Your First Chat - Inktomi93/OrbWeaver GitHub Wiki

From a fresh install to a chat you can steer, redo and fix. It assumes you've already connected a model (Connect a model). If you're coming from SillyTavern, the verbs you know (swipe, continue, impersonate, regenerate) are all here, just regrouped.

Tell the story who you are

The first time you open Orbweaver with a new account, a box asks Who are you in the story? It won't go away until you fill it in.

  • Name is required. It's what characters call you.
  • Description is optional. It's how characters should picture you.

Press Create persona and you're in. This is your first persona. You can rename it, add more and switch between them later. → Characters and personas

Friends who join through an invite link name their persona while they sign up, so they never see this box.

Start a chat with a character

  1. Open the Characters page and pick a character. No characters yet? Make one or bring your SillyTavern library over.
  2. Press New chat.

You land in the chat with the character's greeting (if the card has one) already waiting as the first message. If the card has alternate greetings, flip through them with the arrows on the greeting before you send anything. Once you've sent your first message, the greeting is locked in.

You can also press New in the Chats list, or type /new-chat in the message box. Those open a picker where you tick the characters you want. Tick one for a normal chat or several for a group chat, then start it, or choose Blank chat to start with no character at all. → Rooms and group chats

If no chat connection is set, a quiet line under the message box links to where you pick one, and Send stays off until you do. The same line comes back whenever a reply can't be sent, saying why and how to fix it. The model that wrote a reply can show with that message's actions; Message details & actions in your chat display settings controls it.

Send a message

Type in the message box and press Enter. Your message goes in and the character's reply streams in. While it's streaming, the Send button turns into Stop generating.

  • Shift+Enter starts a new line. Ctrl or Cmd+Enter always sends.
  • Enter to send is a switch under Settings, in Chat behavior, if you'd rather Enter didn't send.
  • Press Enter on an empty box and Orbweaver does the obvious thing. If the character spoke last, it continues their reply. If you spoke last (or there are no messages yet), it asks for a reply. The box's hint text tells you which. Both can be turned off in Chat behavior (Send continues the reply, Empty Enter generates a reply).
  • Drag a picture onto the box, or paste one, to attach it. → Images

The buttons around the message box

The row of buttons is mostly icons. Hover or focus any of them for its name.

Hover name Looks like What it does
Chat options three lines Chat-wide actions: rename or delete the chat, start a new chat with the same characters, select several messages at once, open the gallery of a character you own, and (host only) turn game mode on or off.
Message tools sparkles wand A menu of less common tools, covered below.
Draft your line theatre masks Writes your next line for you. Pick 1st, 2nd or 3rd person. It lands in the message box so you can edit it. Nothing is sent.
Try another reply circular arrow Rerolls the character's latest reply.
Generate reply play triangle Asks for a reply without you writing anything. In a group chat it opens a menu: Let the room pick, or pick which character speaks.
Continue the reply fast-forward Makes the character's latest reply keep going.
Send message paper plane Posts what you wrote and gets a reply.

While a draft is being written, a Stop drafting your line button appears. It keeps whatever has been written so far.

If a button is greyed out, hover it. The tooltip says why. Try another reply and Continue the reply both need a character's reply as the last message, so they wait until one is.

The composer with text typed and the buttons lit up

Typing first turns a button into a steer

This is the idea that makes the row worth learning. Type something in the box and four of the buttons light up. Their tooltips change to say they use your typed text as direction. Press one and your text becomes guidance for that one action instead of a message.

  • Type she hesitates, then lies about the key and press Generate reply: the reply is written with that nudge.
  • Type make it angrier and press Try another reply: a new version of the latest reply, steered that way. The text stays in the box, so you can reroll again and again with the same direction.
  • Type end on a cliffhanger and press Continue the reply: it keeps going, pointed that way.
  • Type agree, but reluctantly and press Draft your line: your text steers the line it writes, then gets replaced by the draft.

Send is the exception. Send always posts your text as your own message.

In the Message tools menu, Input has Recover input, which brings back the steers you've used recently in this session.

Swipes: other versions of a reply

The character's latest reply has a small strip: a left arrow (Previous variant), a counter like 2 / 3, and a right arrow (Next variant). The app calls swipes variants.

  • Step back and forth to compare versions. Whichever one is showing is the one the chat continues from.
  • At the newest version, the right arrow writes a fresh one (Generate a variant). With only one version, the strip is just a Generate button.
  • The left and right arrow keys do the same when you aren't typing in a box.

The strip only appears on the latest reply. Older replies keep whichever version was showing when you moved on.

The Message tools menu

The sparkles button opens a menu with the rest:

  • Corrections... is for when a reply is almost right. See below.
  • Regenerate is a plain reroll of the latest reply. It ignores whatever you've typed. Use Try another reply when you want your text to steer the reroll.
  • Continuation has Undo continuation and Revert continuation. See below.
  • Simple send posts your message without asking for a reply.
  • Offer choices asks for a reply that ends with a few numbered options, this once. To get them on every reply, the host turns on Offer choices under Host controls, Storytelling, in the This chat tab.
  • Recover input and Clear input bring back an earlier steer or empty the box.
  • Attach images & video, plus the image tools, are in the Media group. → Images
  • In a game chat with Plot steering on, a Story group appears with plot steers. → RPG mode

Corrections: fix a reply without a blind reroll

Choose Corrections... and a box titled Rewrite the reply opens. Say what's wrong in the Correction instruction field ("she wouldn't say that, she's furious"), and optionally flip any of the style switches: More concise, Expand, Novella prose, Internet-RP style, Literary style, Past tense, Present tense. Press Rewrite and the latest reply is rewritten as a new version, so the original is still one swipe away. If the message box already had text in it, it's copied into the instruction for you. Closing the box without pressing Rewrite keeps what you wrote.

It needs a reply to work on, and an instruction or at least one style switch.

Undo a continuation

If Continue the reply rambled on too long, open Message tools, then Continuation, then Undo continuation. The reply goes back to how it read before. Revert continuation brings the added text back.

You can do the same on any reply, not just the newest. Open that message's menu (below) and use Undo last continuation or Re-apply continuation. Both stay greyed out until a continuation has happened on that version.

Actions on a message

Hover over a message to see its buttons. On a touch screen they're always showing, and Edit, Fork and reactions live in the three-dot menu.

  • Edit message (pencil) edits the text in place. Enter saves, Shift+Enter adds a line, Esc cancels. There are also Save edit and Cancel edit buttons.
  • Fork chat here (branching lines) copies the chat, up to and including that message, into a new chat and takes you there. The original stays in your Chats list. Use it when you want to try a different direction without losing this one. In a fork, the chat's options menu has Back to parent chat, which opens the chat you forked from. It only shows if you're a member of that chat.
  • Add a reaction (smiley) drops an emoji on the message, unless reactions are switched off for the chat.
  • More message actions (three dots) holds everything above plus:
    • Hide from AI takes the message out of what the model sees, while it stays on screen, dimmed. It becomes Unhide from AI once hidden. For out-of-character chatter or a bad turn you want to forget.
    • Copy copies the message text.
    • Undo last continuation and Re-apply continuation, on character replies.
    • View wire trace... on character replies, for the host. It shows the exact prompt that version was written from. → How a turn is built
    • Delete removes the message, after a confirmation. It can't be undone.

In a shared chat, you can edit, hide and delete your own messages, and the host can do it for anyone. Only the host can fork a chat with more than one person in it.

To clear out several messages at once, use Select messages... in Chat options.

Tool cards. Where a tool card offers Hide from AI, you can leave the card visible while keeping that call out of later prompts. Show to AI restores it. Delete tool call removes the card and its later prompt replay, but does not undo anything the tool already changed. These controls apply to that reply version.

Change things for just this chat

Open the chat's details with Show details. The panel has tabs, and This chat holds everything you can bend for this one chat without touching your characters, presets or global settings. Its sections open and close, and most start closed. In a shared chat, members can read most of it but only the host can change it, except Macro picks.

  • Field overrides swaps in a different main prompt, post-history or scenario for this chat only. Leave a field empty to keep using the character's or preset's own.
  • Injections adds notes of your own to the prompt.
  • Documents and World books (what SillyTavern calls lorebooks) attach reference material to this chat. → Lorebooks
  • Regex shows which regex scripts run here, and lets you attach more.
  • Macro picks is where you answer any questions your preset asks about this chat. It's empty unless the preset asks something.
  • Host controls is for the host, which in your own chats is you. It holds Background (a picture just for this chat), Group behavior (group chats only), Storytelling (including Offer choices for every reply), Reactions, Tool use and Rules. → Automation

The panel's Preview tab, also host-only, shows what the next turn will send to the model before you send it.

Its prompt + history bar shows how much of the available prompt room is used, with separate segments for the card, world info, game state and history. The room is smaller than the full context window: space is held for the reply, and a model-window limit also keeps a safety margin for token-count error. The explainer names each amount. If the window is unknown, the preview says so instead of drawing a made-up capacity ratio. If the reply reserve leaves no room for history, it says only the newest message fits. These are fitting estimates, not provider billing counts.

In shared rooms, the host's Storytelling → Members' persona lore switch controls whether other players' persona-attached lorebooks enter the prompt. Turning it off leaves the host's, room's and characters' lorebooks in use; members see a notice under World books. → Rooms and group chats

There's no model picker for a single chat. Replies run on the host's Chat model, set under Settings, Connections, Model roles. In your own chats that's you, so change it there to change models.

Turn on Memory

Long chats outgrow what a model can read at once. Memory is how Orbweaver keeps hold of the older parts. How summaries get made and recalled is its own page. → How memory works

Memory is off until you turn it on. It costs model calls, so turning it on asks first.

  1. Open Settings, then Chat behavior, then Memory, and switch on Remember earlier in long chats.
  2. A Turn on Memory? box opens. It says what memory costs (about one extra summary call on your Utility model for every 8 messages, once a chat grows long), which Utility model pays for it, and what happens to the chats you already have.
  3. Press Turn on Memory.

It's one switch for your whole account, not for one chat, and in a shared chat the host's switch is the one that counts. Switching it off doesn't ask.

Memory needs a Utility model, a cheaper model for background work, because that's what writes the summaries. If you picked one when you added your first connection, you're set. If not, the box says none is set and offers Open Model roles, where you pick one on a connection with Allow background work on. → Connect a model

With memory on but no Utility model, nothing fails. Summaries are paused, and the Memory section says so beside an Open Model roles button. Orbweaver keeps indexing the scenes themselves, so searching your past scenes still works, but there's nothing to recall into the prompt until a Utility model is set. If a summary ever can't be written, you'll see a notice that long-term memory couldn't update. The reply itself is unaffected.

Turning memory off stops new summaries and stops recall.

Chats you already have

Turning memory on doesn't go through your library unless you ask it to. A chat you already had builds its memory the next time it gets a reply. At that point Orbweaver summarizes the chat's whole backlog, not just the newest part, so the first reply in a long old chat starts a lot of background work at once.

To do them all at once, tick Also build memory for all my existing chats now in the Turn on Memory? box. The box says about how many model calls that is before you agree, and it runs the Memory backfill job in the background. The box only offers it when there's something to build and a Utility model to build it.

When an import brings in chats while memory is on, the import's report offers Build memory for imported chats. It says about how many model calls that is for just those chats and asks before it starts. Skip it and each imported chat builds its memory at its next reply. The free part, the search index for those chats, builds on its own.

You can also run the Memory backfill job yourself under Jobs (the Memory section has a Go to Jobs button). Like every job that calls a model, it shows about how many calls it will make and asks before it starts. It reads through the chats you host, writes their summaries and rolls them up, and reports what it built and how many chats it had to skip because of an error. You can leave and come back, and it can be re-run safely, because anything already built is left alone. It won't start while memory is off, so turn memory on first.

The Corpus's Run the understanding pass button also runs the backfill as one of its steps, when memory is on. It asks first with its call count, and it stays greyed out, with an Open Model roles button beside it, until you have a Utility model. Bringing old chats over in the first place is on Coming from SillyTavern.

Good to know

  • You can only swipe the latest reply. To explore a different timeline, fork the chat from any message.
  • There's no text-completion mode; replies go through a chat API. → Coming from SillyTavern

Where to go next