← Foundations
Codename · Scout
Retrieval · Foundation

Entity search that understands
what you're looking for.

Scout is a semantic search engine for profile-based products. It understands what a user means in plain language — across talent, creators, properties, listings — and returns who, or what, actually fits.

See how it works Not keyword search with a thesaurus bolted on. A three-layer retrieval engine, already running in production.
Indexing
Facet + CLIP
Understanding
LLM planning
Retrieval
kNN + rerank
Gating
LLM verdict
Modality
Text + visual
Domains
3 verticals
01 · The problem

Keyword search was built for documents. Your profiles are people.

Users rarely know the keywords. They only know what they mean.

Keyword search works when users know the exact terms to type. In profile-based products, they rarely do. A recruiter looking for “someone who has scaled a data team through hypergrowth” does not know which keywords to use. A brand looking for “a creator with a warm, editorial aesthetic and a quietly loyal audience” cannot express that in a filter panel. A tenant looking for “something that feels like a loft but in a quieter neighborhood” is not going to find it by bedroom count.

The problem is not that users are imprecise. The architecture was designed for a different kind of query. Keyword and filter search treats profiles as flat documents — but real profiles are multidimensional. The person, the creator, the property each carries a topic dimension, a visual dimension, an audience dimension, a behavioral dimension, a style dimension. Keyword search collapses all of it into one text field and hopes.

Scout does not. It indexes every dimension separately, understands what the query is asking across each of them, and retrieves against all of them at once.

02 · Architecture

Three layers. Each doing something the others can't.

Most semantic search is one layer — embed the query, embed the documents, take the nearest neighbors. Scout is three: it understands the query before it retrieves, retrieves across every dimension at once, and gates the results before you ever see them.

01
Query understanding
reads intent before retrieving
“warm editorial creator…”
TopicsFacet keywordsFiltersIntents
LLM plan
N sections
02
Retrieval & ranking
per section, ≤100 candidates
Retrieve · runs in parallel
Structured filter
hard constraints
Facet keyword
boosted match
kNN vector
facet + CLIP
Then
Neural rerank
semantic re-order
Then
Relevance gate
LLM verdict · binary
output
Streamed sections
03
Index store
what retrieval reads from
Facet embeddings
N dimensions, indexed apart
CLIP visual
images + text, one space
Structured fields
filters, wired day one
Fig. 1 — One query, three layers. Understanding parses the query and plans the page into thematic sections; each section retrieves from the index across three methods at once, reranks, then gates on meaning. Results stream as each section resolves.
03 · The three layers

Each one solving a different part of the problem.

Most semantic search implementations are a single layer. Scout is three — indexing, query understanding, and retrieval — and each is doing work the others cannot.

Profile
bio · posts · images · audience · history
Topic focus
weighted ↑
Visual aesthetic
text + CLIP
Audience
embed
Behavioral
embed
Content style
embed
Identity
embed
Affinities
weighted ↓
Fig. 2 — One profile, decomposed into multiple facets and embedded independently. The visual facet adds a CLIP embedding so text and images share one space. A query that is mostly about aesthetic searches the visual facet without being diluted by topic noise.
01

Indexing

Every profile, indexed across every dimension.

Profiles are not stored as a single blob. They are decomposed into discrete facets — each capturing a different dimension of what a profile is — and each facet is indexed independently. A query about experience is not diluted by location noise. A query about style is not confused with category. On top of text, CLIP visual embeddings place images and language in the same vector space, so a description of what something should look like lands near profiles whose imagery matches — with no labels on the images at all.

02

Query understanding

The query arrives in plain language. Scout reads it like an expert would.

Before any retrieval, Scout parses the query across every dimension at once — extracting topics, generating facet-specific keyword and phrase lists, resolving hard filters expressed in natural language, and detecting special intents that shape how retrieval should run. An LLM planner then decides how to structure the page: not a flat ranked list, but a set of distinct thematic sections, each targeting a different angle and running its own retrieval.

03

Retrieval & ranking

Only the right candidates make it through.

Each section retrieves independently — structured filtering, facet-boosted keyword matching, and kNN vector search running simultaneously. A neural reranker re-orders the raw candidates by true semantic relevance, then an LLM relevance gate makes a final pass, giving each candidate a binary verdict and removing anything that technically matched but clearly does not fit. What remains streams to the interface.

04 · Differentiation

What vector search alone won't give you.

Embedding the query and finding nearest neighbors is table stakes. These four are what separate a retrieval engine from a search box.

01

Facet-level indexing

Standard semantic search embeds the whole profile. Scout embeds each dimension on its own. The payoff is searchability without noise — visual queries are not diluted by topic signals, audience queries are not confused with content format.

02

Dual modality

Text facet embeddings and CLIP visual embeddings run in parallel. Scout retrieves against what a profile says and what it looks like, simultaneously and independently. Most approaches give you one or the other without custom engineering.

03

LLM search planning

Most search returns a ranked list. Scout returns a structured page — multiple thematic sections, each independently targeted. Users see the best results organized by angle. That is the difference between a search engine and a discovery product.

04

Relevance gating

Retrieval and reranking get you close. The LLM relevance gate gets you precise — profiles that matched on keywords but missed on meaning are removed before the user ever sees them. What is left is a set a human expert would be proud to have assembled.

05 · What comes next

Every new modality adds a dimension that was previously invisible.

Scout's architecture is built to grow. The dual embedding strategy — text and visual — is one instance of a broader principle: any modality with a contrastive, language-trained model becomes a searchable dimension.

CLAP does for audio what CLIP does for images — embedding actual sound so a query like "warm, intimate acoustic with a melancholic mood" retrieves against it, with no text labels. Video-language models extend the logic to motion and pacing; document embeddings to portfolios, CVs, and case studies; behavioral and temporal embeddings to trajectory — not just who a profile is, but how they have moved over time.

Every domain Scout has shipped in — talent, creators, properties — added new modalities and new facets, and every deployment made the architecture more capable. A new engagement inherits not just the engine, but everything learned from every domain that came before it.

Keyword search was never going to get you there. The gap between what your users mean and what they type is an architecture problem — not a UX one.

Every engagement that draws on Scout starts with the three-layer architecture already in place. The facet model is designed. The dual embedding strategy is implemented. The query understanding pipeline, the search planner, the reranking and gating logic — all of it has already run in production. What an engagement adds is your domain, your profile structure, and your users' query patterns. That is the part worth spending time on.

Start somewhere

Building a discovery product where users should be able to describe what they want?

Tell us about your domain, your profile structure, and where your current search breaks down. We'll tell you honestly what Scout can do for it — and what it would take.