README.md

AutoFaiss

Automatically create Faiss knn indices with the most optimal similarity search parameters.

It selects the best indexing parameters to achieve the highest recalls given memory and query speed constraints.

Doc and posts and notebooks

Using faiss efficient indices, binary search, and heuristics, Autofaiss makes it possible to automatically build in 3 hours a large (200 million vectors, 1TB) KNN index in a low amount of memory (15 GB) with latency in milliseconds (10ms).

Get started by running this colab notebook, then check the full documentation.
Get some insights on the automatic index selection function with this colab notebook.

Then you can check our multimodal search example (using OpenAI Clip model).

Read the medium post to learn more about it!

Installation

To install run pip install autofaiss

It's probably best to create a virtual env:

python -m venv .venv/autofaiss_env
source .venv/autofaiss_env/bin/activate
pip install -U pip
pip install autofaiss

Using autofaiss in python

If you want to use autofaiss directly from python, check the API documentation and the examples

In particular you can use autofaiss with on memory or on disk embeddings collections:

Using in memory numpy arrays

If you have a few embeddings, you can use autofaiss with in memory numpy arrays:

from autofaiss import build_index
import numpy as np

embeddings = np.float32(np.random.rand(100, 512))
index, index_infos = build_index(embeddings, save_on_disk=False)

query = np.float32(np.random.rand(1, 512))
_, I = index.search(query, 1)
print(I)

Using numpy arrays saved as .npy files

If you have many embeddings file, it is preferred to save them on disk as .npy files then use autofaiss like this:

from autofaiss import build_index

build_index(embeddings="embeddings", index_path="my_index_folder/knn.index",
            index_infos_path="my_index_folder/index_infos.json", max_index_memory_usage="4G",
            current_memory_available="4G")

Memory-mapped indices

Faiss makes it possible to use memory-mapped indices. This is useful when you don't need a fast search time (>50ms) and still want to reduce the memory footprint to the minimum.

We provide the should_be_memory_mappable boolean in build_index function to generate memory-mapped indices only. Note: Only IVF indices can be memory-mapped in faiss, so the output index will be a IVF index.

To load an index in memory mapping mode, use the following code:

import faiss
faiss.read_index("my_index_folder/knn.index", faiss.IO_FLAG_MMAP | faiss.IO_FLAG_READ_ONLY)

You can have a look to the examples to see how to use it.

Technical note: You can create a direct map on IVF indices with index.make_direct_map() (or directly from the build_index function by passing the make_direct_map boolean). Doing this speeds up a lot the .reconstruct() method, function that gives you the value of one of your vector given its rank. However, this mapping will be stored in RAM... We advise you to create your own direct map in a memory-mapped numpy array and then call .reconstruct_from_offset() with your custom direct_map.

Using autofaiss with pyspark

Autofaiss allows you to build indices with Spark for the following two use cases:

• To build a big index in a distributed way
• Given a partitioned dataset of embeddings, building one index per partition in parallel and in a distributed way.

Prerequisities:

1. Install pyspark: pip install pyspark.
2. Prepare your embeddings files (partitioned or not).
3. Create a Spark session before calling autofaiss. If no Spark session exists, a default session will be creaed with a minimum configuration.

Creating a big index in a distributed way

See distributed_autofaiss.md for a complete guide.

It is possible to generate an index that would require more memory than what's available. To do so, you can control the number of index splits that will compose your index with nb_indices_to_keep. For example, if nb_indices_to_keep is 10 and index_path is knn.index, the final index will be decomposed into 10 smaller indexes:

• knn.index01
• knn.index02
• knn.index03
• ...
• knn.index10

A concrete example shows how to produce N indices and how to use them.

Creating partitioned indexes

Given a partitioned dataset of embeddings, it is possible to create one index per partition by calling the method build_partitioned_indexes.

See this example that shows how to create partitioned indexes.

Using the command line

Create embeddings

import os
import numpy as np
embeddings = np.random.rand(1000, 100)
os.mkdir("embeddings")
np.save("embeddings/part1.npy", embeddings)
os.mkdir("my_index_folder")

Generate a Knn index

autofaiss build_index --embeddings="embeddings" --index_path="my_index_folder/knn.index" --index_infos_path="my_index_folder/index_infos.json" --metric_type="ip"

Try the index

import faiss
import glob
import numpy as np

my_index = faiss.read_index(glob.glob("my_index_folder/*.index")[0])

query_vector = np.float32(np.random.rand(1, 100))
k = 5
distances, indices = my_index.search(query_vector, k)

print(list(zip(distances[0], indices[0])))

How are indices selected ?

To understand better why indices are selected and what are their characteristics, check the index selection demo

Command quick overview

Quick description of the autofaiss build_index command:

embeddings -> Source path of the embeddings in numpy.
index_path -> Destination path of the created index. index_infos_path -> Destination path of the index infos. save_on_disk -> Save the index on the disk. metric_type -> Similarity distance for the queries.

index_key -> (optional) Describe the index to build.
index_param -> (optional) Describe the hyperparameters of the index.
current_memory_available -> (optional) Describe the amount of memory available on the machine.
use_gpu -> (optional) Whether to use GPU or not (not tested).

Command details

The autofaiss build_index command takes the following parameters:

Flag available	Default	Description
--embeddings	required	directory (or list of directories) containing your .npy embedding files. If there are several files, they are read in the lexicographical order. This can be a local path or a path in another Filesystem e.g. hdfs://root/... or s3://...
--index_path	required	Destination path of the faiss index on local machine.
--index_infos_path	required	Destination path of the faiss index infos on local machine.
--save_on_disk	required	Save the index on the disk.
--file_format	"npy"	File format of the files in embeddings Can be either npy for numpy matrix files or parquet for parquet serialized tables
--embedding_column_name	"embeddings"	Only necessary when file_format=parquet In this case this is the name of the column containing the embeddings (one vector per row)
--id_columns	None	Can only be used when file_format=parquet. In this case these are the names of the columns containing the Ids of the vectors, and separate files will be generated to map these ids to indices in the KNN index
--ids_path	None	Only useful when id_columns is not None and file_format=parquet. This will be the path (in any filesystem) where the mapping files Ids->vector index will be store in parquet format
--metric_type	"ip"	(Optional) Similarity function used for query: ("ip" for inner product, "l2" for euclidian distance)
--max_index_memory_usage	"32GB"	(Optional) Maximum size in GB of the created index, this bound is strict.
--current_memory_available	"32GB"	(Optional) Memory available (in GB) on the machine creating the index, having more memory is a boost because it reduces the swipe between RAM and disk.
--max_index_query_time_ms	10	(Optional) Bound on the query time for KNN search, this bound is approximative.
--min_nearest_neighbors_to_retrieve	20	(Optional) Minimum number of nearest neighbors to retrieve when querying the index. Parameter used only during index hyperparameter finetuning step, it is not taken into account to select the indexing algorithm. This parameter has the priority over the max_index_query_time_ms constraint.
--index_key	None	(Optional) If present, the Faiss index will be build using this description string in the index_factory, more detail in the Faiss documentation
--index_param	None	(Optional) If present, the Faiss index will be set using this description string of hyperparameters, more detail in the Faiss documentation
--use_gpu	False	(Optional) Experimental, gpu training can be faster, but this feature is not tested so far.
--nb_cores	None	(Optional) The number of cores to use, by default will use all cores
--make_direct_map	False	(Optional) Create a direct map allowing reconstruction of embeddings. This is only needed for IVF indices. Note that might increase the RAM usage (approximately 8GB for 1 billion embeddings).
--should_be_memory_mappable	False	(Optional) Boolean used to force the index to be selected among indices having an on-disk memory-mapping implementation.
--distributed	None	(Optional) If "pyspark", create the index using pyspark. Otherwise, the index is created on your local machine.
--temporary_indices_folder	"hdfs://root/tmp/distributed_autofaiss_indices"	(Optional) Folder to save the temporary small indices, only used when distributed = "pyspark"
--verbose	20	(Optional) Set verbosity of logging output: DEBUG=10, INFO=20, WARN=30, ERROR=40, CRITICAL=50
--nb_indices_to_keep	1	(Optional) Number of indices to keep at most when distributed is "pyspark".

Install from source

First, create a virtual env and install dependencies:

python3 -m venv .env
source .env/bin/activate
make install

python -m pytest -x -s -v tests -k "test_get_optimal_hyperparameters" to run a specific test