From ef9c7f907b67428806f76cfe33764d2960259447 Mon Sep 17 00:00:00 2001 From: LuyaoZhuang Date: Sat, 3 Jan 2026 07:59:37 -0500 Subject: fix add edges --- LinearRAG_upload/src/LinearRAG.py | 623 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 623 insertions(+) create mode 100644 LinearRAG_upload/src/LinearRAG.py diff --git a/LinearRAG_upload/src/LinearRAG.py b/LinearRAG_upload/src/LinearRAG.py new file mode 100644 index 0000000..76b65ad --- /dev/null +++ b/LinearRAG_upload/src/LinearRAG.py @@ -0,0 +1,623 @@ +from src.embedding_store import EmbeddingStore +from src.utils import min_max_normalize +import os +import json +from collections import defaultdict +import numpy as np +import math +from concurrent.futures import ThreadPoolExecutor +from tqdm import tqdm +from src.ner import SpacyNER +import igraph as ig +import re +import logging +import torch +logger = logging.getLogger(__name__) + + +class LinearRAG: + def __init__(self, global_config): + self.config = global_config + logger.info(f"Initializing LinearRAG with config: {self.config}") + retrieval_method = "Vectorized Matrix-based" if self.config.use_vectorized_retrieval else "BFS Iteration" + logger.info(f"Using retrieval method: {retrieval_method}") + + # Setup device for GPU acceleration + self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + if self.config.use_vectorized_retrieval: + logger.info(f"Using device: {self.device} for vectorized retrieval") + + self.dataset_name = global_config.dataset_name + self.load_embedding_store() + self.llm_model = self.config.llm_model + self.spacy_ner = SpacyNER(self.config.spacy_model) + self.graph = ig.Graph(directed=False) + + def load_embedding_store(self): + self.passage_embedding_store = EmbeddingStore(self.config.embedding_model, db_filename=os.path.join(self.config.working_dir,self.dataset_name, "passage_embedding.parquet"), batch_size=self.config.batch_size, namespace="passage") + self.entity_embedding_store = EmbeddingStore(self.config.embedding_model, db_filename=os.path.join(self.config.working_dir,self.dataset_name, "entity_embedding.parquet"), batch_size=self.config.batch_size, namespace="entity") + self.sentence_embedding_store = EmbeddingStore(self.config.embedding_model, db_filename=os.path.join(self.config.working_dir,self.dataset_name, "sentence_embedding.parquet"), batch_size=self.config.batch_size, namespace="sentence") + + def load_existing_data(self,passage_hash_ids): + self.ner_results_path = os.path.join(self.config.working_dir,self.dataset_name, "ner_results.json") + if os.path.exists(self.ner_results_path): + existing_ner_reuslts = json.load(open(self.ner_results_path)) + existing_passage_hash_id_to_entities = existing_ner_reuslts["passage_hash_id_to_entities"] + existing_sentence_to_entities = existing_ner_reuslts["sentence_to_entities"] + existing_passage_hash_ids = set(existing_passage_hash_id_to_entities.keys()) + new_passage_hash_ids = set(passage_hash_ids) - existing_passage_hash_ids + return existing_passage_hash_id_to_entities, existing_sentence_to_entities, new_passage_hash_ids + else: + return {}, {}, passage_hash_ids + + def qa(self, questions): + retrieval_results = self.retrieve(questions) + system_prompt = f"""As an advanced reading comprehension assistant, your task is to analyze text passages and corresponding questions meticulously. Your response start after "Thought: ", where you will methodically break down the reasoning process, illustrating how you arrive at conclusions. Conclude with "Answer: " to present a concise, definitive response, devoid of additional elaborations.""" + all_messages = [] + for retrieval_result in retrieval_results: + question = retrieval_result["question"] + sorted_passage = retrieval_result["sorted_passage"] + prompt_user = """""" + for passage in sorted_passage: + prompt_user += f"{passage}\n" + prompt_user += f"Question: {question}\n Thought: " + messages = [ + {"role": "system", "content": system_prompt}, + {"role": "user", "content": prompt_user} + ] + all_messages.append(messages) + with ThreadPoolExecutor(max_workers=self.config.max_workers) as executor: + all_qa_results = list(tqdm( + executor.map(self.llm_model.infer, all_messages), + total=len(all_messages), + desc="QA Reading (Parallel)" + )) + + for qa_result,question_info in zip(all_qa_results,retrieval_results): + try: + pred_ans = qa_result.split('Answer:')[1].strip() + except: + pred_ans = qa_result + question_info["pred_answer"] = pred_ans + return retrieval_results + + def retrieve(self, questions): + self.entity_hash_ids = list(self.entity_embedding_store.hash_id_to_text.keys()) + self.entity_embeddings = np.array(self.entity_embedding_store.embeddings) + self.passage_hash_ids = list(self.passage_embedding_store.hash_id_to_text.keys()) + self.passage_embeddings = np.array(self.passage_embedding_store.embeddings) + self.sentence_hash_ids = list(self.sentence_embedding_store.hash_id_to_text.keys()) + self.sentence_embeddings = np.array(self.sentence_embedding_store.embeddings) + self.node_name_to_vertex_idx = {v["name"]: v.index for v in self.graph.vs if "name" in v.attributes()} + self.vertex_idx_to_node_name = {v.index: v["name"] for v in self.graph.vs if "name" in v.attributes()} + + # Precompute sparse matrices for vectorized retrieval if needed + if self.config.use_vectorized_retrieval: + logger.info("Precomputing sparse adjacency matrices for vectorized retrieval...") + self._precompute_sparse_matrices() + e2s_shape = self.entity_to_sentence_sparse.shape + s2e_shape = self.sentence_to_entity_sparse.shape + e2s_nnz = self.entity_to_sentence_sparse._nnz() + s2e_nnz = self.sentence_to_entity_sparse._nnz() + logger.info(f"Matrices built: Entity-Sentence {e2s_shape}, Sentence-Entity {s2e_shape}") + logger.info(f"E2S Sparsity: {(1 - e2s_nnz / (e2s_shape[0] * e2s_shape[1])) * 100:.2f}% (nnz={e2s_nnz})") + logger.info(f"S2E Sparsity: {(1 - s2e_nnz / (s2e_shape[0] * s2e_shape[1])) * 100:.2f}% (nnz={s2e_nnz})") + logger.info(f"Device: {self.device}") + + retrieval_results = [] + for question_info in tqdm(questions, desc="Retrieving"): + question = question_info["question"] + question_embedding = self.config.embedding_model.encode(question,normalize_embeddings=True,show_progress_bar=False,batch_size=self.config.batch_size) + seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores = self.get_seed_entities(question) + if len(seed_entities) != 0: + sorted_passage_hash_ids,sorted_passage_scores = self.graph_search_with_seed_entities(question_embedding,seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores) + final_passage_hash_ids = sorted_passage_hash_ids[:self.config.retrieval_top_k] + final_passage_scores = sorted_passage_scores[:self.config.retrieval_top_k] + final_passages = [self.passage_embedding_store.hash_id_to_text[passage_hash_id] for passage_hash_id in final_passage_hash_ids] + else: + sorted_passage_indices,sorted_passage_scores = self.dense_passage_retrieval(question_embedding) + final_passage_indices = sorted_passage_indices[:self.config.retrieval_top_k] + final_passage_scores = sorted_passage_scores[:self.config.retrieval_top_k] + final_passages = [self.passage_embedding_store.texts[idx] for idx in final_passage_indices] + result = { + "question": question, + "sorted_passage": final_passages, + "sorted_passage_scores": final_passage_scores, + "gold_answer": question_info["answer"] + } + retrieval_results.append(result) + return retrieval_results + + def _precompute_sparse_matrices(self): + """ + Precompute and cache sparse adjacency matrices for efficient vectorized retrieval using PyTorch. + This is called once at the beginning of retrieve() to avoid rebuilding matrices per query. + """ + num_entities = len(self.entity_hash_ids) + num_sentences = len(self.sentence_hash_ids) + + # Build entity-to-sentence matrix (Mention matrix) using COO format + entity_to_sentence_indices = [] + entity_to_sentence_values = [] + + for entity_hash_id, sentence_hash_ids in self.entity_hash_id_to_sentence_hash_ids.items(): + entity_idx = self.entity_embedding_store.hash_id_to_idx[entity_hash_id] + for sentence_hash_id in sentence_hash_ids: + sentence_idx = self.sentence_embedding_store.hash_id_to_idx[sentence_hash_id] + entity_to_sentence_indices.append([entity_idx, sentence_idx]) + entity_to_sentence_values.append(1.0) + + # Build sentence-to-entity matrix + sentence_to_entity_indices = [] + sentence_to_entity_values = [] + + for sentence_hash_id, entity_hash_ids in self.sentence_hash_id_to_entity_hash_ids.items(): + sentence_idx = self.sentence_embedding_store.hash_id_to_idx[sentence_hash_id] + for entity_hash_id in entity_hash_ids: + entity_idx = self.entity_embedding_store.hash_id_to_idx[entity_hash_id] + sentence_to_entity_indices.append([sentence_idx, entity_idx]) + sentence_to_entity_values.append(1.0) + + # Convert to PyTorch sparse tensors (COO format, then convert to CSR for efficiency) + if len(entity_to_sentence_indices) > 0: + e2s_indices = torch.tensor(entity_to_sentence_indices, dtype=torch.long).t() + e2s_values = torch.tensor(entity_to_sentence_values, dtype=torch.float32) + self.entity_to_sentence_sparse = torch.sparse_coo_tensor( + e2s_indices, e2s_values, (num_entities, num_sentences), device=self.device + ).coalesce() + else: + self.entity_to_sentence_sparse = torch.sparse_coo_tensor( + torch.zeros((2, 0), dtype=torch.long), torch.zeros(0, dtype=torch.float32), + (num_entities, num_sentences), device=self.device + ) + + if len(sentence_to_entity_indices) > 0: + s2e_indices = torch.tensor(sentence_to_entity_indices, dtype=torch.long).t() + s2e_values = torch.tensor(sentence_to_entity_values, dtype=torch.float32) + self.sentence_to_entity_sparse = torch.sparse_coo_tensor( + s2e_indices, s2e_values, (num_sentences, num_entities), device=self.device + ).coalesce() + else: + self.sentence_to_entity_sparse = torch.sparse_coo_tensor( + torch.zeros((2, 0), dtype=torch.long), torch.zeros(0, dtype=torch.float32), + (num_sentences, num_entities), device=self.device + ) + + def graph_search_with_seed_entities(self, question_embedding, seed_entity_indices, seed_entities, seed_entity_hash_ids, seed_entity_scores): + if self.config.use_vectorized_retrieval: + entity_weights, actived_entities = self.calculate_entity_scores_vectorized(question_embedding,seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores) + else: + entity_weights, actived_entities = self.calculate_entity_scores(question_embedding,seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores) + passage_weights = self.calculate_passage_scores(question_embedding,actived_entities) + node_weights = entity_weights + passage_weights + ppr_sorted_passage_indices,ppr_sorted_passage_scores = self.run_ppr(node_weights) + return ppr_sorted_passage_indices,ppr_sorted_passage_scores + + def run_ppr(self, node_weights): + reset_prob = np.where(np.isnan(node_weights) | (node_weights < 0), 0, node_weights) + pagerank_scores = self.graph.personalized_pagerank( + vertices=range(len(self.node_name_to_vertex_idx)), + damping=self.config.damping, + directed=False, + weights='weight', + reset=reset_prob, + implementation='prpack' + ) + + doc_scores = np.array([pagerank_scores[idx] for idx in self.passage_node_indices]) + sorted_indices_in_doc_scores = np.argsort(doc_scores)[::-1] + sorted_passage_scores = doc_scores[sorted_indices_in_doc_scores] + + sorted_passage_hash_ids = [ + self.vertex_idx_to_node_name[self.passage_node_indices[i]] + for i in sorted_indices_in_doc_scores + ] + + return sorted_passage_hash_ids, sorted_passage_scores.tolist() + + def calculate_entity_scores(self,question_embedding,seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores): + actived_entities = {} + entity_weights = np.zeros(len(self.graph.vs["name"])) + for seed_entity_idx,seed_entity,seed_entity_hash_id,seed_entity_score in zip(seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores): + actived_entities[seed_entity_hash_id] = (seed_entity_idx, seed_entity_score, 1) + seed_entity_node_idx = self.node_name_to_vertex_idx[seed_entity_hash_id] + entity_weights[seed_entity_node_idx] = seed_entity_score + used_sentence_hash_ids = set() + current_entities = actived_entities.copy() + iteration = 1 + while len(current_entities) > 0 and iteration < self.config.max_iterations: + new_entities = {} + for entity_hash_id, (entity_id, entity_score, tier) in current_entities.items(): + if entity_score < self.config.iteration_threshold: + continue + sentence_hash_ids = [sid for sid in list(self.entity_hash_id_to_sentence_hash_ids[entity_hash_id]) if sid not in used_sentence_hash_ids] + if not sentence_hash_ids: + continue + sentence_indices = [self.sentence_embedding_store.hash_id_to_idx[sid] for sid in sentence_hash_ids] + sentence_embeddings = self.sentence_embeddings[sentence_indices] + question_emb = question_embedding.reshape(-1, 1) if len(question_embedding.shape) == 1 else question_embedding + sentence_similarities = np.dot(sentence_embeddings, question_emb).flatten() + top_sentence_indices = np.argsort(sentence_similarities)[::-1][:self.config.top_k_sentence] + for top_sentence_index in top_sentence_indices: + top_sentence_hash_id = sentence_hash_ids[top_sentence_index] + top_sentence_score = sentence_similarities[top_sentence_index] + used_sentence_hash_ids.add(top_sentence_hash_id) + entity_hash_ids_in_sentence = self.sentence_hash_id_to_entity_hash_ids[top_sentence_hash_id] + for next_entity_hash_id in entity_hash_ids_in_sentence: + next_entity_score = entity_score * top_sentence_score + if next_entity_score < self.config.iteration_threshold: + continue + next_enitity_node_idx = self.node_name_to_vertex_idx[next_entity_hash_id] + entity_weights[next_enitity_node_idx] += next_entity_score + new_entities[next_entity_hash_id] = (next_enitity_node_idx, next_entity_score, iteration+1) + actived_entities.update(new_entities) + current_entities = new_entities.copy() + iteration += 1 + return entity_weights, actived_entities + + def calculate_entity_scores_vectorized(self,question_embedding,seed_entity_indices,seed_entities,seed_entity_hash_ids,seed_entity_scores): + """ + GPU-accelerated vectorized version using PyTorch sparse tensors. + Uses sparse representation for both matrices and entity score vectors for maximum efficiency. + Now includes proper dynamic pruning to match BFS behavior: + - Sentence deduplication (tracks used sentences) + - Per-entity top-k sentence selection + - Proper threshold-based pruning + """ + # Initialize entity weights + entity_weights = np.zeros(len(self.graph.vs["name"])) + num_entities = len(self.entity_hash_ids) + num_sentences = len(self.sentence_hash_ids) + + # Compute all sentence similarities with the question at once + question_emb = question_embedding.reshape(-1, 1) if len(question_embedding.shape) == 1 else question_embedding + sentence_similarities_np = np.dot(self.sentence_embeddings, question_emb).flatten() + + # Convert to torch tensors and move to device + sentence_similarities = torch.from_numpy(sentence_similarities_np).float().to(self.device) + + # Track used sentences for deduplication (like BFS version) + used_sentence_mask = torch.zeros(num_sentences, dtype=torch.bool, device=self.device) + + # Initialize seed entity scores as sparse tensor + seed_indices = torch.tensor([[idx] for idx in seed_entity_indices], dtype=torch.long).t() + seed_values = torch.tensor(seed_entity_scores, dtype=torch.float32) + entity_scores_sparse = torch.sparse_coo_tensor( + seed_indices, seed_values, (num_entities,), device=self.device + ).coalesce() + + # Also maintain a dense accumulator for total scores + entity_scores_dense = torch.zeros(num_entities, dtype=torch.float32, device=self.device) + entity_scores_dense.scatter_(0, torch.tensor(seed_entity_indices, device=self.device), + torch.tensor(seed_entity_scores, dtype=torch.float32, device=self.device)) + + # Initialize actived_entities + actived_entities = {} + for seed_entity_idx, seed_entity, seed_entity_hash_id, seed_entity_score in zip( + seed_entity_indices, seed_entities, seed_entity_hash_ids, seed_entity_scores + ): + actived_entities[seed_entity_hash_id] = (seed_entity_idx, seed_entity_score, 0) + seed_entity_node_idx = self.node_name_to_vertex_idx[seed_entity_hash_id] + entity_weights[seed_entity_node_idx] = seed_entity_score + + current_entity_scores_sparse = entity_scores_sparse + + # Iterative matrix-based propagation using sparse matrices on GPU + for iteration in range(1, self.config.max_iterations): + # Convert sparse tensor to dense for threshold operation + current_entity_scores_dense = current_entity_scores_sparse.to_dense() + + # Apply threshold to current scores + current_entity_scores_dense = torch.where( + current_entity_scores_dense >= self.config.iteration_threshold, + current_entity_scores_dense, + torch.zeros_like(current_entity_scores_dense) + ) + + # Get non-zero indices for sparse representation + nonzero_mask = current_entity_scores_dense > 0 + nonzero_indices = torch.nonzero(nonzero_mask, as_tuple=False).squeeze(-1) + + if len(nonzero_indices) == 0: + break + + # Extract non-zero values and create sparse tensor + nonzero_values = current_entity_scores_dense[nonzero_indices] + current_entity_scores_sparse = torch.sparse_coo_tensor( + nonzero_indices.unsqueeze(0), nonzero_values, (num_entities,), device=self.device + ).coalesce() + + # Step 1: Sparse entity scores @ Sparse E2S matrix + # Convert sparse vector to 2D for matrix multiplication + current_scores_2d = torch.sparse_coo_tensor( + torch.stack([nonzero_indices, torch.zeros_like(nonzero_indices)]), + nonzero_values, + (num_entities, 1), + device=self.device + ).coalesce() + + # E @ E2S -> sentence activation scores (sparse @ sparse = dense) + sentence_activation = torch.sparse.mm( + self.entity_to_sentence_sparse.t(), + current_scores_2d + ) + # Convert to dense before squeeze to avoid CUDA sparse tensor issues + if sentence_activation.is_sparse: + sentence_activation = sentence_activation.to_dense() + sentence_activation = sentence_activation.squeeze() + + # Apply sentence deduplication: mask out used sentences + sentence_activation = torch.where( + used_sentence_mask, + torch.zeros_like(sentence_activation), + sentence_activation + ) + + # Step 2: Apply sentence similarities (element-wise on dense vector) + weighted_sentence_scores = sentence_activation * sentence_similarities + + # Implement per-entity top-k sentence selection (more aggressive pruning) + # For vectorized efficiency, we use a tighter global approximation + num_active = len(nonzero_indices) + if num_active > 0 and self.config.top_k_sentence > 0: + # Calculate adaptive k based on number of active entities + # Use per-entity top-k approximation: num_active * top_k_sentence + k = min(int(num_active * self.config.top_k_sentence), len(weighted_sentence_scores)) + if k > 0: + # Get top-k sentences + top_k_values, top_k_indices = torch.topk(weighted_sentence_scores, k) + # Zero out all non-top-k sentences + mask = torch.zeros_like(weighted_sentence_scores, dtype=torch.bool) + mask[top_k_indices] = True + weighted_sentence_scores = torch.where( + mask, + weighted_sentence_scores, + torch.zeros_like(weighted_sentence_scores) + ) + + # Mark these sentences as used for deduplication + used_sentence_mask[top_k_indices] = True + + # Step 3: Weighted sentences @ S2E -> propagate to next entities + # Convert to sparse for more efficient computation + weighted_nonzero_mask = weighted_sentence_scores > 0 + weighted_nonzero_indices = torch.nonzero(weighted_nonzero_mask, as_tuple=False).squeeze(-1) + + if len(weighted_nonzero_indices) > 0: + weighted_nonzero_values = weighted_sentence_scores[weighted_nonzero_indices] + weighted_scores_2d = torch.sparse_coo_tensor( + torch.stack([weighted_nonzero_indices, torch.zeros_like(weighted_nonzero_indices)]), + weighted_nonzero_values, + (num_sentences, 1), + device=self.device + ).coalesce() + + next_entity_scores_result = torch.sparse.mm( + self.sentence_to_entity_sparse.t(), + weighted_scores_2d + ) + # Convert to dense before squeeze to avoid CUDA sparse tensor issues + if next_entity_scores_result.is_sparse: + next_entity_scores_result = next_entity_scores_result.to_dense() + next_entity_scores_dense = next_entity_scores_result.squeeze() + else: + next_entity_scores_dense = torch.zeros(num_entities, dtype=torch.float32, device=self.device) + + # Update entity scores (accumulate in dense format) + entity_scores_dense += next_entity_scores_dense + + # Update actived_entities dictionary (only for entities above threshold) + next_entity_scores_np = next_entity_scores_dense.cpu().numpy() + active_indices = np.where(next_entity_scores_np >= self.config.iteration_threshold)[0] + for entity_idx in active_indices: + score = next_entity_scores_np[entity_idx] + entity_hash_id = self.entity_hash_ids[entity_idx] + if entity_hash_id not in actived_entities or actived_entities[entity_hash_id][1] < score: + actived_entities[entity_hash_id] = (entity_idx, float(score), iteration) + + # Prepare sparse tensor for next iteration + next_nonzero_mask = next_entity_scores_dense > 0 + next_nonzero_indices = torch.nonzero(next_nonzero_mask, as_tuple=False).squeeze(-1) + if len(next_nonzero_indices) > 0: + next_nonzero_values = next_entity_scores_dense[next_nonzero_indices] + current_entity_scores_sparse = torch.sparse_coo_tensor( + next_nonzero_indices.unsqueeze(0), next_nonzero_values, + (num_entities,), device=self.device + ).coalesce() + else: + break + + # Convert back to numpy for final processing + entity_scores_final = entity_scores_dense.cpu().numpy() + + # Map entity scores to graph node weights (only for non-zero scores) + nonzero_indices = np.where(entity_scores_final > 0)[0] + for entity_idx in nonzero_indices: + score = entity_scores_final[entity_idx] + entity_hash_id = self.entity_hash_ids[entity_idx] + entity_node_idx = self.node_name_to_vertex_idx[entity_hash_id] + entity_weights[entity_node_idx] = float(score) + + return entity_weights, actived_entities + + def calculate_passage_scores(self, question_embedding, actived_entities): + passage_weights = np.zeros(len(self.graph.vs["name"])) + dpr_passage_indices, dpr_passage_scores = self.dense_passage_retrieval(question_embedding) + dpr_passage_scores = min_max_normalize(dpr_passage_scores) + for i, dpr_passage_index in enumerate(dpr_passage_indices): + total_entity_bonus = 0 + passage_hash_id = self.passage_embedding_store.hash_ids[dpr_passage_index] + dpr_passage_score = dpr_passage_scores[i] + passage_text_lower = self.passage_embedding_store.hash_id_to_text[passage_hash_id].lower() + for entity_hash_id, (entity_id, entity_score, tier) in actived_entities.items(): + entity_lower = self.entity_embedding_store.hash_id_to_text[entity_hash_id].lower() + entity_occurrences = passage_text_lower.count(entity_lower) + if entity_occurrences > 0: + denom = tier if tier >= 1 else 1 + entity_bonus = entity_score * math.log(1 + entity_occurrences) / denom + total_entity_bonus += entity_bonus + passage_score = self.config.passage_ratio * dpr_passage_score + math.log(1 + total_entity_bonus) + passage_node_idx = self.node_name_to_vertex_idx[passage_hash_id] + passage_weights[passage_node_idx] = passage_score * self.config.passage_node_weight + return passage_weights + + def dense_passage_retrieval(self, question_embedding): + question_emb = question_embedding.reshape(1, -1) + question_passage_similarities = np.dot(self.passage_embeddings, question_emb.T).flatten() + sorted_passage_indices = np.argsort(question_passage_similarities)[::-1] + sorted_passage_scores = question_passage_similarities[sorted_passage_indices].tolist() + return sorted_passage_indices, sorted_passage_scores + + def get_seed_entities(self, question): + question_entities = list(self.spacy_ner.question_ner(question)) + if len(question_entities) == 0: + return [],[],[],[] + question_entity_embeddings = self.config.embedding_model.encode(question_entities,normalize_embeddings=True,show_progress_bar=False,batch_size=self.config.batch_size) + similarities = np.dot(self.entity_embeddings, question_entity_embeddings.T) + seed_entity_indices = [] + seed_entity_texts = [] + seed_entity_hash_ids = [] + seed_entity_scores = [] + for query_entity_idx in range(len(question_entities)): + entity_scores = similarities[:, query_entity_idx] + best_entity_idx = np.argmax(entity_scores) + best_entity_score = entity_scores[best_entity_idx] + best_entity_hash_id = self.entity_hash_ids[best_entity_idx] + best_entity_text = self.entity_embedding_store.hash_id_to_text[best_entity_hash_id] + seed_entity_indices.append(best_entity_idx) + seed_entity_texts.append(best_entity_text) + seed_entity_hash_ids.append(best_entity_hash_id) + seed_entity_scores.append(best_entity_score) + return seed_entity_indices, seed_entity_texts, seed_entity_hash_ids, seed_entity_scores + + def index(self, passages): + self.node_to_node_stats = defaultdict(dict) + self.entity_to_sentence_stats = defaultdict(dict) + self.passage_embedding_store.insert_text(passages) + hash_id_to_passage = self.passage_embedding_store.get_hash_id_to_text() + existing_passage_hash_id_to_entities,existing_sentence_to_entities, new_passage_hash_ids = self.load_existing_data(hash_id_to_passage.keys()) + if len(new_passage_hash_ids) > 0: + new_hash_id_to_passage = {k : hash_id_to_passage[k] for k in new_passage_hash_ids} + new_passage_hash_id_to_entities,new_sentence_to_entities = self.spacy_ner.batch_ner(new_hash_id_to_passage, self.config.max_workers) + self.merge_ner_results(existing_passage_hash_id_to_entities, existing_sentence_to_entities, new_passage_hash_id_to_entities, new_sentence_to_entities) + self.save_ner_results(existing_passage_hash_id_to_entities, existing_sentence_to_entities) + entity_nodes, sentence_nodes,passage_hash_id_to_entities,self.entity_to_sentence,self.sentence_to_entity = self.extract_nodes_and_edges(existing_passage_hash_id_to_entities, existing_sentence_to_entities) + self.sentence_embedding_store.insert_text(list(sentence_nodes)) + self.entity_embedding_store.insert_text(list(entity_nodes)) + self.entity_hash_id_to_sentence_hash_ids = {} + for entity, sentence in self.entity_to_sentence.items(): + entity_hash_id = self.entity_embedding_store.text_to_hash_id[entity] + self.entity_hash_id_to_sentence_hash_ids[entity_hash_id] = [self.sentence_embedding_store.text_to_hash_id[s] for s in sentence] + self.sentence_hash_id_to_entity_hash_ids = {} + for sentence, entities in self.sentence_to_entity.items(): + sentence_hash_id = self.sentence_embedding_store.text_to_hash_id[sentence] + self.sentence_hash_id_to_entity_hash_ids[sentence_hash_id] = [self.entity_embedding_store.text_to_hash_id[e] for e in entities] + self.add_entity_to_passage_edges(passage_hash_id_to_entities) + self.add_adjacent_passage_edges() + self.augment_graph() + output_graphml_path = os.path.join(self.config.working_dir,self.dataset_name, "LinearRAG.graphml") + os.makedirs(os.path.dirname(output_graphml_path), exist_ok=True) + self.graph.write_graphml(output_graphml_path) + + def add_adjacent_passage_edges(self): + passage_id_to_text = self.passage_embedding_store.get_hash_id_to_text() + index_pattern = re.compile(r'^(\d+):') + indexed_items = [ + (int(match.group(1)), node_key) + for node_key, text in passage_id_to_text.items() + if (match := index_pattern.match(text.strip())) + ] + indexed_items.sort(key=lambda x: x[0]) + for i in range(len(indexed_items) - 1): + current_node = indexed_items[i][1] + next_node = indexed_items[i + 1][1] + self.node_to_node_stats[current_node][next_node] = 1.0 + + def augment_graph(self): + self.add_nodes() + self.add_edges() + + def add_nodes(self): + existing_nodes = {v["name"]: v for v in self.graph.vs if "name" in v.attributes()} + entity_hash_id_to_text = self.entity_embedding_store.get_hash_id_to_text() + passage_hash_id_to_text = self.passage_embedding_store.get_hash_id_to_text() + all_hash_id_to_text = {**entity_hash_id_to_text, **passage_hash_id_to_text} + + passage_hash_ids = set(passage_hash_id_to_text.keys()) + + for hash_id, text in all_hash_id_to_text.items(): + if hash_id not in existing_nodes: + self.graph.add_vertex(name=hash_id, content=text) + + self.node_name_to_vertex_idx = {v["name"]: v.index for v in self.graph.vs if "name" in v.attributes()} + self.passage_node_indices = [ + self.node_name_to_vertex_idx[passage_id] + for passage_id in passage_hash_ids + if passage_id in self.node_name_to_vertex_idx + ] + + def add_edges(self): + existing_edges = set() + for edge in self.graph.es: + source_name = self.graph.vs[edge.source]["name"] + target_name = self.graph.vs[edge.target]["name"] + existing_edges.add(frozenset([source_name, target_name])) + + new_edges = [] + new_weights = [] + + for node_hash_id, node_to_node_stats in self.node_to_node_stats.items(): + for neighbor_hash_id, weight in node_to_node_stats.items(): + if node_hash_id == neighbor_hash_id: + continue + edge_key = frozenset([node_hash_id, neighbor_hash_id]) + if edge_key not in existing_edges: + new_edges.append((node_hash_id, neighbor_hash_id)) + new_weights.append(weight) + existing_edges.add(edge_key) + + if new_edges: + self.graph.add_edges(new_edges) + start_idx = len(self.graph.es) - len(new_edges) + for i, weight in enumerate(new_weights): + self.graph.es[start_idx + i]['weight'] = weight + + def add_entity_to_passage_edges(self, passage_hash_id_to_entities): + passage_to_entity_count ={} + passage_to_all_score = defaultdict(int) + for passage_hash_id, entities in passage_hash_id_to_entities.items(): + passage = self.passage_embedding_store.hash_id_to_text[passage_hash_id] + for entity in entities: + entity_hash_id = self.entity_embedding_store.text_to_hash_id[entity] + count = passage.count(entity) + passage_to_entity_count[(passage_hash_id, entity_hash_id)] = count + passage_to_all_score[passage_hash_id] += count + for (passage_hash_id, entity_hash_id), count in passage_to_entity_count.items(): + score = count / passage_to_all_score[passage_hash_id] + self.node_to_node_stats[passage_hash_id][entity_hash_id] = score + + def extract_nodes_and_edges(self, existing_passage_hash_id_to_entities, existing_sentence_to_entities): + entity_nodes = set() + sentence_nodes = set() + passage_hash_id_to_entities = defaultdict(set) + entity_to_sentence= defaultdict(set) + sentence_to_entity = defaultdict(set) + for passage_hash_id, entities in existing_passage_hash_id_to_entities.items(): + for entity in entities: + entity_nodes.add(entity) + passage_hash_id_to_entities[passage_hash_id].add(entity) + for sentence,entities in existing_sentence_to_entities.items(): + sentence_nodes.add(sentence) + for entity in entities: + entity_to_sentence[entity].add(sentence) + sentence_to_entity[sentence].add(entity) + return entity_nodes, sentence_nodes, passage_hash_id_to_entities, entity_to_sentence, sentence_to_entity + + def merge_ner_results(self, existing_passage_hash_id_to_entities, existing_sentence_to_entities, new_passage_hash_id_to_entities, new_sentence_to_entities): + existing_passage_hash_id_to_entities.update(new_passage_hash_id_to_entities) + existing_sentence_to_entities.update(new_sentence_to_entities) + return existing_passage_hash_id_to_entities, existing_sentence_to_entities + + def save_ner_results(self, existing_passage_hash_id_to_entities, existing_sentence_to_entities): + with open(self.ner_results_path, "w") as f: + json.dump({"passage_hash_id_to_entities": existing_passage_hash_id_to_entities, "sentence_to_entities": existing_sentence_to_entities}, f) -- cgit v1.2.3