Install what you need

pip install vectrixdb gives vector search and nothing else. Everything a deployment adds, the server, sign-in, document readers, cloud stores, comes as an extra, so a machine carries only what it uses.

The core

pip install vectrixdb

Python 3.9 or later. It brings five packages:

Package	For
numpy	Vectors
usearch	The vector index
onnxruntime, below 2	Running the models on this machine
rich, typer	The vectrixdb command

The wheel carries the English embedding model, the reranker, the sparse model and the late-interaction model, so a collection can be made, filled and searched in every mode with no network and no key. The multilingual models are fetched once with vectrixdb download-models; see Run without a network.

With the core alone you can open a Vectrix, add text and Markdown, HTML, CSV, PowerPoint and plain text documents, search, and run vectrixdb check and vectrixdb doctor.

Extras go in square brackets, several at once:

pip install "vectrixdb[api,signin,documents]"

The server and its parts

Extra	Adds	For
api	fastapi, uvicorn[standard], pydantic	The REST API and the dashboard: vectrixdb serve, and vectrixdb extract-serve
signin	pyjwt[crypto], qrcode	Signing people in, with VECTRIXDB_SIGNIN: single sign-on, email, authenticator apps and passkeys
mcp	mcp 2 or later, httpx	The server's /mcp endpoint with VECTRIXDB_MCP=1, and vectrixdb mcp
tracing	opentelemetry-api, opentelemetry-sdk, the OTLP HTTP exporter	A span for every search, ingestion and evaluation run; see Trace searches and ingestion
client	httpx	vectrixdb.connect, VectrixClient and AsyncVectrixClient, and the vectrixdb command on a server; see Use it from Python, TypeScript, Go or Rust
jobs-azure	azure-storage-blob, azure-storage-queue, azure-identity	The extraction service's long jobs on Azure, with VECTRIXDB_EXTRACT_JOBS=azure

A server for a team is usually api, signin, mcp and tracing, plus the readers and the store it needs. The container image installs api,signin,mcp,tracing,documents,feeds,azure,aws,postgres,jobs-azure.

Documents and recordings

The readers for Markdown, HTML, CSV, PowerPoint and text need nothing. The rest are named for what the file becomes.

Extra	Adds	Reads
documents	pypdf, pypdfium2, python-docx, openpyxl, olefile, xlrd	PDFs with a text layer, Word, Excel, and Word and Excel 97-2003
ocr	rapidocr-onnxruntime, pypdfium2, pillow, pypdf	Pictures and scanned PDF pages, on this machine
asr	faster-whisper	Recordings, on this machine. The speech model downloads the first time
ffmpeg	imageio-ffmpeg	ffmpeg alone, about 90 MB: the sound out of a video, and a long recording cut into pieces, for a host whose transcribing is done by a service
video	ffmpeg and faster-whisper	Videos, on this machine
extract	documents, ocr, asr and video	Every local reader
youtube	yt-dlp	YouTube addresses
feeds	feedparser	RSS 2.0, RSS 1.0 and Atom feeds a collection keeps up with. JSON Feed and pages need nothing
ocr-azure	azure-ai-documentintelligence	Pictures and scans through Azure Document Intelligence
asr-azure	nothing	Azure Speech is one HTTPS request, so it needs no package
ocr-aws, asr-aws	boto3	The same jobs through AWS
masking	presidio-analyzer, spacy	Masking names, addresses and national ids before a document is indexed, in this process. The language models are a separate step: vectrixdb download-models --type masking

vectrixdb check says which kinds of file this machine reads and which extra reads the rest; see Check and doctor.

Storage

Extra	Adds	For
azure	azure-cosmos, azure-search-documents, azure-identity, openai	Cosmos DB and Azure AI Search, keyless sign-in to Azure, the second vector from Azure OpenAI, and cosmos:// stores for sign-in, collection records and chunks
aws	opensearch-py, boto3, requests-aws4auth, psycopg2-binary	OpenSearch, Aurora PostgreSQL, dynamodb:// stores, s3:// addresses, Amazon Comprehend
postgres	psycopg2-binary	PostgreSQL, Lakebase, and postgresql:// stores for sign-in, collection records and audit
databricks	databricks-sql-connector	Delta Lake

See Choose a storage backend for each.

Models

Extra	Adds	For
hf	sentence-transformers	Embedding models from Hugging Face by name, through sentence-transformers
fastembed	fastembed	Embedding models through FastEmbed
embeddings	fastembed and sentence-transformers	Both
nlp	spacy	Entity extraction for graph mode on ordinary prose. Without it a pattern reader finds capitalised phrases, acronyms, quoted terms and repeated phrases. The model is a separate step: python -m spacy download en_core_web_sm
setup-models	torch, transformers, optimum[onnxruntime]	Converting a model to ONNX, which vectrixdb download-models asks for when it is missing

Tools

Extra	Adds	For
viz	umap-learn, scikit-learn	Nothing in VectrixDB 2.2 imports them; install them for code of your own
all	api, signin, mcp, documents, viz, hf, fastembed, aws, azure, databricks, postgres, tracing, feeds	Everything at run time except the local readers in extract, whose speech models are large enough to ask for by name

Which command needs which

You run	Install
Vectrix(...), add, search	the core
add_document on a PDF, Word or Excel file	documents
add_document on a picture, a scan, a recording or a video	ocr, asr or video, or an extraction service
vectrixdb serve	api, and signin with sign-in on, mcp with VECTRIXDB_MCP=1, tracing with tracing on
vectrixdb extract-serve	api and documents, with the readers and services it uses
vectrixdb mcp, vectrixdb-mcp	mcp
vectrixdb sources refresh on an XML feed	feeds
vectrixdb download-models --type masking	masking
vectrixdb.connect, VectrixClient, AsyncVectrixClient	client
vectrixdb.load_youtube	youtube
vectrixdb check, vectrixdb doctor	the core; each says which extra a setting needs

A call that needs an extra that is not installed raises DependencyError, which names the package and the extra: see Handle errors.