Understanding these concepts will help you utilize the Mixpeek Multimodal Warehouse offerings.
Mixpeek organizes data in a structured hierarchy designed for flexibility and performance of multimodal content processing and retrieval.
Mixpeek TermDescriptionData Warehouse Analogy
NamespaceQuery boundaries that isolate environmentsDatabase/Schema
BucketStorage containers for raw objects and filesRaw Data Lake/Storage Layer
ObjectCollections of related input filesRaw Data Files/Source Documents
BlobIndividual raw files within ObjectsBinary Data/Single File
CollectionGroups of processed documents with consistent schemaTable
DocumentStructured outputs from feature extractorsRow
Feature ExtractorSpecialized components that process inputs to extract specific featuresETL Process/Transformation
FeatureExtracted data elements stored in feature storesColumn/Field
Feature StoreSpecialized storage for extracted features optimized for efficient retrievalIndexed Columns/Materialized Views
RetrieverQuery engines that search feature stores to find relevant documentsSQL Query Engine
Retriever StageComponents of search pipelines that perform specific operations in the retrieval processQuery Execution Plan Step
TaxonomyMultimodal equivalent of SQL JOIN operationsJOIN Operation
ClusteringMultimodal equivalent of SQL GROUP BY operationsGROUP BY Operation
ResearchMulti-step process that explores topics through iterative searches, generates structured reports with sections, and combines retrieved information into cohesive contentBusiness Intelligence Report

Component Relationships

The different components in Mixpeek relate to each other in specific ways:

Understanding the Relationships

  • Buckets contain Objects; Objects are grouped into Batches for processing.
  • Submitting a Batch runs Feature Extractors to produce Documents and Feature Stores.
  • Documents (features + metadata) live in Collections.
  • Taxonomies enrich Documents; Clusters group Documents.
  • Retrievers query Feature Stores and return ranked Documents.

Processing Components

Feature Extractors

Specialized components that process inputs to extract features (embeddings, faces, scenes, objects, transcripts). They populate feature stores and produce documents in collections.

Retrievers

Configurable pipelines that combine KNN over feature stores, metadata filters, grouping, and reranking to return relevant documents.

SQL Analogs

  • JOIN → Taxonomies: similarity joins that enrich documents (flat or hierarchical).
  • GROUP BY → Clustering: groups similar documents; can output centroids/artifacts.

Metadata and Document Properties

All documents in Mixpeek collections include standard metadata properties:
{
  "document_id": "doc_123",           // Unique document identifier
  "collection_id": "col_456def",      // Collection this document belongs to
  "source_details": [                   // Lineage and processing info for this doc
    {
      "type": "bucket",               // Immediate origin type (bucket/collection/taxonomy/cluster)
      "source_id": "bkt_789",         // Immediate source identifier (e.g., bucket_id)
      "object_id": "obj_123abc",      // Originating object id
      "feature_extractor_task_id": "ray_task_fx_1", // Ray task that ran the extractor
      "object_task_id": "ray_task_obj_1",           // Ray task that orchestrated the object
      "feature_extractor_id": "clip_vit_l_14@1.0.0" // Extractor instance that produced features
    }
  ],
  "metadata": {},                       // User-provided metadata
  "internal_metadata": {},              // System-generated metadata
  "vector": null,                       // Optional vector embedding in this response
  "presigned_url": null,                // Optional signed URL for the primary asset
  "enrichments": null,                  // Optional taxonomy/cluster enrichments
  "source_blobs": []                    // Optional list of source blob descriptors
}

Next Steps

Now that you understand the core concepts of Mixpeek, you’re ready to start building with the platform: