What we use representations for
A user representation is the answer to “what is this user like?” — and any product surface that has to make that decision can pull from it.
LCM representations are consumable by LLMs, so you can use them in any LLM application.
Below are three of the most common uses, each with the actual input the representation contributes and the actual output it helps produce.
Re-ranking — same retrieval, smarter order
Existing search returns a candidate list ordered by popularity. The representation re-orders it for this specific user, using their food preferences and temporal patterns. Cheapest first bet because it sits on top of what you already have.
Push notifications — what to say, when, to whom
A CRM agent decides which item to surface, what tone to write in, and when to send. Persona, temporal patterns, and food preferences from the representation drive the targeting; the LLM generates the copy.
Query expansion — turn a two-word search into the right result
A short query like “pizza” tells you almost nothing about what the user actually wants. The representation fills the gap: the LLM expands the query into an “ideal match” description for the specific user, and that expansion is what hits semantic search.