Put your material in — characters, chapters, rules, bills, history — and Liminalis reads it, connects it, and works
alongside you inside it. It doesn't answer from a general model's guesses about your world. It answers
from your documents, and it can tell you exactly which ones.
Three separate occasions, and none of them told her the whole of it.
Marisol mentions it obliquely in the amber hall — a warning dressed up as a favour. Later,
Corvin is more direct, but leaves out who commissioned it…
What's in context6,108 tok
Project Overview—found by always-include
Wren—found by reverse index · named in query
14.9 — What the Map Omits0.91found by rerank · bm25 + dense
The Amber Hall0.78found by graph walk · entity hub
Vignette — the borrowed name0.74found by motif · debts-unpaid
pinned recalled history free
The “what's in context” panel — every document that fed the answer, its score, and how it was found.
Representative composite of the application UI.
§ 00 Measured, not asserted
3.3×recall@5 0.269 → 0.876
1.000hit@10 from 0.318
−49%context tokens vs. the baseline
44hand-authored bench queries
Measured on an isolated copy of a real 135-document project. Pinned scaffold documents are excluded from rank positions, so
recall@5 measures retrieval rather than padding. The full ablation ladder — and the findings that were retracted —
are on Engineering.
§ 01
the product thesis
Testeda third lens in an unrelated domain needed no engine change
§ 01 The lens
One document engine. Swappable domain lenses.
A project is viewed through a lens. The same document set means different things depending on what you're
doing with it — and the lens changes the vocabulary, the interface, the generation stages, the validation rules, and the
retrieval barriers, over one shared corpus.
Underneath, nothing is domain-specific. Chunking, extraction, association, recall and context assembly don't know what a
chapter is, or a bill. That claim was tested: after the abstraction was factored out, a third lens was built over
roughly ten thousand pieces of legislation, and it required no special case in ingestion, chunking, extraction, retrieval,
assembly, the access barriers, the job model, or the agent API.
Neveron the save path — the editor never waits on a model
§ 02 What it builds while you write
The index isn't a black box. It's a document you can correct.
Background workers read every document as it changes and derive a knowledge layer from it — incrementally, resumably, and
never on the save path. What comes out is not an opaque embedding blob. It's readable, editable material that sits in your
project alongside everything else.
01
Entities & relationships
A fixed ontology of around fifty categories and two hundred typed relations. Each relationship carries two directional
summary sentences, so it reads back as memory prose rather than as a graph label.
02
Facts, atomic and sourced
Conservative statements about named things, each linked back to the passage it came from. Nothing floats free of its
source.
03
Searchable summaries
Section-level and document-level, and themselves retrievable — so a long chapter can be recalled at the right altitude
instead of all-or-nothing.
04
Motifs — retrieval by theme
Theme-language tags like debts-unpaid or made-not-born, so you can pull thematically-linked
material together regardless of where it sits in the folder tree. This is the piece nothing else in the category does.
Reverse Index.mdgenerated · deterministic
By character
Wrenin Wren, 14.9 — What the Map Omits, The Compassmentioned in 31 documents
Marisolin Marisol, The Amber Hallmentioned in 18 documents
By place
The Amber Hallin The Amber Hall, 7.2 — Small Merciesmentioned in 9 documents
By motif
debts-unpaidin 14.9 — What the Map Omits, Vignette — the borrowed name, The Compass
made-not-bornin 3.1 — The Guild Table, Corvin
The Reverse Index — every person, place, thing, thread and motif, showing what each is in versus merely
mentioned in. Built by deterministic string matching, no model involved, so it is robust to exactly the failures
a vector search has.
Why deterministic matters
A graph walk can miss an entity the extractor never found. A vector search can put a name far away in embedding space.
String matching fails at neither. When a user names a thing, the reverse index guarantees its document is
pinned into context — regardless of what the probabilistic layers thought.
§ 03
recall & context
Invariantnever head-slice a retrieved document — show what matched
§ 03 Retrieval, then context
A hybrid stack that was tuned against a bench, not against a vibe.
Content-defined chunking with sentence-window vectors. BM25 fused with dense retrieval by reciprocal rank, so the
cross-encoder gets to see lexical-only matches too. A reranker over passages rather than whole documents. Contextual
embeddings, where each chunk is embedded aware of its surrounding document — which matters enormously for prose full of
pronouns and implicit referents.
Progressive disclosure
Tiered context
Background documents render as a summary plus the passages that actually matched. Four guards stop it damaging
the answer — including one that refuses to summarize a document if the summary would come out larger.
Graceful degradation
Context budget
Over budget, it degrades in order of least damage: weakest documents first, then clipping the largest. Pins are never
dropped, and every clip is marked in the text so the model knows it holds a partial document.
Five rungs
Thinking levels
From a single deterministic pass, through plan-and-fan-out, to draft → self-critique → revise. Cost
scales with the rung, and the cheap scaffolding work runs on a cheap model.
The recurring lesson
The dominant bottleneck was a config default, not an algorithm.
— retrieval roadmap, twice over
A conversational default of five vector results, inherited unexamined, was starving a document corpus. Widening the
candidate pool moved recall@5 from 0.269 to 0.439 on its own. The reranker in isolation was worth +0.011; paired
with the widened pool it took the same metric to 0.871. Neither number is interesting alone. Together they're the whole
story.
§ 04
resonance
ModelACT-R activation over a Hebbian association graph
§ 04 The cognitive layer
Search is stateless. Memory isn't.
Everything above is machinery for “what best matches this query?” — and it gives the same answer forever. That is
right for a search engine and wrong for something that is meant to know your material. So underneath the retrieval stack
there is a second thing: a memory model that changes as you work.
A passage, an extracted fact, an entity, a summary and a conversation turn are all the same kind of object in one graph.
Every one carries an activation level computed the way the
ACT-R cognitive architecture computes it — recency and frequency in a single number. Connections strengthen
when both ends are used together, weaken when they aren't, and are eventually pruned. The graph learns the shape of
your project, and forgets the parts of itself nobody ever travels.
A
Recall with intuition
Because reach is gated by how warm a thing already is, the system can surface something that shares no keyword and sits
in no obvious embedding neighbourhood — purely because in this project, over time, those two things became connected.
B
Agreement is the signal
The query gets one pass. Every entity named in it gets its own, independently. Material reached from
two or more directions at once is promoted above material that merely scored well in one — so a hub connected
to everything can't win by volume.
C
It forgets on purpose
Unreinforced associations decay and are removed. Without that, years of extraction give you a dense mesh where
everything is faintly connected to everything and no signal survives. Forgetting is what keeps the graph usable.
Said plainly
This is the one part of the system with no benchmark number attached, and that's structural rather than
evasive: a fixed-corpus, single-shot bench scores a cold graph, and the whole point of this layer is that it is
history-dependent — it is supposed to get better in a specific project as that project gets worked in. Designing an
evaluation that can actually see that is open work. Until it exists, no performance claim is made for it.
The prompt has a shape, and the shape is the point.
Every turn is assembled into a deliberate order: an unchanging system block, then the stable pinned knowledge artifacts,
then the volatile query-dependent recall, then history, then the question. The first two segments are marked for
provider-side caching. The rest churns freely without ever invalidating them.
Staticsystem prompt · canon precedence · lens persona
The Project Overview and Project Structure documents are generated in the background and always pinned — so every
reply carries the project's shape without retrieval having to go and find it. They live inside the cached half.
B
Cache-aware by construction
On-demand expansions are routed into the volatile half specifically so they cannot invalidate the prefix for
every later turn in the conversation. That's a design constraint, not an accident.
C
Memory that persists between turns
A per-chat rolling context is serialized to disk: document scores, entity scores, section provenance and a
hierarchical summary. Recent turns stay verbatim; older ones compress. A follow-up keeps its material without
re-mentioning it.
Honest framing
Is this CAG?
Partly, and it's worth being precise. Liminalis is a cache-augmented prefix over a measured RAG stack:
precomputed knowledge artifacts that live permanently in the window, inside an explicitly cached prompt segment,
verified live to bill at roughly a tenth from the second turn onward.
What it is not is corpus-in-context. The whole tiering-and-budget effort exists because a real corpus does not
fit in a window, and a long chat will walk itself out of one no matter how tight retrieval is. Anyone claiming otherwise
at this corpus size is selling something.
Agents write into a draft queue as generated provenance. Direct write needs an explicit scope that is off by default,
and scopes only ever narrow.
Barriers before assembly
Access control runs before context is assembled, not as an output filter. A prompt-injected agent cannot coax out
material that never entered its context.
Never a silent clip
Truncation is declared both in-band and as a structured field. Retrieval that doesn't lie to you about what it showed
you.
§ 07 Where to go next
Read the parts that interest you.
There is a narrative explanation, a deep technical one with the bench attached, and three lens pages showing what the same
engine looks like pointed at three different problems.
Liminalis