k-NN index

Preview

The k-NN plugin introduces a custom data type, the knn_vector, that allows users to ingest their k-NN vectors into an Lucenia index and perform different kinds of k-NN search. The knn_vector field is highly configurable and can serve many different k-NN workloads. For more information, see k-NN vector.

To create a k-NN index, set the settings.index.knn parameter to true:

PUT /test-index
{
  "settings": {
    "index": {
      "knn": true
    }
  },
  "mappings": {
    "properties": {
      "my_vector1": {
        "type": "knn_vector",
        "dimension": 3,
        "method": {
          "name": "hnsw",
          "space_type": "l2",
          "engine": "lucene",
          "parameters": {
            "ef_construction": 128,
            "m": 24
          }
        }
      }
    }
  }
}

Lucene byte vector

Starting with k-NN plugin version 2.9, you can use byte vectors with the lucene engine to reduce the amount of storage space needed. For more information, see Lucene byte vector.

SIMD optimization for the Faiss engine

Starting with version 2.13, the k-NN plugin supports Single Instruction Multiple Data (SIMD) processing if the underlying hardware supports SIMD instructions (AVX2 on x64 architecture and Neon on ARM64 architecture). SIMD is supported by default on Linux machines only for the Faiss engine. SIMD architecture helps boost overall performance by improving indexing throughput and reducing search latency.

SIMD optimization is applicable only if the vector dimension is a multiple of 8.

x64 architecture

For the x64 architecture, two different versions of the Faiss library are built and shipped with the artifact:

• libluceniaknn_faiss.so: The non-optimized Faiss library without SIMD instructions.
• libluceniaknn_faiss_avx2.so: The Faiss library that contains AVX2 SIMD instructions.

If your hardware supports AVX2, the k-NN plugin loads the libluceniaknn_faiss_avx2.so library at runtime.

To disable AVX2 and load the non-optimized Faiss library (libluceniaknn_faiss.so), specify the knn.faiss.avx2.disabled static setting as true in lucenia.yml (default is false). Note that to update a static setting, you must stop the cluster, change the setting, and restart the cluster. For more information, see Static settings.

ARM64 architecture

For the ARM64 architecture, only one performance-boosting Faiss library (libluceniaknn_faiss.so) is built and shipped. The library contains Neon SIMD instructions and cannot be disabled.

Method definitions

A method definition refers to the underlying configuration of the approximate k-NN algorithm you want to use. Method definitions are used to either create a knn_vector field (when the method does not require training) or create a model during training that can then be used to create a knn_vector field.

A method definition will always contain the name of the method, the space_type the method is built for, the engine (the library) to use, and a map of parameters.

Supported nmslib methods

HNSW parameters

For nmslib, ef_search is set in the index settings.

Supported Faiss methods

For hnsw, “innerproduct” is not available when PQ is used.

HNSW parameters

IVF parameters

For more information about setting these parameters, refer to the Faiss documentation.

IVF training requirements

The IVF algorithm requires a training step. To create an index that uses IVF, you need to train a model with the Train API, passing the IVF method definition. IVF requires that, at a minimum, there are nlist training data points, but it is recommended that you use more than this. Training data can be composed of either the same data that is going to be ingested or a separate dataset.

Supported Lucene methods

HNSW parameters

Lucene HNSW implementation ignores ef_search and dynamically sets it to the value of “k” in the search request. Therefore, there is no need to make settings for ef_search when using the Lucene engine.

"method": {
    "name":"hnsw",
    "engine":"lucene",
    "space_type": "l2",
    "parameters":{
        "m":2048,
        "ef_construction": 245
    }
}

Supported Faiss encoders

You can use encoders to reduce the memory footprint of a k-NN index at the expense of search accuracy. The k-NN plugin currently supports the flat, pq, and sq encoders in the Faiss library.

The following example method definition specifies the hnsw method and a pq encoder:

"method": {
  "name":"hnsw",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "encoder":{
      "name":"pq",
      "parameters":{
        "code_size": 8,
        "m": 8
      }
    }
  }
}

The hnsw method supports the pq encoder. The code_size parameter of a pq encoder with the hnsw method must be 8.

PQ parameters

SQ parameters

For more information and examples, see Using Faiss scalar quantization.

Examples

The following example uses the ivf method without specifying an encoder (by default, Lucenia uses the flat encoder):

"method": {
  "name":"ivf",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "nlist": 4,
    "nprobes": 2
  }
}

The following example uses the ivf method with a pq encoder:

"method": {
  "name":"ivf",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "encoder":{
      "name":"pq",
      "parameters":{
        "code_size": 8,
        "m": 8
      }
    }
  }
}

The following example uses the hnsw method without specifying an encoder (by default, Lucenia uses the flat encoder):

"method": {
  "name":"hnsw",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "ef_construction": 256,
    "m": 8
  }
}

The following example uses the hnsw method with an sq encoder of type fp16 with clip enabled:

"method": {
  "name":"hnsw",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "encoder": {
      "name": "sq",
      "parameters": {
        "type": "fp16",
        "clip": true
      }
    },
    "ef_construction": 256,
    "m": 8
  }
}

The following example uses the ivf method with an sq encoder of type fp16:

"method": {
  "name":"ivf",
  "engine":"faiss",
  "space_type": "l2",
  "parameters":{
    "encoder": {
      "name": "sq",
      "parameters": {
        "type": "fp16",
        "clip": false
      }
    },
    "nprobes": 2
  }
}

Choosing the right method

There are several options to choose from when building your knn_vector field. To determine the correct methods and parameters, you should first understand the requirements of your workload and what trade-offs you are willing to make. Factors to consider are (1) query latency, (2) query quality, (3) memory limits, and (4) indexing latency.

If memory is not a concern, HNSW offers a strong query latency/query quality trade-off.

If you want to use less memory and increase indexing speed as compared to HNSW while maintaining similar query quality, you should evaluate IVF.

If memory is a concern, consider adding a PQ encoder to your HNSW or IVF index. Because PQ is a lossy encoding, query quality will drop.

You can reduce the memory footprint by a factor of 2, with a minimal loss in search quality, by using the fp_16 encoder. If your vector dimensions are within the [-128, 127] byte range, we recommend using the byte quantizer to reduce the memory footprint by a factor of 4. To learn more about vector quantization options, see k-NN vector quantization.

Memory estimation

In a typical Lucenia cluster, a certain portion of RAM is reserved for the JVM heap. The k-NN plugin allocates native library indexes to a portion of the remaining RAM. This portion’s size is determined by the circuit_breaker_limit cluster setting. By default, the limit is set to 50%.

Having a replica doubles the total number of vectors.

For information about using memory estimation with vector quantization, see the vector quantization documentation.

HNSW memory estimation

The memory required for HNSW is estimated to be 1.1 * (4 * dimension + 8 * M) bytes/vector.

As an example, assume you have a million vectors with a dimension of 256 and M of 16. The memory requirement can be estimated as follows:

1.1 * (4 * 256 + 8 * 16) * 1,000,000 ~= 1.267 GB

IVF memory estimation

The memory required for IVF is estimated to be 1.1 * (((4 * dimension) * num_vectors) + (4 * nlist * d)) bytes.

As an example, assume you have a million vectors with a dimension of 256 and nlist of 128. The memory requirement can be estimated as follows:

1.1 * (((4 * 256) * 1,000,000) + (4 * 128 * 256))  ~= 1.126 GB

Index settings

Additionally, the k-NN plugin introduces several index settings that can be used to configure the k-NN structure as well.

At the moment, several parameters defined in the settings are in the deprecation process. Those parameters should be set in the mapping instead of the index settings. Parameters set in the mapping will override the parameters set in the index settings. Setting the parameters in the mapping allows an index to have multiple knn_vector fields with different parameters.