Browse the corpus through its topical facets.
A proof of concept, under test.
Topical Facets searches a collection by the ideas running through it rather than by the words you happen to type. Every document becomes a network of word pairs; the pairs that carry information gather into facets, and a query travels those facets to reach documents that may never use your words at all.
Accounts are approved by hand, and each one is given the collections it may read.
Documents that matched the query but did not make the answer. Every other view lives at or below the ranking, so these were unreachable — and a document that answers part of your question can lose the scoring contest and still be the one you wanted. They are clustered here by what they share, and kept apart from the answer on purpose.
A query of several parts is several questions. Each concept below starts its own path — the documents carrying it, then one step along the nearest-neighbour clusters — and what is worth seeing is where the paths meet. A document two paths reach while carrying none of the concepts itself was asked for by nobody.
A tick is a document that carries the concept — it says the words. A ring is a document the concept only reaches, through its cluster. The two are never counted together: reach is inference, and a single seed inside a large cluster pulls all of it in.
Reach, not evidence: documents two or more paths arrive at after one step along the clusters. Worth reading for matters no one asked for, but which — through transitivity — might be relevant.
Exactly what would be sent to the model for the current query, before anything is sent. Building this costs nothing and reaches no third party.
Everything you kept, arranged. Place findings into sections, order them, and the note renders with its references — each one still carrying the query that surfaced it and the trail that reached it. Press read beside anything you kept to open it and clip a passage into the note. Your own words open and close it and introduce each section; the rest is quoted from what you kept.
A note is an arrangement of what you kept, not a document you type. It holds no content of its own: every block names something in your notebook and is read fresh each time, so correcting a keep corrects every note that cites it. Three of the four steps below happen somewhere else.
On the Search tab, press ☆ keep on a document, a cluster or the answer itself. What is saved is not just the item — it is the route: the query that surfaced it, the trail you followed, and the reason the engine retrieved it. That is what lets the finished note say where each claim came from.
Keeps go into a dated notebook. You can run several: the box beside kept names the one in use.
Select any text in a document and a ✂ clip this passage button appears at the selection. A clip is a quotation that stands on its own, so a twenty-page report does not have to travel into your note to carry one sentence.
Several clips from one document share a single reference number — they are one source, quoted more than once.
You can clip here too. Press read beside anything you kept — a document, a clip, a brief, a cluster or a group from the Content overview, a trail — and it opens over the note. Select a passage and ✂ clip into … files it straight into the section you are filling. A group opens as its docparts, a brief as its text with the documents it cites: open one to clip the source itself. A clip of the brief's own text is allowed, and the note labels it as the model's words — never as a source.
Read it through renders the note with its references below. Copy as Markdown, Download .md and Print / PDF each take it out of the application.
Draft it with the model → hands your arrangement to the LLM tab, where you can see exactly what would be sent before anything is sent. Building that pack costs nothing and reaches no third party; the model is told to follow your outline rather than regroup the material.
A dropped keep shows as “Missing”. It does not vanish quietly — if you delete something a note cites, the note says so where it stood.
You cannot paste a paragraph among the quotations. That is on purpose, and it is the whole claim this makes: a pasted sentence would be indistinguishable from one the corpus attests. What the corpus says enters as a clip that names where it was cut from; what you say goes in the opening, the closing and each section's text. Everything in a note can be traced back to a document you can open.
What this part of the corpus is about, before you know what to ask. The subjects that recur in it are pulled together into groups you can read, fold away, keep, and hand to a search.