Sources

A twin answers from your sources and nothing else. This is how material gets in.

Sources belong to your account, not to one agent: add something once and any of your agents can draw on it. This page is about getting material in; Knowledge and Agent knowledge is about sharing it between agents and keeping some of it off your public twin.

Where this lives#

Knowledge is two panels, and they're listed across the top of the page whichever one you're on:

Two panels, not four. What it knows and Guardrails are questions about one agent — what that agent has made of your material, and what it will decline — and this page is your account's, with no agent in context to answer them for. They live on the agent's own Knowledge page instead: see Agent knowledge — one agent's view.

The sidebar also lists the two underneath Knowledge while you're on it, when the sidebar is showing names rather than icons — that's a shortcut to the same two places, not the only way to reach them.

Each panel has its own address, so the one you use is linkable and bookmarkable: /app?view=knowledge&panel=sources and …&panel=add. A plain /app?view=knowledge link — including every one you saved before the panels existed — still works and opens Sources, and so does any panel name the hub doesn't have. Buttons that mean add something take you straight to Add.

Opening a source#

Press a source's title and it opens in a panel over the page, with the list still beside it: the same search, the same kind chips, the same grouping. You can move from one source to the next without closing anything, which is the point of it — checking six sources is six presses rather than six rounds of open, read, close.

What the panel shows you, for the source you're on:

"What we read from it" is a sample, not the document. You get the first part of it — enough to see whether it came out right — and when there's more than that the panel says so underneath, with the source's full word count. That's what it's for: a PDF whose text layer came out as garbled ligatures has a perfectly normal word count and lists exactly like a good one, and this is the only screen that would show you.

Closing it. The ✕ in the corner, a tap on the dimmed area above or below it, or Esc. On a phone the panel opens straight onto the source, with All sources taking you back to the list — and the list keeps its place, so you land where you left rather than at the top of it. Pressing the source you're already on takes you straight back to it.

Sources under From Knowledge Hub don't open here. They belong to the account rather than to this agent, so this agent's screen offers the one thing it can decide about them — stop using — and the rest lives on Knowledge.

Ways to add material#

Starting from nothing — until your twin has its first source, your Overview opens with four starting points rather than a list of instructions: import a page you wrote, paste something you wrote, upload a PDF or a doc, or bring what ChatGPT or Claude already knows. Each one opens into a single field and adds the source from that page, so the first source never depends on finding this tab. They go away as soon as one exists.

What happens after you add one#

Adding a source doesn't rebuild your twin on the spot. The source is stored and counted straight away, and it joins what your twin knows at the next train. Until then the twin keeps answering from the material it already had — adding something never takes a working twin offline.

A train runs when you start one, from Train → Train now, and it takes minutes. Train overnight — a switch on the Train stage, off unless you turn it on — hands that job to the night in your own timezone, gathering everything waiting into one rebuild. With it off, what is waiting stays waiting until you press Train. Nothing is lost in between either way: the queue just grows, and the Train stage shows how many changes are in it.

A small first source is the exception. An agent with no index yet can't answer anything at all, so if what you add is small enough to build on the spot, it is — you add it, the page comes back, and it is answerable. A big first batch waits like everything else: material you imported but haven't looked at yet is never trained on because a later, unrelated add happened to find an empty index. The message you get on adding always tells you which happened, so you never have to guess.

While a change is waiting, the row carries a Waiting chip and asking about it can still get an honest I don't know. The chip clears when the train lands. Bulk import waits like everything else, and that is a change: it used to train once at the end by itself. It runs on our servers, reads every page you picked, and leaves the lot Waiting — because a site usually holds a good deal you would not want answered from, and the moment to say so is when you can see what actually came in. Your agent is not answering from any of it until you press Train.

A source you untick in Train reads Held back instead, and keeps reading it until you tick it back on. Held material stays in your list and stays out of every train — it is not waiting for anything, and it costs nothing against what your plan holds. Ticking it back on puts it in the queue like any other change, so the next train picks it up.

A source with no chip is live and answerable, which is most of your list most of the time. That's why the settled state is the one with nothing to say: a badge on every row would be furniture, and the state worth interrupting you for is the one where your twin can't answer yet.

Removing works the same way, and that matters more. A removed source reads Removing and leaves the index at the next train — until then your twin can still answer from it. If you need it gone now, remove it and then press Train now; that rebuild is what actually takes the material out.

Removing is never paused. Your plan includes a number of trains a month, and running out pauses learning — but never a train that is taking something out. A source you delete leaves the index at the next train whatever your month looks like, and Train now stays available for it. Taking your own material down is not something you can run out of.

Deleting the last source empties the index and takes the page off the air at the same time — at that train, not at the moment you press delete, so a twin you have shared keeps answering until the rebuild lands.

If a rebuild fails#

Rebuilds retry on their own, and most problems clear themselves. When one runs out of retries, Knowledge says so at the top of the page: Last rebuild failed, when it happened, and — the part that actually matters — the date of the index your twin is still answering from.

Nothing is lost and your twin does not go offline. The previous index keeps answering the whole time, because the swap only happens at the end of a build that worked. What you lose is the newest material: anything added since that date is not in its answers yet. Try the rebuild again does not start one from here: Knowledge holds no training quote, so the same words take you to the Train stage, which states what a rebuild would add and what your month has left before anything is spent. The notice clears itself once a train finishes.

If it keeps failing, the source added just before it is the usual cause — remove it, rebuild, and add it back on its own to see.

Sources and citations#

A grounded answer cites the passage it used, and a visitor can open it and read the surrounding text. hiy's own summary of your material is never quoted to a visitor and doesn't count as one of your sources — only what you actually wrote does.

You choose whether visitors see the sources behind answers at all. Hiding them hides the citations from visitors — your own test chat still shows you everything. What no setting hides is an honest gap: when your twin says it doesn't know, it still shows how many searches it ran first.

When your sources don't cover something#

The twin says so rather than guessing, and shows how many searches it ran first. That question then queues for you, and answering it once turns your answer into a source the twin cites from then on. What goes into that source is your answer and the subject of what was asked — never the message somebody typed, which stays in your queue where only you read it.

Removing a source takes it out of the next rebuild — it and everything built from it are gone once that build swaps in.

Correcting one works differently, and the difference matters. Your correction is stored as a source of its own, and it is authoritative: where the two come up together, your twin answers from the correction — its wording, its numbers, its dates — and doesn't repeat the outdated claim or split the difference. It's matched to a question like any other source, and it's promoted above the rest whenever it's relevant enough to be worth citing, so write it in the words somebody would actually use to ask.

What a correction does not do is delete anything. The material you corrected stays in your sources, stays searchable and stays citable — so a citation somebody already opened still resolves, and you can still see what your twin used to answer from. It's out-ranked, not erased. If you want it gone as well, remove it too.