Indexing vectors

This page explains how to store vectors in hashes. Hashes provide an efficient way to store vectors in Redis.

Data serialization

Before storing vectors in a Redis hash, vectors need to be converted into a format that Redis understands. It requires vector serialization into binary blobs where the size equals the data type's byte size (e.g., 4 for FLOAT32) multiplied by the vector's number of dimensions. A popular choice for numerical vectors is the Python NumPy library:

Connect to Redis

Before storing the vector in a hash, establish a connection to your Redis instance using a client like redis-py:

Store the vector in a hash

Redis hashes are like dictionaries, with key-value pairs. Use the Redis HSET command to store your serialized vector:

import numpy as np
import redis

# Sample vector
vector = np.array([1.2, 3.5, -0.8]) # 3-dimensional vector

# Serialize to a binary blob
serialized_vector = vector.tobytes()

redis_client = redis.Redis(host='your_redis_host', port=6379)

redis_client.hset('vector_storage', 'vector_key', serialized_vector)  # 'vector_key' is a unique identifier
  • For successful indexing, your vector data must adhere to the dimensions and data type set in the index schema.

Backfilling Indexes

Backfilling indexes may occur in one of the following scenarios:

  • Once an index is created, the backfilling procedure scans through Redis keyspace for entries that meet the index filter criteria.
  • Vector indexes and their data are persisted in RDB snapshots. When an RDB file is loaded, an automatic index backfilling process is triggered. This process actively detects and integrates any new or modified entries into the index since the RDB snapshot was created, maintaining index integrity and ensuring current results.