Skip to main content

​
General

Mixpeek is a multimodal data processing and retrieval platform. It ingests raw files (video, images, audio, PDFs, text), extracts features (embeddings, transcriptions, structured data), and enables semantic search across all modalities through a unified API.
Objects are raw inputs registered in buckets (e.g., a video file, PDF, JSON payload). They’re validated but not processed.Documents are processed outputs created by feature extractors. They live in collections, include vectors/embeddings, and are queryable via retrievers.Flow: Object → Batch → Engine → Document
No. Mixpeek abstracts away model selection, infrastructure, and scaling. You declare what features you want (e.g., text embeddings, scene detection) via simple JSON configurations. We handle training, hosting, and optimization.
Custom models are available for Enterprise customers. Contact us via “Talk to Engineers” to discuss integration options.

​
Ingestion & Processing

Depends on:
  • Object count: 100 objects typically process in 1-5 minutes
  • File size: Large videos (>1GB) take longer
  • Extractors: Video/audio extractors are slower than text
  • Ray cluster size: More workers = faster throughput
Monitor progress via Task API:
GET /v1/tasks/{task_id}
Partial results are written to Qdrant. Documents will have:
  • __fully_enriched: false
  • __missing_features: ["list", "of", "failed", "features"]
You can:
  1. Query partial results (filter by __fully_enriched: true)
  2. Reprocess failed objects individually
  3. Inspect Engine logs for failure reasons
Yes. Use the Collection Documents API:
PATCH /v1/collections/{collection_id}/documents/{document_id}
{ "metadata": { "status": "reviewed" } }
Or batch update:
POST /v1/collections/{collection_id}/documents/batch-update
Note: Updating metadata doesn’t re-run feature extractors. To reprocess, create a new batch with the source object.
Two approaches:1. Delete specific documents:
DELETE /v1/collections/{collection_id}/documents/{document_id}
2. Delete all documents from an object:
DELETE /v1/buckets/{bucket_id}/objects/{object_id}?cascade=true
This removes the object and all derived documents across collections.
TypeFormats
VideoMP4, MOV, AVI, MKV, WebM
ImageJPEG, PNG, GIF, WebP, TIFF
AudioMP3, WAV, FLAC, OGG, M4A
DocumentPDF, DOCX, TXT, Markdown
StructuredJSON, CSV
For unsupported formats, pre-convert or contact support for custom extractors.

​
Search & Retrieval

Unlimited. Specify multiple collection IDs:
{ "collection_ids": ["col_text", "col_images", "col_videos"] }
Results are fused across collections based on stage configurations.
No. The X-Namespace header enforces hard isolation. Each request operates within a single namespace.Workaround: Make separate requests per namespace and merge results client-side.
  1. Create a collection with image_extractor
  2. Build a retriever with KNN search on clip_embedding
  3. Pass image URL in inputs:
POST /v1/retrievers/{retriever_id}/execute
{
  "inputs": {
    "query_image": "s3://my-bucket/sample.jpg"
  }
}
Per query: 10,000 documents (pagination required for larger result sets)Per stage: Configurable limit parameter (e.g., retrieve 100 candidates, rerank top 20)Best practice: Use filters and sorts to narrow results before pagination.

​
Taxonomies & Enrichment

Flat: Single-level classification. Each document maps to one or more nodes.
  • Example: Product categories (Electronics, Clothing, Home)
Hierarchical: Multi-level tree structure with inheritance.
  • Example: Animal taxonomy (Kingdom → Phylum → Class → Order)
Hierarchical taxonomies require compatible features at each level (e.g., coarse vs fine-grained embeddings).
Materialized (post-ingestion):
  • Taxonomy stable, changes infrequently
  • Enrichment cost amortized across many queries
  • Low-latency retrieval required
On-demand (query-time):
  • Taxonomy updates frequently
  • Personalized enrichment per query
  • Cost-sensitive (only pay when enrichment used)
Yes, for minor changes:
  1. Create a new taxonomy version
  2. Update collection’s taxonomy_applications to reference new version
  3. New queries use updated taxonomy; existing documents unchanged
For major changes (e.g., new hierarchy levels), reprocess affected collections.

​
Namespaces & Multi-Tenancy

Use separate namespaces for:
  • Multi-tenancy: Isolate customer data
  • Environments: dev, staging, production
  • Access control: Restrict team/service access
Single namespace is sufficient for simple use cases.
No. Collections are namespace-scoped. To share data across namespaces:
  1. Replicate objects to both namespaces
  2. Process into identical collections
  3. Query via namespace-specific retrievers
Enterprise: Contact us for cross-namespace join capabilities.
Minimal. Namespaces map 1:1 to Qdrant collections, which are optimized for isolation. Overhead is primarily storage (each namespace has its own vectors/payloads).

​
Cost & Billing

Credits are consumed by:
  • Document creation: 1 credit per document
  • Inference: 1-500 credits depending on model (embeddings, LLMs, OCR)
  • Search: 0.1-10 credits per query (vector search, web search)
  • Storage: 100 credits per GB/month
See Rate Limits & Quotas for full breakdown.
No. Cache hits are free. This includes:
  • Retriever-level caching
  • Stage-level caching
  • Document reads (after initial creation)
Optimization tip: Aggressive caching can reduce costs by 5-10x.
Operations are blocked until:
  1. Monthly quota resets (1st of month)
  2. You upgrade tier
  3. You purchase additional credits
Set up alerts at 80% usage:
POST /v1/organizations/webhooks
{
  "event_types": ["usage.threshold_exceeded"],
  "filters": { "threshold_percentage": 80 }
}
Credits are non-refundable but roll over month-to-month within the same tier. Downgrades forfeit unused credits.

​
Security & Compliance

Yes.
  • In transit: TLS 1.3 for all API calls
  • At rest: AES-256 encryption for Qdrant, MongoDB, Redis, S3
Enterprise: Bring-your-own-key (BYOK) available.
Yes (Enterprise only). Choose deployment region:
  • US East (Virginia)
  • US West (Oregon)
  • EU (Frankfurt)
  • Custom (contact us)
See Deployment for regional options.
SOC 2 Type II in progress. Expected certification: Q1 2026.Currently available: GDPR compliance, HIPAA-ready architecture (BAA upon request).
Default retention:
  • Objects: Indefinitely (or until deleted)
  • Documents: Indefinitely
  • Tasks: 24 hours in Redis, 90 days in MongoDB
  • Cache: TTL-based (default 5 minutes)
Custom retention: Configure per-bucket or per-collection.

​
Advanced Use Cases

Yes (Enterprise). We support:
  • Fine-tuning embedding models (text, image)
  • Custom classification heads
  • Domain-specific NER models
Requires minimum 10K labeled examples. Contact us for pricing.
Yes (Enterprise). We provide:
  • Docker Compose deployment
  • Kubernetes Helm charts
  • Full source code access (with license)
See Deployment for self-hosted options.
No, only REST API. GraphQL support planned for 2026.Workaround: Build a GraphQL wrapper around REST endpoints.
Yes. Use the Documents API to export vectors and metadata:
POST /v1/collections/{collection_id}/documents/list
{ "return_vectors": true, "limit": 10000 }
Paginate through full collection and load into your database.

​
Support & Community

  1. Documentation: You’re here! Start with Quickstart
  2. Talk to Engineers: Use CTA in top bar for 1:1 support
  3. GitHub Issues: For bug reports and feature requests
  4. Discord: Community support and discussions (link in footer)
TierResponse Time
FreeBest effort, 48-72 hours
Pro<24 hours business days
Enterprise<4 hours, 24/7
Critical issues (P0): Escalated immediately for Enterprise.
Yes.
  • Pro: 2-hour onboarding call included
  • Enterprise: Dedicated solutions architect, quarterly reviews
Paid consulting available for:
  • Custom integration development
  • Performance optimization audits
  • Training for internal teams

​
Still have questions?

Can’t find what you’re looking for? Reach out:

Talk to Engineers

Get 1:1 support from our team

Join Discord

Community support and discussions