Use it from LangChain, LlamaIndex, any embedding API, or your own plugin¶

VectrixDB is a library, so the useful question is how it plugs into what you already run. Four answers.

LangChain¶

from vectrixdb.integrations.langchain import VectrixVectorStore

store = VectrixVectorStore("docs", path="./data")      # VectrixDB's bundled model, offline
store.add_texts(["Basalt forms when lava cools."], metadatas=[{"topic": "rocks"}])
store.similarity_search("volcanic rock", k=3)
retriever = store.as_retriever(search_kwargs={"k": 5, "mode": "hybrid"})

pip install langchain-core is the only extra. Pass embedding= with any LangChain Embeddings to use that model for documents and queries instead; the collection is created at its dimension. filter= takes the VectrixDB filter grammar, and every search() keyword (mode, rerank, token_budget, parents) passes through search_kwargs. store.db is the underlying Vectrix for memory, ingestion and everything else.

LlamaIndex¶

from llama_index.core import StorageContext, VectorStoreIndex
from vectrixdb.integrations.llamaindex import VectrixLlamaStore

store = VectrixLlamaStore("docs", path="./data")
index = VectorStoreIndex.from_documents(
    documents, storage_context=StorageContext.from_defaults(vector_store=store)
)

pip install llama-index-core. LlamaIndex embeds with its own embed_model and hands the vectors over, so the store never embeds and the collection takes that model's dimension on the first add. Deleting by ref_doc_id removes every node of the document.

Any OpenAI-compatible embedding endpoint¶

from vectrixdb import Vectrix

db = Vectrix("docs", dense_model="openai:text-embedding-3-small")     # api.openai.com, OPENAI_API_KEY

The same class reaches any server that speaks the OpenAI embeddings API:

from vectrixdb.models import OpenAIEmbedder

ollama = OpenAIEmbedder("nomic-embed-text", base_url="http://localhost:11434/v1")
db = Vectrix("docs", embed_fn=ollama, dimension=ollama.dimension)

Ollama, vLLM, LM Studio and Text Embeddings Inference all work this way. The dimension is asked of the endpoint once. This is the one embedder in the package that needs a network, and the model name is recorded in the collection like any other, so a later open with a different model warns.

Plugins¶

A package can add an embedder, reranker, extractor or storage backend by declaring an entry point; VectrixDB finds it by name and nothing in VectrixDB has to change.

# the plugin's pyproject.toml
[project.entry-points."vectrixdb.embedders"]
my-encoder = "my_package.encoders:MyEncoder"

[project.entry-points."vectrixdb.storage"]
my-store = "my_package.storage:MyStorage"

db = Vectrix("docs", dense_model="plugin:my-encoder")
config = StorageConfig(backend="plugin:my-store")

An embedder is anything with embed(texts) -> ndarray and, ideally, a dimension attribute. A storage backend subclasses vectrixdb.core.storage.BaseStorage; the storage contract suite in the repository is the definition of done. vectrixdb.plugins.available("embedders") lists what is installed, and register() adds one in code for tests or for a program that ships its own.

Timing hooks¶

Hosts meter their own way. Two callbacks give them the numbers without VectrixDB importing a metrics client:

def meter(event):
    statsd.timing(f"vectrixdb.{event['op']}", event["ms"], tags=[f"collection:{event['collection']}"])

db = Vectrix("docs", on_search=meter, on_add=meter)

Each event is a small dict: op (search or add), ms, count, mode and collection. A hook that raises is logged and never fails the call.

The command line¶

Every command and option is in the command line reference.

vectrixdb ingest ./docs --name manual --chunk markdown --parent-size 600 --dedupe 0.9
vectrixdb query "how do I export" --name manual --limit 5 --parents
vectrixdb stats --name manual
vectrixdb serve --port 7337

ingest takes files or directories (PDF, DOCX, HTML, Markdown, text) with every add_document() option; query prints a table or --json; stats shows count, mode, model and size on disk; serve starts the REST API and dashboard.