Explorations

An exploration is a conversation with the agent — and the core working surface of Ronja. Questions, analyses, and builds all start as an exploration.

Starting one

Click New chat in the sidebar. A fresh exploration opens on the start screen — “Let’s start analyzing” — with example prompts you can click, a composer to type into, and (for the owner) the mode picker. You can also type directly into Ask anything… on the home page to start a chat with your first question. After your first message, the exploration names itself from what you asked; you’ll see a brief “naming…” shimmer, then the title. You can rename it later, or clear the title to let auto-naming take over again.

While you work

Beyond typing prompts, you can:

Attach files. Use the “+” menu (File), drag files onto the chat, or paste images. Data files like CSV and Excel become tables; images (up to 20 MB), documents (up to 100 MB), and text files (up to 5 MB) ride along as context — very long text files are clipped to roughly their first 50 KB, and the attachment pill shows a “truncated” badge when that happens. A zip is unpacked automatically while it’s small, and listed rather than unpacked when it isn’t — see Zip archives.
Attach tables. “+” → Table opens a picker; attached tables show as chips and link back to the table in your message. Clicking a table’s name in an answer opens it beside the chat, and its Map tab lets you walk the tables around it and attach them from there — see Tables.
Queue the next prompt. While Ronja is working, the send button becomes Queue. One prompt can wait in the “Up next” pill — run it early, or cancel it. Stopping Ronja clears the queue and puts the queued prompt back in the composer, so you can edit it and send it again. Only one prompt can be queued at a time, and it can’t carry attachments.
Comment. Every assistant reply can host a comment thread, with copyable deep links — useful when several people read the same exploration.

What accumulates

As Ronja works, a floating rail — the ledger — collects every openable artifact the conversation produced, grouped by feature. Pinned entries appear as they become relevant: Plan (Build mode), Methodology (Deep analysis), and Runs (this conversation’s workflow and agent runs). Clicking a reference pill or ledger entry opens the resource in a side panel, so the conversation stays your home base.

When a conversation gets very long

A conversation can outgrow what the model is able to read in one go. Ronja handles that herself: she keeps a running summary of the earlier part of the thread and works from that, so a long exploration keeps going instead of stopping.

When it happens mid-reply you’ll see a short note from Ronja saying so — it begins “Heads up: this conversation had grown too long for the model, so I summarised the earlier turns and carried on from that summary.” and tells you whether she had already started on your message. That note is for you; it is not part of the conversation Ronja reads back later.

The chat marks where the summary ends with a thin line reading “Conversation history above this point has been summarized — the assistant may not recall every detail verbatim”. That point often falls inside a work block, and then the line sits among the steps. While the block is collapsed it says “History summarized inside” instead; while Ronja is still working and the step it marks has scrolled out of the visible list, it says “History summarized above” until you click Show all N steps.

Nothing is deleted. The full transcript stays on the page for you to scroll, and Ronja can go back and read an earlier part of it when you ask her to — so if the summary lost a detail you need, say so and she’ll fetch it. She can read a different conversation the same way, which is the other half of Finding and keeping explorations. If you’re starting on a genuinely new topic, a new exploration (or a tangent) is usually better than growing one thread forever.

A long conversation opens on its most recent messages, and earlier ones load as you scroll up — Loading earlier messages… shows at the top while they arrive. If they can’t be loaded, it reads Couldn’t load earlier messages instead; click Try again.

When Ronja’s connection to the model drops

Ronja’s connection to the model can drop part-way through a reply. She starts that part again rather than leaving it unfinished, and what you see depends on what she was doing at the time.

If she was in the middle of a step, the step she was on is greyed out and marked stopped, a short note appears beginning “The connection to the model dropped part-way through that step…”, and the same step starts again below it. Seeing the same step twice is the expected shape of this, not a duplicate: the greyed-out one never finished, so nothing it was doing took effect, and the one below the note is the attempt that counts.

If she was in the middle of writing, the half-finished sentence usually disappears and a note takes its place, beginning “The connection to the model dropped part-way through my answer.” Ronja then writes that part again from the start. The disappearance is deliberate — half a sentence is worse than none, and rewriting from the start is the only way to avoid ending up with the same paragraph twice — so nothing has gone wrong when text vanishes here. Not every drop can be handled that way. If she had already finished a step in the same stretch of writing, that step is real work she is owed a result for, so nothing is thrown away: the half-finished sentence stays on screen and the reply ends there with a note saying the answer could not be finished. Send the message again to pick it up.

While she is trying again the status line above the reply reads “Retrying after a dropped connection…” instead of Thinking, so you can tell a retry from a slow reply. If the model itself is busy rather than unreachable, that line names the pause she is taking before her next attempt — “Waiting for the AI provider — pausing 8s before attempt 4…”. It is not a countdown: it describes the wait she has settled in for, and the waits get longer the longer the model stays busy, so a pause that grows from seconds to a minute is the sign of an outage rather than a blip.

Nothing is asked of you in either case; the retry is already running by the time the note appears. It happens once per message: if the connection drops again on the same message, Ronja tells you the answer could not be finished, and you can send it again. Like the summary note above, these notes are written for you and are not part of the conversation Ronja reads back later.

A stopped step is not the same as stopping Ronja yourself with the Stop generating button, which ends the whole reply rather than one step.

Documents you can download

Ask Ronja to write something up — “turn this into a one-page summary”, “draft a monthly report” — and she produces a formatted document that opens in a side panel right beside the chat, so you can read it without leaving the conversation. Ask for a change (“add a section on churn”, “make the title bigger”) and she updates the same document in place.

When Ronja links a file in her reply, it appears as a chip carrying the file’s name. Click a document’s chip and it opens in the same side panel; click any other file’s chip and it opens in a new browser tab. A file you don’t have access to stays a plain label with nothing to click.

Ask for the PDF again after a round of edits and Ronja replaces the same file rather than making a second one, so a download link you already shared keeps working and serves the new version.

To take it with you, open the document’s side panel and use the Download as menu:

PDF — a print-ready file.
Word (.docx) — an editable Word document.

Both are generated on the spot from the document you’re viewing, so a download always reflects the latest version. You can also open the document in its own browser tab. The document is saved as a file you can reopen or share later — by default it’s in your My Files, which only you and your organization’s admins can reach, so ask Ronja to save it to a shared folder if you want others to see it (a User Read-Only member’s documents always stay in their My Files). (File privacy works a little differently from exploration privacy: your saved files sit in Files, where an admin can turn on Everyone’s files and browse them — whereas an unshared exploration shows up only in an admin’s Activity page, which lists every chat in the organization, and an admin can open one there or by its link for governance.)

A Word file Ronja exported for you opens a little differently, because the document side panel shows only documents and PDFs. Ask her to open one and the panel shows the document it was exported from, headed with the Word file’s name and the caption Preview of the source document. Its Download (.docx) button gives you the saved Word file itself, not a fresh export. If the document has changed since the Word file was made, the panel says so — ask Ronja to export it again for the latest. A Word file with no document behind it, such as one you uploaded, doesn’t open in the side panel at all: download it from Files.

Version history

Every time Ronja changes a document or a PDF in place — an edit you asked for, or a fresh export to the same file — the version it replaces is kept. Click the History button (the clock icon) in the side panel’s header to see them: Current at the top, then each earlier version with when it was written and who wrote it. Choose one and the panel shows it, with a strip across the top: Back to current returns you to the latest content, Download gets a copy of that version, and people who can edit the file also see Restore this version and Delete this version, each of which asks you to confirm first. While an earlier version is on screen, the header’s own download button — Download as on a document, Download on a PDF — and Open in new tab are hidden, because they act on the current content; use the strip’s Download instead.

Restoring makes that version the current one again, and the content it replaces is kept as a version in its turn — so a restore can itself be undone from History. It’s still one document: its chip in the conversation and its entry in the ledger stay a single entry and open the restored content. Ronja keeps the 25 most recent versions of each document; older ones drop off the list. Anyone who can open the document can open its earlier versions too, so an edit that removed something sensitive doesn’t remove it from the history: to do that, delete the earlier versions that contain it (Delete this version) — a deleted version can’t be recovered. History appears only once a document has an earlier version, and not on the preview of an exported Word file. Restoring and deleting a version are done here in Ronja, not through the API — see Actions the API can’t take.

Finding and keeping explorations

The sidebar lists your explorations — those in the Space you’re in, if your organization has more than one — in four sections: Favorites (star an exploration to pin it there), Projects (each project expanding to the chats filed into it), Chats (your own recent history), and Shared (explorations others shared with you). Unread dots mark explorations with activity since your last visit. A conversation filed into a project is listed under that project rather than in Chats. Deleting an exploration is owner-only and also cleans up any scratch tables imported just for that conversation. Charts made in it are not deleted with it, but they stop being shared: after the conversation is gone a chart still opens for whoever made it (and for administrators), and no longer for the colleagues the conversation had been shared with.

Explorations are searchable too. Press ⌘K (or click Search in the sidebar), type at least two characters, and conversations whose title matches appear under Chats beside the features, tables, notes and apps that match — picking one opens that conversation. You only ever see explorations you could already open from the sidebar: your own, the ones shared with you, and the chats in a project you belong to.

You can also just ask Ronja to find one — “find the conversation where we worked out the refund window” — which goes by what a conversation was about rather than by its title. Ask her to read it in and she pulls the relevant exchanges from that conversation into the one you’re in, so a decision you reached weeks ago is in front of her without you retelling it. She reads only explorations you could open yourself, and she searches your recent history rather than all of it — see Limits and retention. What comes back is history: if it names a table or a number, it is worth checking that it still holds.

Explorations are private by default. Someone without access who opens your link sees only a denial: “You don’t have access to this chat.” A colleague gets into a conversation you started when you give them access: by sharing it directly, or by filing it in a project they belong to. Until you do one of those, your exploration never appears in anyone else’s lists or search. There are two exceptions. Administrators can view every chat in the organization, including private ones, for support and governance: they can open any exploration by its link, and the admin Activity page lists every chat with its theme, a short summary, who started it and how it ended. Unshared explorations still stay out of an administrator’s sidebar and search (see the roles capability matrix). And the conversation an automation or Saved Agent run produces opens for anyone who can see that automation or Saved Agent, from its run history. Either way it opens as View: you can read the conversation but not prompt in it, favorite it, or branch a tangent from it.

The owner shares via the Share button, which opens the “Share exploration” modal. You can add individual people or whole access groups; everyone starts at View and can be promoted to Full access per row. The difference:

View — can read the transcript, but the composer is replaced with “View-only chat”. While the conversation is paused for someone’s approval, a View reader sees “Waiting for approval to continue” instead of the approval buttons.
Full access — can prompt and work in the conversation alongside you.

The owner always has full access. Copy link is disabled until at least one share exists — a link to a private exploration is useless by design. “Make private (remove all)” reverts everything. Sharing also lets those people open the charts made in the exploration on their own pages, and removing someone’s share closes those pages to them again. An inline app in the conversation opens for them too — with their own access, not yours, so a table you can see and they cannot comes back refused inside it — and removing a person’s share also cuts off any copy of it they still have open.

When two or more people have prompted a shared exploration, a Contributors strip appears and each author’s messages are tinted and labeled, so you always know who asked what.

Branching off

If you want to test an idea without disturbing the thread, fork a tangent — a private branch you can later fold back in.