Vision / routing: loading…

Text camera & debug toggles

Camera for text chat i When checked, JPEG frames are sent with vision_session_id on text chat requests. Use Voice+Video for spoken questions such as “what do you see?”

Every reply uses live camera i Attaches a fresh camera caption on every user turn while vision is on (adds latency). Leave off for faster chat—camera context is merged only when your message matches vision phrases (“see”, “camera”, “what we’re doing”, …). With Voice+Video connected, toggling syncs immediately over the voice link; text+camera syncs on the next send.

LLM verbose (debug) i When enabled, extra technical detail from the assistant response (where available) is shown in the transcript for debugging.

Expand thinking blocks by default i When the model returns a separate thinking / reasoning section, this opens those blocks expanded by default in the chat log. Visibility still depends on the model and API returning that content.

WebSocket URL Saved instructions Session selection

Session memory actions

New session Rename Delete

System Instructions

Sampling for text chat and the next voice connect (toggle voice off/on to re-send). Model context length is set when the stack starts (e.g. local llama --ctx-size in start_current_stack.sh, overridable via LOCAL_LLAMA_CTX_SIZE); Text history lines limits how many prior chat lines are included. Text chat max prompt tokens/chars add separate caps on total prompt size before the request is sent.

Local vs Ollama: effective context is not the same. Long prompts can use more of the window on local llama than on Ollama, where the server may truncate inputs near a fixed token ceiling (large prompts were observed to cap around 32k prompt tokens while the same character padding counted higher on local). Max tokens is only a completion cap—prompt + reply must still fit the active backend’s context.

Loading limits for the active model…

Reset generation defaults

Offline lexical index (SQLite FTS5). Text chat: enable Local RAG below with agent tools. Voice / voice+video: snippets are prepended when VOICE_LOCAL_RAG is on (stack default). Chat attachments are mirrored under from_chat/ for later questions.

Offline only: tools run on this machine. HTTP connectors must use allowlisted hosts (default 127.0.0.1 / localhost; see TOOL_HTTP_ALLOW_* in stack docs). No public internet.

Local RAG tool settings

Local RAG (indexed docs)

First: Select route Second: Select model (used for Ollama route) Available Ollama models Loading models...

Vision: image captions always go through Ollama (VISION_OLLAMA_MODEL in the stack), independent of this dialogue route. With Local dialogue + vision + XTTS on one GPU, VRAM can spike; with Ollama for both dialogue and captions, use smaller tags (e.g. gemma4:e2b-it-q4_K_M) or run ./offline_setup/lisa_stack.sh status to inspect GPU rows.

Apply Routing (Restart Stack)

Runs offline_setup/start_current_stack.sh (ASR, XTTS, LLM, bot on :7861, HTTPS on :7860). Takes several minutes; this page may disconnect until the bot is back.

Restart full stack