_k, include_metadata=True # gets the metadata (dates, text, etc) ).get('matches') def get_results_from_pinecone(query, top_k=3, re_rank=True, verbose=True): results_from_pinecone = query_from_pinecone(query, top_k=top_k) if not results_from_pinecone: return [] if verbose: print("Query:", query) final_results = [] if re_rank: if verbose: print('Document ID (Hash)\t\tRetrieval Score\tCE Score\tText') sentence_combinations = [[query, result_from_pinecone['metadata']['text']] for result_from_pinecone in results_from_pinecone] # Compute the similarity scores for these combinations similarity_scores = cross_encoder.predict(sentence_combinations, activation_fct=nn.Sigmoid()) # Sort the scores in decreasing order sim_scores_argsort = reversed(np.argsort(similarity_scores)) # Print the scores for idx in sim_scores_argsort: result_from_pinecone = results_from_pinecone[idx] final_results.append(result_from_pinecone) if verbose: print(f"{result_from_pinecone['id']}\t{result_from_pinecone['score']:.2f}\t{similarity_scores[idx]:.2f}\t{result_from_pinecone['metadata']['text'][:50]}") return final_results if verbose: print('Document ID (Hash)\t\tRetrieval Score\tText') for result_from_pinecone in results_from_pinecone: final_results.append(result_from_pinecone) if verbose: print(f"{result_from_pinecone['id']}\t{result_from_pinecone['score']:.2f}\t{result_from_pinecone['metadata']['text'][:50]}") return final_results def semantic_search(prompt): final_results = get_results_from_pinecone(prompt, top_k=9, re_rank=True, verbose=True) return '\n\n'.join(res['metadata']['text'].strip() for res in final_results[:3]) cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-12-v2') sentencetransformer_model = SentenceTransformer('sentence-transformers/multi-qa-mpnet-base-cos-v1') pinecone_key = PINECONE_API_KEY INDEX_NAME = 'k8s-semantic-search' NAMESPACE = 'default' pinecone.init(api_key=pinecone_key, environment="gcp-starter") if not INDEX_NAME in pinecone.list_indexes(): pinecone.create_index( INDEX_NAME, # The name of the index dimension=768, # The dimensionality of the vectors metric='cosine', # The similarity metric to use when searching the index pod_type='starter' # The type of Pinecone pod ) index = pinecone.Index(INDEX_NAME) # Set up mistral model base_model_id = 'mistralai/Mistral-7B-Instruct-v0.1' lora_model_id = 'ComponentSoft/mistral-kubectl-instruct' tokenizer = AutoTokenizer.from_pretrained( lora_model_id, padding_side="left", add_eos_token=False, add_bos_token=True, ) tokenizer.pad_token = tokenizer.eos_token bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16 ) base_model = AutoModelForCausalLM.from_pretrained( base_model_id, quantization_config=bnb_config, use_cache=True, trust_remote_code=True, ) model = PeftModel.from_pretrained(base_model, lora_model_id) model.eval() stop_terms=["", "#End"] eos_token_ids_custom = [torch.tensor(tokenizer.encode(term, add_special_tokens=False)).to("cuda") for term in stop_terms] category_terms=["", "\n"] category_eos_token_ids_custom = [torch.tensor(tokenizer.encode(term, add_special_tokens=False)).to("cuda") for term in category_terms] class EvalStopCriterion(StoppingCriteria): def __call__(self, input_ids: torch.LongTensor, score: torch.FloatTensor, **kwargs): return any(torch.equal(e, input_ids[0][-len(e):]) for e in eos_token_ids_custom) class CategoryStopCriterion(StoppingCriteria): def __call__(self, input_ids: torch.LongTensor, score: torch.FloatTensor, **kwargs): return any(torch.equal(e, input_ids[0][-len(e):]) for e in category_eos_token_ids_custom) start_template = '### Answer:' command_template = '# Command:' end_template = '#End' def text_to_text_generation(prompt): prompt = prompt.strip() '' is_kubectl_prompt = ( f"[INST] You are a helpful assistant who classifies prompts into three categories. Respond with 0 if it pertains to a 'kubectl' operation. This is an instruction that can be answered with a 'kubectl' action. Look for keywords like 'get', 'list', 'create', 'show', 'view', and other command-like words. This category is an instruction instead of a question. Respond with 1 only if the prompt is a question, and is about a definition related to Kubernetes, or non-action inquiries. Respond with 2 every other scenario, for example if the question is a general question, not related to Kubernetes or 'kubectl'.\n" f"So for instance the following:\n" f"List all pods in Kubernetes\n" f"Would get a response:\n" f"0 [/INST]" f'text: "{prompt}"' f'response (0/1/2): ' ) model_input = tokenizer(is_kubectl_prompt, return_tensors="pt").to("cuda") with torch.no_grad(): response = tokenizer.decode(model.generate(**model_input, max_new_tokens=8, pad_token_id=tokenizer.eos_token_id, repetition_penalty=1.15, stopping_criteria=StoppingCriteriaList([CategoryStopCriterion()]))[0], skip_special_tokens=True) response = response[len(is_kubectl_prompt):] print('-----------------------------QUERY START-----------------------------') print('Prompt: ' + prompt) print('Classified as: ' + response) response_num = 2 # Default to generic question if '0' in response: response_num = 0 elif '1' in response: response_num = 1 # Check if general question if response_num == 0: prompt = f'[INST] {prompt}\n Lets think step by step. [/INST] {start_template}' elif response_num == 1: retrieved_results = semantic_search(prompt) print('Query:') print(f'[INST] You are an assistant who summarizes results retrieved from a book about Kubernetes. This summary should answer the question. If the answer is not in the retrieved results, use your general knowledge. [/INST] Question: {prompt}\nRetrieved results:\n{retrieved_results}\nResponse:') prompt = f'[INST] You are an assistant who summarizes results retrieved from a book about Kubernetes. This summary should answer the question. If the answer is not in the retrieved results, use your general knowledge. [/INST] Question: {prompt}\nRetrieved results:\n{retrieved_results}\nResponse:' else: prompt = f'[INST] {prompt} [/INST]' # Generate output model_input = tokenizer(prompt, return_tensors="pt").to("cuda") with torch.no_grad(): response = tokenizer.decode(model.generate(**model_input, max_new_tokens=256, pad_token_id=tokenizer.eos_token_id, repetition_penalty=1.15, stopping_criteria=StoppingCriteriaList([EvalStopCriterion()]))[0], skip_special_tokens=True) # Get the relevalt parts start = response.index(start_template) + len(start_template) if start_template in response else len(prompt) start = response.index(command_template) + len(command_template) if command_template in response else start end = response.index(end_template) if end_template in response else len(response) true_response = response[start:end].strip() print('Returned: ' + true_response) print('------------------------------QUERY END------------------------------') return true_response iface = gr.Interface(fn=text_to_text_generation, inputs="text", outputs="text") iface.launch()