Recipes

Turn raw media into structured intelligence

Multi-tier feature extraction that decomposes content into searchable components: embeddings, transcripts, detected objects, OCR text, scene boundaries, and more. The foundation for all downstream retrieval and analysis.

Find anything across video, image, audio, and documents

Unified semantic search across all content types. Query by natural language and retrieve relevant video clips, images, audio segments, and documents based on meaning-not keywords or manual tags.

Auto-label content into structured taxonomies

Assign content to multi-level category hierarchies using embedding-based classification. Define your taxonomy once, then classify new content automatically with confidence scores.

LLMs that cite real clips, frames, and documents

Retrieval-augmented generation across video, images, and text. Retrieve relevant multimodal context, then pass to your LLM with citations back to source timestamps and frames.

Reveal structure you didn't know existed

Unsupervised clustering that groups content into semantic themes using HDBSCAN. Surfaces hidden patterns, content variants, and outliers without requiring predefined labels.

Link extracted content to business context

Bridge extracted content features with business reference data. Join video clips to product catalogs, detected faces to employee directories, or documents to compliance frameworks-all via embedding similarity.

Spot unusual content automatically

Identify outliers and anomalous content using embedding distance from cluster centroids. Flag quality issues, novel content, or items that don't match expected patterns.

Rebuild any dataset state deterministically

Treat versioned object storage as your dataset's source of truth. Capture complete snapshots-raw assets, embeddings, and cluster assignments-for deterministic reconstruction at any point in time.

Track how your data changes over time

Monitor distribution shifts between baseline and current data using cluster comparison. Detect when new content diverges from training data or when content mix changes unexpectedly.

Auto-detect and assemble highlight moments from sports footage

Automatically identify highlight-worthy moments in sports broadcasts using multimodal analysis, visual action detection, audio spike recognition (crowd noise, commentator excitement), and on-screen graphic parsing. Returns timestamped event manifests ready for clip assembly.

Bring your own embeddings and search in 60 seconds

Bring pre-computed embeddings from any provider (OpenAI, Cohere, Together, etc.) and upsert them directly into MVS for instant vector search. No feature extractors, no pipelines -- just embeddings in, results out.

Bring your own vectors to MVS, search them by vector, and know the one setup step BM25 needs

Upsert documents you embedded elsewhere into an MVS namespace and search them by raw vector through the features search endpoint. Hybrid BM25 plus dense is not part of a plain BYO upsert: the documents carry dense vectors only and no text index is created. If you want a lexical leg later, declare a TEXT payload index on the field when you create the namespace; this recipe shows that declaration and the dense search that works today.