Resonance Lattice · Docs

rlat documentation

Your AI assistant is brilliant about the world in general — and confidently wrong about yours. rlat builds a knowledge model from your files: one portable .rlat that knows its own world, so the assistant answers from it — grounded, cited, no invented sources. Same model, same questions: hallucination fell 19.6% → 0% while accuracy rose 56.9% → 92.2% (measured). The file also carries what has been learned about the world your documents cover — stable facts, standing rules, what was tried and failed — with a receipt for every claim.

Learned knowledge that provably changes answers — standing rules cut rule-breaking answers from 62% to 7% (R1), and held in a garden and a law practice (R1-X); recorded failures stopped 7/7 repeat recommendations (R2).
Private by design — capture is opt-in and gated: facts about people are dropped, validated at zero leaks (E2c).
A corpus that improves itself — gaps, contradictions, and stale facts are surfaced for your review; nothing lands without passing its gates (Self-improvement).
Receipts on everything — every passage and every learned claim traces back to its source (Claims).

New to the idea? Read What is rlat? — the whole thing in plain language. Otherwise these pages are written for someone setting rlat up for the first time: start with Getting started; the rest can be read in any order.

Start here

The whole idea in plain language — what it is, why it helps, and how it works. No jargon, no setup. Start here if you've never used it.

The measured case against the alternatives — LLM alone, file tools, vector DBs, pasting docs — and the honest list of when not to use it.

Getting started

Install rlat, build your first knowledge model, run your first query — in about fifteen minutes.

The handful of things rlat is for, and which command serves each one.

Self-improvement

How a knowledge model audits its own shape — gaps, contradictions, stale facts — and grows from trusted sources. The loop, the skills, and when you need an API key.

Concepts

One earned record from two sources — the assistant's sessions and your documents. How claims are earned, trusted, retrieved, and forgotten.

Memory & sessions

Recall at the prompt, capture at session end, zero labour. Hooks, the recall daemon, privacy, and the full command table — one subcommand of seventeen needs a key.

The single embedding recipe behind every knowledge model, and why there are no knobs.

Bundled, local, or remote — where a knowledge model keeps its source, and how to choose.

Reference

Every command and flag, grouped by what you are trying to do.

Which features need an Anthropic key, and how rlat finds it.

Every term in one place — knowledge model, source and insight layers, lens, drift, and the rest.

Short answers to the questions that come up most often.

Tune & integrate

Measured numbers — retrieval quality, hallucination, speed, token cost — and how to reproduce them.

Wiring rlat into AI-assistant skills for live, grounded context.

Querying a knowledge model hosted as a Microsoft Fabric function.