Skip to content

Latest commit

 

History

History
63 lines (44 loc) · 2.38 KB

CATEGORIES.md

File metadata and controls

63 lines (44 loc) · 2.38 KB

cat_hnswlib - Fork of the Hnswlib with support of categorical filtering.

Motivation

New categorical methods:

  • add_tags(labels, tag) - assign tag to the specified labels
  • get_tags(label) - returns list of tags, assigned to the label
  • reset_tags() - drop all tag-related information including additionaly built links
  • index_tagged(tag, m) - build additional navigation graph among tagged points with tag. Ensure connectiviti of conditional search
  • index_cross_tagged(tags, m) - build additional navigation graph among tagged points with tags. Does not create new entrypoints. Useful for creating geo-index and numerical ranges.
  • knn_query(data, k = 1, num_threads = -1, conditions = []) - extended with parameret conditions. It defines what points to include in search results. Performs traversal starting from the first point which fulfills condition. Example (A | !B) & C is represented as [[(0, A), (1, B)], [(0, C)]], where A, B, C loginal clauses if respective tag is assigned to a point. [[(0, 55)]] - means find closest point with tag 55.

Install

pip install --no-binary :all: 'git+https://github.com/generall/cat_hnswlib.git#subdirectory=python_bindings'

Example

import hnswlib
import numpy as np
from collections import defaultdict
import tqdm


dim = 50
elements = 10_000
parts_count = 100

hnsw = hnswlib.Index(space='cosine', dim=dim)
hnsw.init_index(max_elements = elements, ef_construction = 10, M = 16, random_seed=45)

points = np.random.rand(elements, dim)
hnsw.add_items(points)


# Assign tags by divisibility, for example
tags = defaultdict(list)
for i in range(elements):
    tags[i % parts_count].append(i)

for tag, ids in tqdm.tqdm(tags.items()):
    hnsw.add_tags(ids, tag)
    hnsw.index_tagged(tag, m=8)

target = np.float32(np.random.random((1, dim)))
condition = [[(False, 66)]]

# Result will only include points with id % 66 == 0
found_labels, found_dist = hnsw.knn_query(target, k=10, conditions=condition)

ToDo

  • Query planner
    • decide on search strategy depending on the amount of data-points covered by given condition
      • If number of points is small - use full scan
      • If there is large categories - search in them with graph
      • if there is a large number of small or-conditioned categories - search independently and parallel