The Event
// atomic unit of meaning
Your chunk boundary is an event, not a sentence or a token. The boundary lives in the world, so it survives scale.
Open the object →
Domain-constrained RAG infrastructure
Stop trying to solve chunking.
Constrain it instead.
Chunking is not a retrieval problem to be solved elegantly. It is a product decision you lock on day one. Pick one domain, define its atomic unit of meaning, and make that the chunk. Everything downstream gets fast, deterministic, and cheap.
1
atomic unit: the event
4+1
core types + catch-all
50
benchmark queries
0
mixed-version indexes
The structure
Four objects, wired into one structure.
Most RAG stacks are a pile of parts. SibylData is four objects that interlink in a loop: the Event instantiates a Schema, the Schema maps to the Context Slots, the Slots query the Versioned Index, and the Index writes the next Event. No object is complete without the other three. Each one prevents a single, costly failure that teams pay for late.
- Event → Schema
- Something happened, to someone, at a time, with an outcome. Each event instantiates a typed schema.
- Schema → Slots
- Typed fields map to typed positions in the brief the model reads.
- Slots → Index
- Each slot queries a version-tagged index, never a mixed one.
- Index → Event
- New events are written at the current embedding version. The loop closes.
The four objects
Each one is a decision most teams pay for late.
The Schema
// typed contract
Schema-first on what you know, a structured catch-all for what you do not. 80% of the benefit on day one, no blind spots.
Open the object →The Context Slots
// ranked positional brief
Lost-in-the-middle is a ranking problem you own. Assemble typed slots, not a flat pile. You control position.
Open the object →The Versioned Index
// temporal coherence
Tag every chunk with its model version, never query a mixed index. Three config fields now, or a rebuild under live load later.
Open the object →Chunking is not a retrieval problem. It is a symptom of never deciding what your data model is.
You tune token windows, sentence 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. The moment you define your domain's atomic unit, chunking resolves itself. This is the unglamorous decision that the teams who sell data-infrastructure companies make on day one, and the rest pay for in iteration seven, under live load.
How you would prove it yourself, without taking our word for it →
A question before you book
This call is not for everyone.
SibylData is for teams building retrieval over a real domain, who already feel the cost of leaving the unit undefined. One question decides whether the call is worth your time.
The reading
Bring one domain. Leave with a data model.
On a 20-minute technical call we define your atomic unit, sketch the core event schema, and name the one benchmark that will prove your retrieval works.
Book my technical call →