Every research project establishes things. Names, numbers, relationships, claims that held up and claims that did not. LumaVista keeps all of it, attributed to where it came from, and puts it somewhere you can actually go and look.

Not a list. A palace — wings you can walk into, rooms inside them, and the individual pieces of knowledge on the shelves.

Walking the palace

The knowledge page opens onto your wings.

Wings are subject areas, drawn from the material itself rather than from folders you maintained. Research a market for three months and you will have a wing for it, without ever having created one.

Rooms subdivide a wing. They load when you open the wing, not before — a palace with thousands of pieces opens as fast as an empty one, because the entrance only ever reads the index.

Halls cut the other way. Where wings and rooms group by subject, halls group by what kind of thing it is:

Hall	What lives there
Fact	Established, checkable.
Event	Something that happened, with a when.
Discovery	Something the research turned up that nobody asked for.
Claim	Asserted by a source, not independently established.
Advice	A recommendation, with whoever made it attached.
Preference	Something you told LumaVista about how you like to work.
Sentiment	How sources feel about something — useful, and never mistaken for fact.

The distinction between Fact and Claim is the one that earns its keep. A claim carries its source and stays a claim until something corroborates it. Nothing gets promoted quietly.

Nodes are the individual pieces. Open one and you get the substance, the sources, the confidence, and — the part most systems cannot show you — where it has been used: which of your research runs and which conversations this piece of knowledge actually informed.

Availability. The palace view is in beta and off by default. If your knowledge page opens on a flat browsable list instead, the spatial view is not enabled for your account — ask and we will turn it on. Everything below the presentation layer is identical either way.

Browse, and Gaps

Two more tabs sit alongside the palace.

Browse is the flat list: filter, sort, search, no spatial metaphor. Faster when you know exactly what you are looking for, and the default when the palace is off.

Gaps is the inverse view — where your knowledge is thin. Not a scold; a research agenda. It is the most direct route from “what do I have” to “what should I look into next.”

What gets established, and how

Nodes

Each piece of knowledge carries:

A name and kind. Which hall it belongs in.
A description written to be useful to a future search, not to a human skimming.
Aliases. “RSC” and “Server Components” resolve to the same piece. Encounter either in future research and LumaVista knows they are one thing.
Confidence, 0 to 1.
Temporal validity. Some things have a shelf life; a piece can carry “true from” and “true until” so a fact about 2022 is not quietly served as current.
Sources. Every piece tracks the research node, document or conversation it came from.

Relationships

Typed, directed connections: uses, enables, part of, related to. Two matter more than the rest:

Contradicts — when something new conflicts with something established, LumaVista does not overwrite. It records the conflict and shows you both.

Supersedes — set when you decide the newer thing replaces the older. The old piece stays, with its confidence dropped, because “this used to be true” is itself worth knowing.

Where it comes from

Knowledge accumulates from three places, in the background, without slowing anything down:

Research projects — as each node completes, what it established is extracted.
Documents you upload — attributed to the file, so you can trace any piece back to the page it came from. See Uploading Documents.
Conversations — including preferences you state in passing. These start at lower confidence than research findings, because saying something is not the same as establishing it.

You can also write a piece yourself. Manual entries start at full confidence: you stated it, and that is the strongest signal available.

Confidence, and how it moves

Confidence is what stops a knowledge base from becoming a pile.

Set by: the extractor’s own assessment (a clearly-stated finding from a solid source scores higher than an inference), your approval (boosts to at least 0.8), or your edit (resets to 1.0 — you personally checked it).

Decays with disuse, gently: roughly half after a year untouched. Pieces LumaVista keeps drawing on stay sharp; pieces established once and never revisited fade toward the background.

Nothing is ever deleted automatically. Below 0.1 a piece stops being offered up, but it is still there when you go looking. Removal is your decision, not a policy.

How it sharpens the next question

This is the payoff, and it is why the palace is worth building rather than merely archiving.

When you start a project or ask a question, LumaVista searches what it has already established for you and uses it to sharpen what you asked.

You ask: “What’s new in frontend frameworks?”

It has: React 19, Next.js 15, RSC, Svelte 5, and a stated preference for practical examples over theory.

It runs: “What’s new in frontend frameworks, particularly React 19 / Next.js 15 (RSC, compiler changes) and Svelte 5? Focus on practical migration impact and code examples over architectural theory.”

You start from where you left off instead of from the top.

You choose how visible that is

Always preview — see the sharpened question before it runs; accept, edit, or discard.
Research only — preview for projects, silent in chat.
Silent with attribution — no interruption, but answers are tagged with which pieces of your knowledge shaped them.

The third is the one most people settle on. It keeps the work visible without putting a confirmation dialog between you and a question.

Scope

Project knowledge outranks workspace, which outranks global. Specific beats general, so domain knowledge from one project does not bleed into an unrelated one.

When things conflict

New knowledge that contradicts established knowledge is never resolved silently. Both appear side by side, and you choose:

Keep the new — the old is superseded, confidence dropped, still visible.
Keep the old — the new is dismissed.
Keep both — often the honest answer. Two statements can both be true in different contexts, and you can record why.
Merge — one updated piece.

No automatic resolution, because deciding which of two conflicting sources is right is exactly the judgment worth keeping human.