Concept
Built on a different principle.
Knowledge cannot be fixed to a single correct meaning. The same words land differently depending on who reads them — and the distance is wider still between people and AI.
VATES treats knowledge not as fixed meaning, but as structure that each reader brings to life in the moment. Where conventional methods fix meaning in advance, VATES leaves interpretation to the reader. That principle is the foundation it is built on.
Why it is built this way
VATES versus the common method for feeding knowledge to a language model (RAG).
| Aspect | RAG | VATES |
|---|---|---|
| Intermediate processing | Embedding, vector search, re-ranking | None of them — BM25 over plain text, top-k into the model |
| Updates | Re-embed and re-index required | Takes effect the moment an ES is written |
| Meaning | Compressed into coordinates at embedding time, and drifts | Kept as structure; interpreted at observation time |
| Dependencies | Vector-database infrastructure | Standard database and files only |
| Visibility | Numeric vectors; why a fragment was retrieved is opaque (black box) | Structured plain text; the contents are visible (white box) |
| Operation | Re-embed, re-index; deletions may persist in the vectors | Edit or delete the line directly; verifiable by eye |
No intermediate processing
A typical RAG system runs a pipeline — embedding, vector search, re-ranking. VATES has none of them.
A question comes in. Every ES is scored against it with BM25. The top-k are handed to the model. The model answers.
That is the whole path. The only thing prepared in advance is the ES itself; nothing is prepared for retrieval. The scoring is rebuilt for each question, so there is no index to maintain, nothing to re-embed, and no fixed meaning stored anywhere. Retrieval is lexical, over plain text you can read. The model observes the structure directly — meaning is not resolved before it arrives, but at the moment it is read.
Instant updates
RAG must re-embed and rebuild its index whenever knowledge changes. With VATES, once an ES is written it is reflected from the very next query. There is no index to rebuild.
Meaning interpreted at observation time
Embedding flattens meaning into fixed coordinates the moment it is stored. VATES does not fix meaning in advance — retrieval is reassembled per query, and interpretation is left to the reader. From one knowledge base, responses adapt to the question and to the reader's language. This is why knowledge stored in Japanese can answer in French, with no translation data held at all.
Standard database and files only
No dedicated vector database — no Pinecone, Weaviate, or pgvector. Knowledge is plain structured text, running on an ordinary database and files. No GPU, no per-use embedding API. Fewer moving parts means lower cost and fewer points of failure.
White box
A RAG index is a mass of numeric vectors no human can read; you cannot tell why a fragment was retrieved. VATES keeps knowledge as plain text along basic axes of observation — substance, quantity, quality, relation, and so on — and the axes extend freely when a subject does not fit. It is not a fixed schema. You can read it by eye and search it. The model sees syntax; a person sees a table.
Direct edit and delete
A deletion in RAG can leave traces in the vectors even after re-embedding. In VATES you edit or delete the relevant line directly and confirm the result by eye — a structure that also serves data-deletion requests, the right to be forgotten.
What is ES?
ES = EchoScript
The unit in which VATES holds knowledge. Two columns, n rows, separated by pipes (|). The pipe itself carries no meaning; it only marks a boundary. Meaning arises not in the symbol, but in the one who observes it.
IIIS = Inter-Intelligence Intermediary Syntax
An ES is also a syntactic structure designed as an inter-intelligence intermediary syntax (IIIS).
What goes inside
Any natural language, plus logical notation and more. Words or sentences alike.
What is IIIS?
The habitable zone
A syntax through which intelligences of different perception can observe the same structure — within the habitable zone where their perceptions overlap. That is IIIS.
Between intelligences
For now: AI and humans.
Observation across perceptions
The same ES is read by an AI as json, and by a person as a two-column table. Both observe one and the same structure — two columns, n rows, pipe-separated, carrying no meaning of its own — each through its own perception.