1.  [](/)
2.  [Diagrams](/diagrams)
3.  The 3072-Dimension Problem: Reduce, Then Place

[Back to Diagrams](/diagrams)IR Foundations

# The 3072-Dimension Problem: Reduce, Then Place

Representation learning as infrastructure: reduce to named features, then make the reduction a permanent, queryable hierarchy.

![A 3072-dimensional embedding encodes everything about a video and distinguishes nothing: lighting, camera angle, dominant color, who is on screen, the room and the vibe of the room are all smeared across the same floats at equal weight, with no name you can filter on. That is why a search for a person holding a coffee cup returns videos with people, videos with cups, and a surprising number with neither, and why reranking and hybrid BM25 each help only a little. Decomposition is where the dimension reduction actually lives: a shot detector turns a thirty-second video into eight scenes with timestamps, a face identity model turns a frame into bounding boxes and 512-dimension face embeddings, a multimodal extractor turns a scene into a Gemini embedding plus OCR output plus dominant colors, and a custom plugin does the same for a proprietary file type. Each extractor's output becomes its own collection with its own schema and its own embedding space, so the features are named and filterable where a smear is not. Reduction alone leaves a smarter vector database; the leverage is placement. A new ad is placed against a brand taxonomy at ingest, and because the taxonomy that locates it is the same retriever a user invokes to ask for ads in that brand, enrichment and search become one traversal in two directions. Six primitives compose in one direction: bucket, extractor, collection, retriever, taxonomy, cluster.](/diagrams/the-3072-dimension-problem.webp)Full screen

Representation learning as infrastructure: reduce to named features, then make the reduction a permanent, queryable hierarchy.

A 3072-dimensional embedding encodes everything about a video and distinguishes nothing. The lighting, the camera angle, the dominant color, who is on screen, the room they are in and the vibe of that room, all of it smeared across three thousand floats. Cosine similarity treats every dimension equally. Your application does not.

That is why a search for a person holding a coffee cup comes back with videos of people, videos of cups, and a surprising number of videos with neither. Reranking helps a little and hybrid BM25 helps a little more, and somewhere around the fourth round of tuning the suspicion sets in that the retriever was never the problem. The unit of work is wrong.

The reframe is old, and representation learning has always done it: reduce complex things to small interpretable representations. What is rare is running it as managed infrastructure, continuously, across millions of objects, with the reduced features exposed as first-class queryable surfaces.

Decomposition is where the dimension reduction actually lives. A shot detector turns a thirty-second video into eight scenes with timestamps. A face identity model turns a frame into bounding boxes and 512-dimension face embeddings. A multimodal extractor turns a scene into an embedding plus OCR output plus dominant colors. Each output becomes its own collection with its own schema and embedding space, which is what makes a feature filterable.

Reduction alone is not enough. Stop after the extractor stage and you have a smarter vector database, still a system where every query starts from scratch. The leverage is the hierarchy: once a taxonomy exists, every new object is placed against it at ingest, the fingerprint is computed once, its location is computed once, and retrieval afterwards is mostly traversal. That collapses enrichment and search into one operation run in two directions, because the taxonomy that locates a new ad is the same retriever a user invokes to ask for ads in that brand.

## Where this diagram appears

*   [Blog: The 3072-Dimension Problem →](/blog/the-3072-dimension-problem)
*   [Guide: Matryoshka Representation Learning →](/guides/matryoshka-nested-embeddings-adaptive-retrieval)
*   [Glossary: Representation Learning →](/glossary/representation-learning)

Run this on your own data

Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API.

[Search your own data](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=diagram_detail&utm_campaign=mvs_onramp)

## Related diagrams

[

![Diagram placing BM25, SPLADE and dense embeddings on a single axis measuring how much vocabulary survives compression, with what each one stores and why production systems run more than one.](/diagrams/sparse-vs-dense.webp)

Sparse and Dense Are One Spectrum, Not Two Camps



](/diagrams/sparse-vs-dense)[

![Diagram comparing memory-resident vector search against storage-first vector search on object storage, showing the cost structure of a billion 768-dimensional vectors under each model.](/diagrams/s3-vector-warehouse.webp)

The S3 Vector Warehouse: Storage-First Vector Search



](/diagrams/s3-vector-warehouse)[

![An ordinary embedding model spreads information across all dimensions with no ordering guarantee, so dimension 900 may carry as much signal as dimension 5 and truncating the vector discards random pieces of meaning. Matryoshka Representation Learning changes the objective so the model packs the most general information into the front of the vector and progressively finer detail toward the back, which makes every prefix a complete usable coordinate. The nesting is drawn as five bars that all start at dimension 0 and grow: the first 64 dimensions carry coarse semantics, enough to tell a cat photo from a car photo; 128 is the prefix the retrieval cascade indexes the whole corpus at; 256 gives mid resolution, enough to separate breeds of cat or a refund request from a refund confirmation; and the full 1024 gives fine resolution for ranking near-duplicates. The mechanism is a modified loss rather than a special architecture: one forward pass produces one d-dimensional output, the training loss is computed independently at each prefix size in the set 64, 128, 256, 512, 1024, and those losses are summed, so the front of the vector has to work on its own. What it buys is a retrieve-coarse, rerank-fine cascade: index 10,000,000 items at a 128-dimension truncation costing 5.1 GB instead of the full 41 GB, retrieve a top-200 shortlist with the truncated query, fetch the full 1024-dimension vectors for those 200 candidates only, and return a final top 10.](/diagrams/matryoshka-representation-learning.webp)

Matryoshka Representation Learning: One Vector, Five Resolutions



](/diagrams/matryoshka-representation-learning)