The Short Answer
There is a class of queries that no conventional search system can answer: the ones where you do not know what you are looking for yet. Keyword search needs your term, vector search needs your example, and RAG needs your question: all three assume the ontology exists before the query does. Discovered taxonomies invert this: the system clusters the latent structure of your corpus first, labels what it finds, and hands you a navigable hierarchy, so the answer to "what concepts exist in my data?" is computed, not retrieved.
Why do these queries break normal search?
A concept like "trust" in advertising is not a keyword and not a single embedding neighborhood. It is expressed through hundreds of recurring motifs (handshakes, doctors, families, a certain palette, a certain pacing) that no one enumerated in advance. Any system that requires the query to name the target fails in one of two ways: it returns the literal matches (videos containing the word "trust") or it returns one neighborhood of one motif and misses the rest. The fix is structural: cluster the corpus across every modality (the mechanics of clustering across modalities, and clustering the clusterings, are in composite clustering), let the recurring motifs surface as groups, label the groups, and organize them into a taxonomy that emerged from the data (the inverse job, assigning content to a taxonomy you already have, is taxonomy classification) rather than a whiteboard.
Six queries that require a discovered ontology
1. "Show me how this company visually represents trust." A corpus of 500,000 marketing videos expresses "trust" through motifs nobody listed: handshakes, smiling clinicians, families at tables, steady wide shots. Clustering finds these as recurring visual groups; the taxonomy names them; the query becomes navigation.
2. "Find every way our brand has portrayed fathers." Not the keyword "father": the dad fixing a bike, the barbecue, the cheering at soccer, the hug at the door, the voiceover that says "Dad." No human enumerates that list beforehand; the corpus already contains it as structure.
3. "What stereotypes keep appearing across our advertising?" Nobody knows the answer in advance: that is the point. The hierarchy surfaces clusters that were never labels: women cooking, elderly couples, military families. Emergent structure is the only honest answer to a question about patterns you have not noticed.
4. "Show me the evolution of 'luxury' over the last 20 years." Luxury drifts: marble kitchens, then Teslas, then quiet-luxury minimalism. A concept with a moving definition cannot be a stored label: but time-sliced clusters over the same corpus reveal the drift directly.
5. "Find content spiritually similar to this commercial." Not visually similar, not semantically similar, evoking the same nostalgia or anticipation without sharing objects, people, or words. Cross-modal cluster membership captures affinity that no single embedding distance expresses.
6. "What themes exist in my footage that I don't know exist?" The purest form: not a search at all, but "teach me the ontology of my data." A query-first system cannot start; a structure-first system has already finished.
How does the machinery actually work?
The pipeline is embed, cluster, label, promote: every asset is decomposed into per-segment features (visual, transcript, faces, on-screen text), clustering groups the segments across feature spaces, an LLM labels each cluster from its actual contents, and stable clusters are promoted into a hierarchical taxonomy that then classifies new content at ingest, powers filters at query time, and reclassifies retroactively when the hierarchy changes. Every search and correction afterward sharpens the structure: the ontology compounds instead of decaying.
Doing this on Mixpeek
This loop is Mixpeek's core mechanism, not a feature bolted onto search: server-side clustering runs at collection scale over content in your own object storage, the Studio cluster explorer makes the groups browsable with real sample items, one step promotes a cluster to a taxonomy, and retrievers use the taxonomy as filters alongside every other signal. The brand-intelligence version of these six queries (trust motifs, brand portrayals, creative drift) is the Creative DNA solution. Start with how the compounding loop works or explore clustering.