Chat playground

The Chat tab is a full chat client for your local models:

• Reasoning levels: the Configs popover has a Reasoning dropdown whose levels come from the model's own chat template (typically None / Low / Medium / High / Extra high; a template with no ladder is just on/off). Higher = the model thinks longer before answering. Changing the level rewrites the top of the system prompt, so the conversation is re-read - pick one and stay on it. Reasoning streams into a collapsible box. The Thinking Budget (Settings) caps reasoning tokens for models with no ladder; it does not apply to the template levels.
• Attachments (paperclip): images go to vision-capable models, but you can also attach PDFs, Word documents (.docx), and plain-text files (code, Markdown, CSV, JSON, YAML, logs, config...). Those are extracted to text in your browser and folded into your message, so any model can read them, no vision needed. Attaching an audio file transcribes it first with the ASR backend (which swaps that model into VRAM - you're warned before it does). Note: PDF extraction isn't OCR, so a scanned PDF yields no text and says so.
• Tools: several are always on - the help wiki, a clock/date tool, an exact calculator, a unit converter, and YouTube (search, transcripts, comments). Optional toggles in the Configs popover: Web Search (SearXNG; also enables fetch page, so the model reads the real page instead of a snippet), Weather & Feeds (live weather, any RSS/Atom feed), Memory (lasting facts about you, reused in every chat, editable in Settings -> Memory), and QM Tools (inspect and, with your approval, tune this Quartermaster instance - see its own article). Tool calls show up as chips in the transcript.
• Rewrite mode (tools menu): send text with an instruction and get a side-by-side word-level diff (added / removed / changed) instead of a chat reply - good for editing prose or code in place. It's a self-contained transform: no tools run.
• Shopping assistant (tools menu): a staged buying helper. It first pins down the brief - budget, currency and region, brands, must-have specs, condition - asking a short set of questions as clickable chips rather than a form; then it searches and reads actual shop pages for real prices; then it answers with a comparison of product cards (image, price, shop, specs, why it made the list, and a verdict) instead of a wall of text. Shopping preferences (country, currency, preferred shops) are a standing setting in the Configs popover, so it doesn't ask every time. Currency conversion is available to it too.
• Read aloud: the speaker button under any assistant reply speaks it with the TTS model chosen in Settings. It dims while the audio is still being generated and turns bright white once the whole reply is ready; click again to stop. Next to it sit a volume slider and a speed menu (0.25x to 2x, pitch preserved) that apply immediately, mid-playback included, and are remembered per user. While it plays, the sentence being spoken is highlighted orange and everything already read is dimmed, so you can follow along or see where you stopped it. Code blocks, reasoning boxes and diagrams are skipped - only the prose is read.
• Per-chat instructions: standing instructions layered on the built-in prompt, saved with that conversation.
• Runs server-side: once you send, the answer is generated by the server, not your browser tab - so closing or refreshing the tab (or losing the connection) doesn't stop or lose it; reopen the chat and the in-progress answer is still streaming. Searches, tool calls and approvals all continue server-side.
• History & auto-compaction: chats are saved server-side per user. When a conversation reaches ~80% of the model's context window, the oldest turns (all but the most recent ~6 messages) are summarized by the model into a brief so the chat can continue past the raw context limit - the full transcript still shows in the UI, only what's sent to the model is compacted. A live bar shows how full the context is - click that bar (or type /compact) to compact on demand, which is worth doing before a long tool-heavy answer rather than waiting for the 80% trigger.

This page is generated from the help wiki that ships inside the app: the same text you get from the Help button, and the same text the playground assistant searches. Corrections go to wiki_articles.json.