Finding an answer again

Every answer Prism gives is kept, and every answer has a link. Weeks later, the person who acted on one can produce the question, the answer, and the record of where its figures came from — without having kept the tab open or exported a PDF at the time.

This page is about what that means for what your installation stores and who can read it. It is worth reading before somebody asks.

For a reader

Your answers in the header opens your own history, newest first. Each entry is a question, when it was asked, and any marks worth seeing at a glance:

Mark	What it means
no data looked up	The answer made no tool calls and none had been made earlier in that conversation. Whatever figures it states, nothing fetched them.
figures from earlier	No lookups on this turn, but there were lookups earlier in the same conversation.
lookups unknown	No lookups on this turn, and no record of the conversation around it — usually a turn from before a restart.
cut off	The answer stopped part way.
failed	The turn errored before it finished.
no record	The turn ended before its provenance record was built. What it looked up is not known — and this is exactly the turn worth reporting.

Opening one shows the answer as it was given, with its tables and charts, and the provenance footer already unfolded. Copy link gives you an address that keeps working: paste it into a ticket, a message, or a decision record. The same button is on every answer in the chat, so you can take the link at the moment you decide the answer matters.

Export PDF produces the same document the chat produces, dated when the answer was given rather than when you exported it.

What you are looking at, and what you are not

An archived answer is replayed from what was stored, not re-asked. It is what Prism said at the time. It is not a fresh answer, and nothing on the page has been re-checked against today's data — the figures may since have changed, and that is the point of keeping the original.

For an administrator

What is stored, and where

Every turn already went into your datapond's app_events table: the question, the answer, the provenance record, timing, and any error. That has been true since long before this feature — what is new is that it can now be read back. Nothing about what is collected changed.

It is your database. There is no outbound connection and no Tetrate endpoint involved, exactly as with feedback.

Who can read what

• A reader sees their own turns and no others. The scope is applied in the query, not in the page.
• A link to somebody else's answer returns "no such answer" — the same response as an id that never existed. That is deliberate: a "not allowed" would confirm that a colleague asked a question with that id.
• A super admin sees every turn on the installation, and the page says so rather than leaving it ambiguous. This is the same tier that already reads every reported answer in the feedback queue.

Under an identity provider (iap / easyauth) "a reader" means one signed-in person. Where several people share one credential — the internal API key, if your install has one — their turns are one undifferentiated history, because there is nothing finer to separate them by.

"Part of this answer could not be checked automatically"

An answer sometimes carries an amber note above its figures saying that some results were too long to keep a full copy of, so the figures taken from them could not be checked.

It is a statement about Prism's record, never about the answer. The agent received every result whole. What the note reports is that the copy kept beside the answer holds only the beginning of one, so the automatic check that reads figures back against their source had nothing to read them against for that part.

How much of each result is kept is agentRunner.tracePayloadMax — 100,000 characters by default, which clears a normal page of rows with room to spare. If you see the note on most answers rather than on unusual ones, that value has been lowered, or an entry in agentRunner.extraEnv is overriding it: below about 13,000 characters a single page of pull requests no longer fits, and the note stops distinguishing the answers that deserve it.

Raising it costs the size of the app_events and run_events rows and nothing else — it does not change what the agent reads, so it cannot change an answer. The one thing it will not fix is a genuinely enormous result; there the note is correct and the remedy is the one it suggests, which is to ask for less at a time.

What it costs you to keep

Nothing is pruned, and there is no retention control in this release. The history goes back to the day you installed Prism and grows by a row per question. See Limits for what that means for the volume, and for the self test line that watches it.

If you need answers gone before then, that is a database operation today — app_events is one table, and turn_id and ts are both indexed. Deleting a row removes the archived answer and the permalink to it; the log line the turn also wrote is separate and is governed by wherever your logs go.

If your install has no datapond

Then there is no history to read, and the archive says so rather than showing an empty list — "nothing is kept here" and "you have asked nothing" are different sentences, and only one of them sends somebody hunting for an answer they are sure they got. An answer on screen can still be exported as a PDF.