Mixpeek turns raw multimodal inputs—text, images, video, audio, PDFs—into structured documents, reusable features, and query-ready indexes. Skip 6-12 months of infrastructure work: Mixpeek replaces 15+ tools (vector DBs, ML orchestration, feature stores, search infrastructure) with a single API. No ops, no ML deployment, no pipeline code—just configuration.
What Mixpeek Replaces
ML Orchestration
Airflow, Prefect, Temporal, Kubeflow
Vector & Search
Qdrant, Pinecone, Weaviate, Elasticsearch
Feature Infrastructure
Feast, Tecton, custom feature stores
Zero Infrastructure, Maximum Power
No ML Ops Required
Deploy CLIP, Whisper, LayoutLM, or custom models without infrastructure. Mix extractors across collections. Model Registry handles versions, GPU allocation, and autoscaling. You write JSON configs, not Kubernetes manifests.
Retrievers = Semantic JOINs
Search is a JOIN operation using vector similarity instead of foreign keys. Chain stages (search → filter → rank → enrich → transform) into multi-hop graphs. Connect documents across collections—no ETL code required.
Versioned & Immutable
Taxonomies, extractors, and pipelines are versioned snapshots. A/B test retrieval strategies, roll back configs, and track complete lineage from result → document → object → source file.
Production-Grade Observability
ClickHouse analytics, multi-channel notifications (email, Slack, SMS, webhooks), audit logs, and real-time task monitoring. Built-in caching delivers sub-100ms retrieval latency.
Multi-Database Orchestration
One API, five databases underneath: MongoDB (metadata), Qdrant (vectors), ClickHouse (analytics), Redis (cache), S3 (files). Cross-database queries, transactions, and routing handled automatically.
Parallel Enrichment at Scale
Taxonomy enrichment runs as parallel joins across Qdrant collections. Process thousands of documents concurrently. Hierarchical taxonomies and flat tags use the same JOIN stage—just configuration.
Time Saved
| Task | Without Mixpeek | With Mixpeek |
|---|---|---|
| Setup vector search | 2-4 weeks | 5 minutes |
| Build feature extraction pipeline | 4-8 weeks | 1 hour |
| Implement multi-modal search | 6-12 weeks | 30 minutes |
| Add new ML model | 2-4 weeks | Config change |
| Multi-tenant architecture | 4-8 weeks | Built-in |
| Taxonomies & semantic joins | 4-8 weeks | 1 hour |
| Total engineering effort | 6-12 months | Days |
Core Workflows
1
Isolate tenants with namespaces
Provision namespaces (or use the default) to keep tenants and environments separate. Pass
X-Namespace on every authenticated request.2
3
Define collections and extract features
Create Collections to map object fields into feature extractor inputs. The Engine downloads per-extractor artifacts, runs Ray tasks, and writes documents plus vectors to Qdrant.
4
Enrich with taxonomies and clustering
Apply Taxonomies (flat or hierarchical) or Clusters to attach structured metadata. Both reuse the JOIN stage for fast, parallel enrichment.
5
Retrieve with multi-stage pipelines
Compose Retrievers from search, filter, rank, enrich, transform, and compose stages. Fetch presigned URLs, execution metrics, and cache-aware responses in a single API call.
Architecture Snapshot
Request Headers
Next Steps
- Follow the Quickstart to stand up Mixpeek locally or in the cloud
- Review Core Concepts for namespaces, objects, documents, and lineage patterns
- Study the complete Architecture and Caching strategies
- Explore ready-made Recipes or jump straight into the API Reference