Tavina

Chat basics

๐Ÿ“– This content is synced from the Tavina Handbook.
The canonical version always lives at https://tavina.ai/docs/friendly/conversations/chat-basics

Chat basics

Chat is the main way you work with Tavina. You can ask questions, provide context, upload files, and ask Tavina to draft, summarize, plan, compare, or troubleshoot.

Ask with enough context

A useful request usually says:

  • what you want Tavina to produce;
  • who the answer is for;
  • what information Tavina should use;
  • what Tavina should avoid;
  • whether you want a short answer, detailed explanation, table, checklist, or draft.

Example:

> Summarize this Source for a busy homeowner. Focus on maintenance tasks, cite the sections you used, and tell me if any pages were skipped.

Use follow-up questions

You can refine Tavina’s answer by asking follow-up questions such as:

  • “Make it shorter.”
  • “Show the assumptions.”
  • “What evidence supports that?”
  • “What did you not check?”
  • “Turn this into next actions.”
  • “Open my task list.”
  • “Take me back to the panel we were using.”

Visible panel actions

When Tavina opens a work panel from chat, it should leave a visible trail: the panel opens in the app, a toast appears, and a chat line says what Tavina opened. If the same panel is already open, Tavina may keep it open instead of reopening it. If you are in focus mode, Tavina should not force the panel into view; it should only cue the suggestion.

Panel history is short and session-based. “Take me back” returns to the previous work panel when one is available. If there is no previous panel, Tavina should say so instead of pretending it moved.

While Tavina is working

Multi-step tasks can take a while. While Tavina works, the prompt field at the bottom of the chat shows a live elapsed timer (for example 4m 12s) that updates every second, along with recent activity such as which step Tavina is on. The timer keeps moving even when the status label itself stays the same for a few minutes, so a long-but-healthy task looks alive rather than frozen.

If Tavina stops making progress for a longer stretch, the status line switches to a clear may-be-stuck warning with a Nudge option. That warning always takes priority over the timer. If you see the timer advancing, Tavina is still actively working and no action is needed.

While Tavina works through the steps of a task, its progress notes appear as a single status line that updates in place โ€” one line that changes as the work moves forward, rather than a stack of interim messages building up in the chat. When the final answer arrives, that line steps aside: the reply itself lands in the conversation, and the trail of in-between steps stays available in the activity summary attached to that reply. Nothing interim is kept in the chat history itself, so long tasks read cleanly afterward. If your Tavina has not picked up the update yet, interim notes keep appearing the familiar way as separate messages.

Screen-reader users hear liveness updates at most about once per minute rather than on every tick, to keep voice mode quiet.

The context dial: how full the chat is

Next to the disclaimer line under the message box, every chat shows a small round context dial with a percentage. It shows how much room the latest model call used, measured against a standard context display scale of 500,000 tokens (500K) that is the same for every subscriber.

The display scale is a Tavina standard for showing usage honestly. It is not the AI model’s maximum capacity, and it is not the budget Tavina uses to trim long conversations. Tap the dial to see how much room is left and which parts are using the most space; the details list the conversation trimming budget separately under the configured limits, and tell you plainly when a part had to be trimmed to fit. Trimming can happen even when overall usage is low, because sections are trimmed to their own assembly limits while a request is being built.

The dial turns amber when the chat is filling up and red when it is nearly full, with a suggestion to sweep the chat clean so answers stay sharp. When a model call uses more context than the display scale allows, the dial does not show a comforting capped 100%: it says the measured input exceeds the context display scale, shows the true percentage and how many tokens over the scale the call was, and again suggests sweeping the chat clean. When usage is not known for the last model call, the dial shows unknown rather than a made-up number.

Tavina trims its working memory automatically when usage crosses the conversation trimming budget โ€” a separate budget Tavina sets for the chat, which can be smaller than the 500K display scale โ€” so long conversations stay reliable without you babysitting the dial.

The dial is only a size indicator. It shows sizes and percentages, never the text of your prompts or Tavina’s instructions, and it remembers its state when you reload the app.

When answers stream in as they are written

As Tavina updates roll out, answers that do not need more tools can start appearing in the chat while Tavina is still writing them, instead of arriving all at once at the end. If your Tavina has not picked up the update yet, replies keep arriving the familiar way. A few honest limits apply:

  • the streamed text is provisional. If Tavina corrects itself or the final answer differs, the visible text is replaced cleanly, and only the finalized answer is kept in your history;
  • parts of a reply that involve tool calls, reasoning, or uncertain content still wait and arrive once they are final;
  • streaming shows Tavina’s progress sooner, but it does not make the model generate any faster;
  • what you hear in voice mode is still the finalized answer, never a provisional draft.

Replies written around tool steps

When Tavina does work with tools as part of an answer, it may write something substantive for you first โ€” for example an explanation or a plan โ€” before it runs the tools, then confirm the result afterwards. Those substantive parts written before or between tool steps now appear in the chat as their own parts of the reply, so nothing you were meant to read gets tucked away.

Short transitional notes while working (a quick “checking that now”) still stay in the thinking section described below rather than as chat parts, and the parts you see are always text that was finalized, never a provisional draft. On Tavina versions before this change, only the text after the last tool step was shown as the reply.

Reviewing Tavina’s thinking after a reply

Some models show their reasoning while they work. That thinking used to disappear as soon as the reply finished. Now every completed reply keeps its reasoning in a collapsed section attached to that exact message.

The collapsed section is labelled with how long the thinking phase took, for example Thought for 12s. Choose the section to expand it and read the full reasoning trace. Choose it again to collapse it back. When no duration is available, the section simply reads Thought process.

The thinking stays with the message:

  • after you reload the page or switch devices;
  • when you return to older conversations from history;
  • on multi-part replies, where each part keeps its own reasoning and its own duration;
  • when a task is interrupted or a safety confirmation replaces the visible text, the reasoning shown up to that point is preserved.

Reasoning is shown for transparency, so you can see how an answer was approached. It is not a transcript of private data, and it does not replace checking an answer before you rely on it.

Seeing what Tavina did: the Actions summary

Replies can also summarize what Tavina actually did while answering. When a reply involved memory or tools, the collapsed section attached to it gains a short hint, for example Thought for 12s ยท Memory, Searched the web, so you can tell at a glance whether an answer came from recalled context, searching, saving, or tools.

Expanding the section shows an Actions list before any reasoning text. Each line covers one kind of action with honest counts:

  • memories supplied automatically are described as supplied, never as a search;
  • memory searches show how many matches were found, including zero matches;
  • saves only claim success when Tavina confirmed them, and failures show separately;
  • tool calls summarize how they ended.

Replies with no recorded activity show nothing extra. The summary keeps only counts and categories โ€” no message content, prompts, or memory text โ€” and it stays with the reply after reloads and when you return to older conversations. If a reply was interrupted, the summary lists only the actions that actually completed.

Files and Sources in chat

When a file is document-like, Tavina may handle it through the Knowledge Library as a Source instead of placing the full extracted text directly into the prompt. This keeps long documents from flooding the conversation and lets Tavina use retrieval and chunking.

If Tavina answers from a Source, ask for citations or coverage when the answer matters.

You can attach up to 20 files to one message. Tavina should use every accepted attachment with that prompt, not just the first. If you select too many, Tavina should tell you the limit and leave the extra files out instead of silently dropping them.

Images show as separate thumbnails. Choose a thumbnail to open a full-size preview for that image. Non-image files show as compact file cards or chips. Those visible cards should not expose private upload paths, hidden metadata, or full extracted document text.

If one attachment fails validation, upload, or security scanning, Tavina should show the error and exclude that file from the message context. Safe attachments in the same prompt can still be sent.

Image analysis recovery

If image analysis gets stuck or is unavailable, Tavina should explain the problem in plain language instead of showing raw technical error text. When analysis is only temporarily unavailable, Tavina now sends the attached photo directly to your assistant’s model whenever that model can view images, so the reply still uses your photo. If the model cannot view images, the rest of your message still goes through and the assistant asks you to resend the image later. You may see recovery options in chat, including:

  • Discard stuck image and retry โ€” retries the last prompt without the image that could not be analyzed.
  • Start fresh conversation โ€” starts a new server-side conversation when the current conversation context is blocked or no longer useful.

Use Discard stuck image and retry when you still want Tavina to answer the text of your prompt without relying on that image. Use Start fresh conversation when you want to leave the blocked context behind and begin again. These recovery actions, and the temporary-unavailable fallback, never send raw image data to a model that cannot view images.

Targeted retrieval is not whole-document review

For long Sources, Tavina can use targeted retrieval to find relevant chunks. That is useful for specific questions, but it does not mean Tavina read every page.

When you need full coverage, ask for whole-Source traversal or a coverage-aware summary. Tavina should report what was processed, skipped, failed, or still pending.

Memory in chat

Tavina may use memory to remember useful preferences or context. Memory is not a perfect record. Correct Tavina when it recalls something incorrectly, and avoid relying on memory as the only source for important facts.

Conversation boundaries and archived segments

Long-running chats can move into a new conversation segment after a day change, a long idle period, or a manual reset. When that happens, Tavina keeps the current chat focused on the new segment instead of automatically reloading the older messages into the live transcript.

If a previous segment is available, the chat may show an archived-segment banner. Use the banner when you want to review the earlier messages. The archived view is read-only and separate from the active chat, so new messages still belong to the current segment.

If you want Tavina to use something from the earlier segment, mention what you want Tavina to carry forward or ask Tavina to look back at the previous context.

Sweeping the chat clean

When the context dial says the chat is getting full, you can start fresh without losing anything. The broom icon next to the disclaimer line under the message box is labeled Sweep chat clean.

Before the sweep runs, Tavina asks you to confirm and shows exactly what happens: the working context for the current chat โ€” the in-flight conversation and Tavina’s working notes โ€” is cleared, while your long-term memory, files, artifacts, and the chat title are all kept. Canceling is always safe.

After the sweep, the visible chat starts fresh and shows the notice “Chat session refreshed. All conversation history remains saved.” Your previous conversation is not gone: use the View previous conversation link in the notice (or Settings > Data > Conversations > Recent Conversation Log) to review it any time. The next reply comes from a clean slate, and the context dial drops back to low usage.

Review before acting

Before you rely on an answer, check:

  • whether Tavina used the right files, Sources, Catalogs, or app connections;
  • whether citations are present when you expected them;
  • whether the answer says anything was partial, failed, or unavailable;
  • whether the topic needs professional review or current support guidance.

Related pages

Scroll to Top