from datasets import load_dataset from sentence_transformers import SentenceTransformer, CrossEncoder, util import torch from huggingface_hub import hf_hub_download embedding_path = "abokbot/wikipedia-embedding" def load_embedding(): print("Loading embedding...") path = hf_hub_download(repo_id="abokbot/wikipedia-embedding", filename="wikipedia_en_embedding.pt") wikipedia_embedding = torch.load(path, map_location=torch.device('cpu')) print("Embedding loaded!") return wikipedia_embedding wikipedia_embedding = load_embedding() def load_encoders(): print("Loading encoders...") bi_encoder = SentenceTransformer('msmarco-MiniLM-L-6-v3') bi_encoder.max_seq_length = 512 cross_encoder = CrossEncoder('cross-encoder/ms-marco-TinyBERT-L-2-v2') print("Encoders loaded!") return bi_encoder, cross_encoder bi_encoder, cross_encoder = load_encoders() def load_wikipedia_dataset(): print("Loading wikipedia dataset...") dataset = load_dataset("abokbot/wikipedia-first-paragraph")["train"] print("Dataset loaded!") return dataset dataset = load_wikipedia_dataset() def search(query): print("Input question:", query) ##### Semantic Search ##### print("Semantic Search") # Encode the query using the bi-encoder and find potentially relevant passages top_k = 32 question_embedding = bi_encoder.encode(query, convert_to_tensor=True) hits = util.semantic_search(question_embedding, wikipedia_embedding, top_k=top_k) hits = hits[0] # Get the hits for the first query ##### Re-Ranking ##### print("Re-Ranking") cross_inp = [[query, dataset[hit['corpus_id']]["text"]] for hit in hits] cross_scores = cross_encoder.predict(cross_inp) # Sort results by the cross-encoder scores for idx in range(len(cross_scores)): hits[idx]['cross-score'] = cross_scores[idx] hits = sorted(hits, key=lambda x: x['cross-score'], reverse=True) # Output of top-3 hits from re-ranker print("\n-------------------------\n") print("Top-3 Cross-Encoder Re-ranker hits") results = [] for hit in hits[:3]: results.append( { "score": round(hit['cross-score'], 3), "title": dataset[hit['corpus_id']]["title"], "abstract": dataset[hit['corpus_id']]["text"].replace("\n", " "), "link": dataset[hit['corpus_id']]["url"] } ) return results