A LARGE LANGUAGE
RULE SYSTEM
WITHOUT A LARGE MODEL.
The first TORI Core is deterministic JavaScript. It is broad about how humans phrase recurring business questions and strict about which facts it is allowed to return.
01 / Normalization
Casing, punctuation, repeated spaces, common abbreviations, chat shorthand, Roman-Hindi forms and a curated typo map are normalized before interpretation.
02 / Intent families
Price, purpose, availability, delivery, purchase, link, contact, email, phone, website, location, hours, payment, refund, discount, support, order status, compatibility and comparison.
03 / Phrase families
Each intent carries many phrase forms. The goal is not one keyword but a large neighborhood of language that can point to the same intent.
04 / Entity resolution
Exact aliases, partial matches, token similarity and context identify the relevant product or service.
05 / Context scoring
Signals near intent phrases matter. Multiple signals can combine into a multi-intent result rather than overwriting one another.
06 / Fact retrieval
Once an intent and entity are known, only the corresponding configured field is retrieved. A price question retrieves price; a purpose question retrieves purpose.
07 / Contradiction checks
When two entities are similarly likely or a required fact is missing, confidence drops and the answer moves to review.
08 / Formal composition
Response templates differ by intent and entity type. The output is generated from real configured facts, not a single universal sentence.
09 / Teaching
A user can teach a new phrase as an explicit local rule. This changes the knowledge pack; it does not silently train a neural network.
Every decision is traceable.
The Core is meant to be inspectable. The user should be able to see the intent, entity, signals, missing facts and reason behind the confidence state.
message + business knowledge
normalize()
↓
intentCandidates()
↓
resolveEntity()
↓
scoreContext()
↓
factFor()
↓
compose()
CONFIDENCE
intent stack
entity
evidence
missing facts
formal response
GREEN = strong grounded result
AMBER = review required
NEUTRAL = not enough information