Dataset Scorer
For the anchor: like-for-like relevance you can benchmark and refine. The same rigorous scorer class — nDCG, Recall, MRR — behind our existing evals.
Coming to Pro
AI creativity, grounded in your knowledge
Vectors that don't just find — they inspire.
Most vector search returns the closest match. Enscrive returns the closest match and the unexpected, grounded connections around it — so your AI creates from your knowledge, not from thin air.
The Idea Underneath
One anchor. Grounded looseness.
A tight, deterministic anchor for precision, plus grounded looseness for creative divergence — tuned on two knobs: representation (how many layers you store) and precision (how much of the vector you search). Retrieval supplies grounded, diverse material; your agent turns it into creativity anchored in what you actually know.
Two distinct offerings, one idea — below.
Offering One · Representation
The anchor, plus your layers.
Store content in layers: a deterministic anchor — your text, exactly — plus interpretive layers you choose. Story beats. Themes. A reference corpus. Whatever angle matters to you.
One search returns the anchor — the like-for-like closest meaning — and the intersections: loosely-associated passages from your other layers that hint at the query from a different angle.
A reasoning agent uses the anchor for precision and the intersections for creative divergence — leaps that stay grounded in your own corpus.
1
Anchor layer
Deterministic. Your text, embedded exactly as written. Returns the closest, like-for-like meaning — the result you'd expect from ordinary vector search.
2
Your layers
Interpretive. You define the angle — story beats, themes, a reference corpus. Returns intersections: passages that hint at the query from elsewhere in what you know.
One search, two kinds of results
The anchor gives your agent precision. The intersections give it creative divergence — grounded, because every intersection still traces back to a layer you defined over your own corpus, never to thin air.
Offering Two · Precision
Coarse to fine, in one vector.
One vector, searchable coarse-to-fine. A short prefix is a fast, broad sweep; the full vector refines to precision. Fast when you want speed, deep when you want nuance.
The space between coarse and fine is grounded serendipity — near-neighbors the fine ranking would sharpen away, surfaced instead as creative fuel.
1
Coarse prefix
A short slice of the vector. A fast, broad sweep across the whole corpus.
2
The middle — grounded serendipity
Near-neighbors a fully-refined search would sharpen away. Surfaced as creative fuel instead of discarded.
3
Full vector
The complete resolution. Precise, fine-grained ranking when nuance matters most.
Evaluation, Two Ways
Prove the anchor. Judge the spark.
The anchor and the loose layers answer different questions, so they're measured differently. Precision is a benchmark. Creative lift is a judgment call — and we plan to give both a real, first-class eval path.
LLM / Human Judge
For the creative layers: does the surfaced material actually lift the output? A different question, answered a different way — a new eval class we're building purpose-built for the loose layers.
We don't sell these as benchmarks — we sell AI creativity, well grounded in your knowledge.
Professional
Coming to Professional.
Multi-Layer Vectors and Matryoshka Vectors are headed to the Professional tier — alongside eval-gated promotion and the rest of the first-class evaluation surface.