Milvex.Index
(milvex v0.8.0)
Copy Markdown
Builder for Milvus index configurations.
Provides a fluent API for constructing index definitions with validation. Supports all common Milvus index types and distance metrics.
Examples
# Using builder
index = Index.new("embedding", :hnsw, :cosine)
|> Index.name("my_index")
|> Index.params(%{M: 16, efConstruction: 256})
# Using smart constructors
Index.flat("embedding", :l2)
Index.ivf_flat("embedding", :ip, nlist: 1024)
Index.hnsw("embedding", :cosine, m: 16, ef_construction: 256)
Index.autoindex("embedding", :l2)Summary
Functions
Creates an AUTOINDEX.
Creates a DiskANN index.
Creates a FLAT index (brute-force search).
Creates an HNSW index.
Returns list of all supported index types.
Creates an IVF_FLAT index.
Creates an IVF_PQ index.
Creates an IVF_SQ8 index.
Returns list of all supported metric types.
Sets the index name.
Creates a new index configuration.
Sets or merges additional index parameters.
Creates a SCANN index.
Creates a SPARSE_INVERTED_INDEX for BM25 full-text search.
Converts the index configuration to protobuf extra_params.
Validates the index configuration.
Validates the index and raises on error.
Types
@type index_type() ::
:flat
| :ivf_flat
| :ivf_sq8
| :ivf_pq
| :hnsw
| :autoindex
| :diskann
| :gpu_ivf_flat
| :gpu_ivf_pq
| :scann
| :sparse_inverted_index
@type inverted_index_algo() :: :daat_maxscore | :daat_wand | :taat_naive
@type metric_type() ::
:l2
| :ip
| :cosine
| :hamming
| :jaccard
| :max_sim_cosine
| :max_sim_ip
| :bm25
@type t() :: %Milvex.Index{ field_name: String.t(), index_type: index_type(), metric_type: metric_type(), name: String.t() | nil, params: map() }
Functions
@spec autoindex(String.t(), metric_type(), keyword()) :: t()
Creates an AUTOINDEX.
Lets Milvus automatically choose the best index type and parameters based on data characteristics.
Options
:name- Index name (optional)
Examples
Index.autoindex("embedding", :l2)@spec diskann(String.t(), metric_type(), keyword()) :: t()
Creates a DiskANN index.
Disk-based ANN index for very large datasets that don't fit in memory.
Options
:name- Index name (optional)
Examples
Index.diskann("embedding", :l2)@spec flat(String.t(), metric_type()) :: t()
Creates a FLAT index (brute-force search).
FLAT provides 100% recall but is slower for large datasets. Best for small datasets or when perfect accuracy is required.
Examples
Index.flat("embedding", :l2)
Index.flat("embedding", :cosine)@spec hnsw(String.t(), metric_type(), keyword()) :: t()
Creates an HNSW index.
Hierarchical Navigable Small World graph. Excellent performance for high-dimensional vectors with good recall.
Options
:m- Maximum number of connections per node (default: 16):ef_construction- Search depth during index building (default: 256):name- Index name (optional)
Examples
Index.hnsw("embedding", :cosine)
Index.hnsw("embedding", :l2, m: 32, ef_construction: 512)@spec index_types() :: [index_type()]
Returns list of all supported index types.
@spec ivf_flat(String.t(), metric_type(), keyword()) :: t()
Creates an IVF_FLAT index.
Inverted File index with flat quantization. Good balance of speed and accuracy for medium-sized datasets.
Options
:nlist- Number of cluster units (default: 1024):name- Index name (optional)
Examples
Index.ivf_flat("embedding", :l2)
Index.ivf_flat("embedding", :ip, nlist: 2048)@spec ivf_pq(String.t(), metric_type(), keyword()) :: t()
Creates an IVF_PQ index.
IVF with product quantization. Very memory-efficient but with lower accuracy. Best for very large datasets.
Options
:nlist- Number of cluster units (default: 1024):m- Number of subquantizers (default: 8):nbits- Bits per subquantizer (default: 8):name- Index name (optional)
Examples
Index.ivf_pq("embedding", :l2)
Index.ivf_pq("embedding", :ip, nlist: 2048, m: 16)@spec ivf_sq8(String.t(), metric_type(), keyword()) :: t()
Creates an IVF_SQ8 index.
IVF with scalar quantization. More memory-efficient than IVF_FLAT with slight accuracy trade-off.
Options
:nlist- Number of cluster units (default: 1024):name- Index name (optional)
Examples
Index.ivf_sq8("embedding", :l2)
Index.ivf_sq8("embedding", :ip, nlist: 2048)@spec metric_types() :: [metric_type()]
Returns list of all supported metric types.
Sets the index name.
If not set, Milvus will auto-generate a name.
@spec new(String.t(), index_type(), metric_type()) :: t()
Creates a new index configuration.
Parameters
field_name- Name of the vector field to indexindex_type- Type of index to createmetric_type- Distance metric to use
Examples
Index.new("embedding", :hnsw, :cosine)
Index.new("vectors", :ivf_flat, :l2)Sets or merges additional index parameters.
Parameters are index-type specific. Common parameters include:
- IVF indexes:
nlist(number of cluster units) - HNSW:
M(max connections),efConstruction(search depth during build) - IVF_PQ:
m(number of subquantizers),nbits(bits per subquantizer)
@spec scann(String.t(), metric_type(), keyword()) :: t()
Creates a SCANN index.
Google's ScaNN (Scalable Nearest Neighbors) implementation. Good balance of speed and accuracy.
Options
:nlist- Number of cluster units (default: 1024):name- Index name (optional)
Examples
Index.scann("embedding", :l2)
Index.scann("embedding", :cosine, nlist: 2048)Creates a SPARSE_INVERTED_INDEX for BM25 full-text search.
Used for SPARSE_FLOAT_VECTOR fields that receive BM25 function output. Supports inverted index algorithms optimized for sparse vectors.
Options
:inverted_index_algo- Algorithm to use (default::daat_maxscore):daat_maxscore- Document-at-a-time with MaxScore optimization:daat_wand- Document-at-a-time with WAND optimization:taat_naive- Term-at-a-time naive approach
:bm25_k1- BM25 k1 parameter (default: 1.2):bm25_b- BM25 b parameter (default: 0.75):drop_ratio_build- Drop ratio during index build (default: 0.2):name- Index name (optional)
Examples
Index.sparse_bm25("text_sparse")
Index.sparse_bm25("text_sparse", inverted_index_algo: :daat_wand)
Index.sparse_bm25("text_sparse", bm25_k1: 1.5, bm25_b: 0.8)
Index.sparse_bm25("text_sparse", drop_ratio_build: 0.1)@spec to_extra_params(t()) :: [Milvex.Milvus.Proto.Common.KeyValuePair.t()]
Converts the index configuration to protobuf extra_params.
Returns a list of KeyValuePair structs for use in CreateIndexRequest.
@spec validate(t()) :: {:ok, t()} | {:error, Milvex.Error.t()}
Validates the index configuration.
Returns {:ok, index} if valid, {:error, error} otherwise.
Validates the index and raises on error.