Domain embeddings
Meaning that knows the law.
The same base model, adapted to one domain. This retrieves the true continuation of a court passage from a pool of held-out Supreme Court judgments — running Quanfire.ai’s legal adapter next to the raw base it was trained over, so you can see what domain adaptation buys, not take our word for it.
Find the passage that comes next
The task: given a passage from a Supreme Court judgment, retrieve its true continuation from a pool of 220 held-out passages. Pick a passage below — both models rank the same pool, and we show where each one placed the real next passage. The adapter was trained only on other judgments; nothing here was seen in training.
Embedded live, in real time · ⌘/Ctrl + Enter to rank · free text has no known answer, so no continuation rank is shown
How this stays honest
The base column is the exact intfloat/multilingual-e5-small checkpoint served through the identical stack as the adapter, with an untrained (zero-effect) adapter attached — so the only thing that changes between the columns is the trained legal weights. Every judgment in the retrieval pool was held out of training. The menu shows curated wins for a clean demo, but the aggregate under it reports all held-out queries, losses included — nothing dropped from the measurement, only from the showcase. Every rank you see is computed the moment you press the button.