Features are fields inside documents produced by feature extractors. They include vectors and payload attributes that power retrieval, enrichment, and analytics.
Overview
- Where features live: Inside collection documents (not standalone). Documents are written by extractors during processing.
- What features are: Model outputs such as vectors (dense/sparse/multi) and structured payload fields.
- Configured by: The collection’s feature extractors and index definitions.
Feature anatomy (in a document)
A document stores features as payload fields and vectors. Exact paths depend on the collection’s extractor configuration and feature addresses.- Vectors are stored under configured vector field names and indexed per the collection’s vector indexes.
- Non-vector outputs (payload) are stored as normal document fields under
features
or other configured paths.
Describe features for a collection
Get feature addresses and metadata available in a collection.- API: Describe Collection Features
- Method: GET
- Path:
/v1/collections/{collection_identifier}/features
- Reference: API Reference
Discover feature extractors
Browse supported extractors and their input/output schemas.- List Extractors: API Reference
- Get Extractor by Name: API Reference
Used by
- Retrievers: Execute vector/payload search over features in documents (Retrievers).
- Taxonomies: Join/enrich using configured feature fields (Taxonomies).
- Clusters: Build similarity groups over feature vectors (Clusters).
Best practices
- Pick extractors per modality: Text, images, audio, and video often need different models.
- Version your models: Prefer new collections or reprocessing when changing model versions.
- Name feature fields clearly: Stable field names simplify retriever and taxonomy configs.