Find any scene in your video library.
Mixpeek understands your video, image, audio, and documents, and returns the most accurate, timestamped results from the storage you already have. Every search teaches it: relevance keeps improving the more you and your agents use it.
Explore the live demos above with zero setup: then start here when you want Mixpeek running on your own data.
Bring your own vectors
You already have embeddingsMixpeek Vector Store (MVS): an agent-native vector store that runs on your object storage. Dense, sparse, and BM25 search. From $25/mo.
Connect your files
You have raw filesManaged indexing: point us at video, audio, or documents and we extract scenes, faces, OCR, transcripts, and embeddings. No pipeline to build. From $25/mo.
What the agent sees.
Every object you index becomes structured, searchable features: faces and objects in a frame, layout regions in a document, speakers in audio. Those are the same features an agent queries, and joins across modalities.
Hover or tap a card to preview the search it powers.
Person · 0.97Face · 0.95Handbag · 0.92Video · 00:04:12“A woman carrying a tan tote bag walks past a red storefront.”
transcript · “…meet me at the corner in five.”
HeaderChartBodySignaturePDF · resume.pdfHeader, body, charts, and signature detected as typed regions.
OCR · 1 header · 3 sections · 1 signature
Who spoke when, aligned to the transcript and the timeline.
audio · 2 speakers · matched at 00:01:30
One query across every modality.
Real questions rarely fit one feature. “Find the moment our CEO said guidance while the slide read Q4 outlook” needs a face, a spoken phrase, and on-screen text to line up at the same instant.
Mixpeek ties those features to the same object and timestamp, so an agent gets back the exact clip instead of three unrelated matches.
The CEO says “guidance” as the slide behind her reads “Q4 outlook.”
"What concepts exist in my data that nobody has labeled yet?"
No keyword, example, or prompt can answer that: they all assume you already know what you're looking for. Mixpeek clusters what belongs together on its own, so a natural hierarchy surfaces instead of a flat pile of tags: a taxonomy built from your data, organized around what your business actually cares about.
That taxonomy is your ground truth, and it feeds back. Every search, every correction sharpens the features, the clusters, and the relationships. Your competitors' metadata decays. Yours compounds.
- Creative moments644
- └Unboxing214
- └Hands-on close-up121
- └Reveal + reaction93
- └Night driving88
- └Product on white342
every search + correction feeds back → sharper features, tighter clusters, truer taxonomy
In production right now.
Search video by what happens on screen
Type what you want to see, not what someone said. Every query fans out across three vector spaces (one for the picture, one for what a vision model says is happening, one for the dialogue) and reciprocal-rank fusion merges them server-side. Results are scenes with timecodes, not whole files.
Try video search →Visual search across 45k artworks
Upload any image and find visually similar paintings across 45,000+ artworks, or just describe what you're looking for. Hybrid image and text retrieval, ranked with RRF.
Try gallery search →Face search across video
Drop in a headshot and find every clip a person appears in across 63 video ads and 2,600+ faces. Full trace for takedown evidence.
Try face search →One install. Two paths.
Most retrieval stacks mean gluing together a vector DB, a file pipeline, and an agent layer. Mixpeek is one install with two ways in.
Bring embeddings
Plugs into your existing stack.
Connect your storage, point Mixpeek at it, and every file becomes searchable by what's inside it. No migration, no code changes.
Mux
Every Mux upload becomes searchable by face, scene, transcript, and on-screen text, with no manual tagging.
View integration →
Backblaze B2
S3-compatible extraction at 1/5th the cost. Store on B2, extract with Mixpeek, zero egress fees.
View integration →
Iconik
Every asset in your DAM becomes findable by what's inside it: scenes, faces, spoken words, on-screen text.
View integration →Pick your file types. Choose what to search by.
Video, image, audio, documents, text, or web: connect a bucket, pick the features you want to search by, and these pipelines run as they are. Every one is documented and open source in the extractor cookbook.
Video · Image · Audio · Text
Search by: Scenes, speech & visual similarity
Unified embeddings for video, audio, image, and text. Scene and silence chunking, Whisper transcription, thumbnails.
Multimodal (Video/Audio/Image)
Any file
Search by: Everything in one pass, any file
One extractor for image, video, audio, and documents. Auto-detects modality and applies the right pipeline.
Universal All-in-One
Image · PDF
Search by: Visual similarity, described in words
Dense 768-D image embeddings with Google SigLIP for text-to-image search in one contrastive space.
Image Embeddings (SigLIP)
Text
Search by: Meaning, not keywords, in any language
Multilingual dense text embeddings with E5-Large for semantic search and RAG out of the box.
Text Embeddings (E5-Large)
Any file
Search by: Whole objects: all their files as one
Embed ALL files of an object (images, PDFs, video, audio, text) into one 3072-D Gemini vector.
Multi-File Object Embeddings (Gemini)
Image · Video · PDF
Search by: The same face, across your whole library
Production face recognition that detects, aligns, and embeds faces to 512-D ArcFace vectors.
Face Identity (SCRFD + ArcFace)
What we shipped lately
- Aug 20APIInteraction reads and learned-fusion weights now answer from your own signals collectionListing interactions and reading Thompson Sampling arm stats both prefer the documents in your namespace's signals collection, with the analytics store answering whenever the collection cannot. The cutover is decided by a coverage rule rather than a guess: the collection is provably complete when it was created within a day of the namespace itself, or when it is older than the retention window, so a namespace with older history keeps being served from the store that has all of it. New namespaces read from their own collection immediately. Weight responses now carry a source field naming which store answered, so support and verification can tell without inferring. Filter translation is bounded on purpose, covering equality, membership and ranges, and falling back rather than approximating anything it cannot express exactly.
- Aug 20APIYour annotations and resource events land in the same collection as your search signalsHuman relevance judgments made through the annotations API now also become documents in your namespace's signals collection, with the label, confidence and reasoning at the payload root and the structured payload nested under its own filterable paths. Lifecycle follows: the mirrored document is keyed by the annotation id, so editing rewrites it in place and deleting removes it, and bulk operations inherit all three behaviours. Separately, events on your own resources such as bucket uploads, syncs, extractions, cluster and taxonomy runs land there too, filterable by resource id, status and event type. The result is one queryable timeline covering what was asked, what was retrieved, what a human judged, and what happened to the underlying resources.
- Aug 20PlatformSignal documents expire on a retention window instead of growing foreverA daily sweep deletes signal documents older than the retention window, matching the expiry the analytics store already applied and keeping the mirrored collection from growing without bound. The window defaults to 90 days and can be overridden per collection. Signal documents now carry an epoch timestamp beside the readable one, because the filter grammar ranges over numbers rather than dates. The delete is scoped to the signals collection and bounded by the time range, both clauses load-bearing since your own documents share the namespace, and a failure on one namespace logs and continues rather than stopping the sweep for everyone else.
- Aug 20APIAsking for vectors back gives you one naming system and tells you which keys are vectorsRequesting embedding values when listing documents returned them keyed by internal storage names, while the vectors manifest and the single-document read used the output names you configured, so one response contradicted itself and there was no way to line the names up with the values. The list path now maps to output names exactly as the single-document path already did. Both read endpoints also return a manifest naming the vector keys actually present when values were requested, and an empty list when hydration returned none, because values arrive as root fields named after each vector and nothing previously told you which of those fields were vectors.
From $25/mo. Usage-based everything.
Two products, one model: a monthly minimum that acts as a floor, with usage above it billed at the same transparent rates. MVS is priced by the vector, Managed by the object.
Bring your own embeddings and pay by the vector. Dense, sparse, and BM25 search on your own object storage. Build starts at $25/mo with up to 1M vectors; Scale ($250/mo) covers 25M.
Start with MVSBring raw objects and pay by the object: credits at $0.001 cover extraction, embedding, indexing, enrichment, and retrieval. Build covers 100K objects/mo; Scale ($250/mo) covers 1M.
Start with ManagedDedicated infrastructure, self-hosted options, SSO, SLA, security reviews, and hands-on architecture support.
Talk to usCommon questions.
Do I have to move my data?
No. Mixpeek reads from your existing S3, GCS, R2, Azure, or S3-compatible bucket. Your storage stays the system of record, and nothing leaves your cloud.
How fast is retrieval?
Hybrid queries (dense, sparse, and BM25) return in well under 100ms p95, even with vectors persisted on object storage rather than held in RAM.
Do I need embeddings to start?
No. Bring your own vectors with MVS, or point Managed at raw files and it generates embeddings and features for you.
What can Managed extract?
Faces, scenes, transcripts, OCR, labels, and embeddings from video, images, audio, PDFs, and documents, all indexed at the object level.
Can I self-host?
Yes. Deploy in your own cloud (BYO-Cloud) with SOC 2-ready and HIPAA-ready controls, SSO, audit trails, and namespaces.
How does pricing work?
Both MVS and Managed start at $25/mo minimum. Usage counts toward the minimum: pay the greater of metered usage or the floor. MVS bills storage + queries; Managed bills in credits covering extraction, embedding, indexing, and retriever execution.