Stop trying to solve chunking. Constrain it instead.

Chunking is not a retrieval problem. It is a product decision.

The chunking debate is a symptom of never deciding what your data model is. It looks like a retrieval engineering problem. It is actually a product definition problem. The moment you define your domain's atomic unit, chunking resolves itself.

The symptom you keep optimising

You tune token windows, paragraph splits, and overlap for months. Retrieval quality still drifts as the corpus grows. The fix was never a smarter splitter, and waiting for a smarter model will not save you either.

Lock the atomic unit on day one: the event.

For your domain the atomic unit is almost certainly an event. Something happened, to someone, at a time, with an outcome. A financial transaction, a life decision, a health event, a career move. Every event carries a schema, a timestamp, and a set of typed fields.

Then chunking is solved, permanently

You define that schema in week one and never change it without a formal migration. Not sentences, not tokens, not paragraphs. Events. That single decision makes retrieval fast, deterministic, and cheap.

The whole system in one sentence

Schema-first on the events you know, a structured catch-all for the rest, typed slots into the model, and version-tagged indexes, all proven by a 50-query benchmark. Five decisions, locked early, that most teams pay for late and at ten times the cost. We help you make them on day one, before there is a pipeline to unwind. That is the entire product, and it is deliberately unglamorous.