wiki_mpnet_index.py

import faiss
import json
import numpy as np
import zipfile

from sentence_transformers import SentenceTransformer


class Retriever:
    def __init__(self):
        self.archive = zipfile.ZipFile('data/en/paragraphs.zip', 'r')
        self.index = faiss.read_index("data/en/embs_IVF16384_HNSW32_2lvl_full.idx")
        self.index.nprobe = 128
        self.model = SentenceTransformer('sentence-transformers/all-mpnet-base-v2', device='cuda')
        self.model.max_seq_length = 512


    def get_paragraph_by_vec_idx(self, vec_idx):
        chunk_id = vec_idx // 100000
        line_id = vec_idx % 100000
        with self.archive.open('enwiki_paragraphs_clean/enwiki_paragraphs_%03d.jsonl' % chunk_id) as f:
            for i,l in enumerate(f):
                if i == line_id:
                    data = json.loads(l)
                    break
        return data


    def search(self, query, k=5):
        emb = self.model.encode(query)
        _, neighbors = self.index.search(emb[np.newaxis, ...], k)
        results = []
        for n in neighbors[0]:
            data = get_paragraph_by_vec_idx(n)
            results.append(data)
        return results