llama-index-vector-stores-ngt

LlamaIndex vector store integration for NGT (Neighborhood Graph and Tree).

Installation

pip install llama-index-vector-stores-ngt

Parameters

NGTVectorStore accepts the following parameters:

• index_path: Directory path where NGT persists its index.
• dimension: Dimensionality of the vectors to be stored.
• distance_type: Distance function. Default: "L2".
  • Available values: "L1", "L2", "Normalized L2", "Angle", "Normalized Angle", "Cosine", "Normalized Cosine", "Hamming", "Jaccard", "Inner Product".
• edge_size_for_creation: Number of edges used when building the graph. Default: 10.
• edge_size_for_search: Number of edges explored during search. Default: 40.
• epsilon: Exploration range for approximate search. Larger values are more accurate but slower. Default: 0.1.

Usage

from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.core.schema import TextNode
from llama_index.vector_stores.ngt import NGTVectorStore

vector_store = NGTVectorStore(
    index_path="ngt_index",
    dimension=128,
    distance_type="L2",
)

storage_context = StorageContext.from_defaults(vector_store=vector_store)

nodes = [
    TextNode(text="apple is red and sweet"),
    TextNode(text="banana is yellow and sweet"),
]

index = VectorStoreIndex(nodes, storage_context=storage_context)
retriever = index.as_retriever(similarity_top_k=2)
results = retriever.retrieve("yellow fruit")

for node in results:
    print(node.score, node.text)

Query options

• similarity_top_k: Number of nearest neighbors to retrieve. Pass it to as_retriever() or as_query_engine().
• epsilon: Override the exploration range for a single query. Set vector_store.epsilon before querying, or pass epsilon directly to vector_store.query().
• edge_size: Override the search graph exploration size. Set vector_store.edge_size_for_search before querying, or pass edge_size directly to vector_store.query().

Persistence

Call persist() before exiting to save both the NGT index and LlamaIndex metadata.

vector_store.persist(persist_dir="ngt_index_persisted")

Later, or in another process, reopen the persisted index:

vector_store = NGTVectorStore(
    index_path="ngt_index_persisted",
    dimension=128,
    distance_type="L2",
)

storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex([], storage_context=storage_context)

results = index.as_retriever(similarity_top_k=2).retrieve("yellow fruit")

Use persist_dir="" to save into the original index_path instead.

Notes

• NGT stores only vectors and integer IDs. Text and metadata are kept by the wrapper and serialised alongside the NGT index.
• Metadata filtering is not supported by NGT itself. If you need it, pre-filter IDs on the Python side or combine with another metadata store.