peardata/user-guide/qvac.md

QVAC (local AI)

The QVAC tab is an on-device assistant for the connected agent. Models run on your desktop; the agent only answers metrics tools over P2P.

First-time setup

1. Open QVAC in the sidebar.
2. Continue through the welcome screen.
3. Review the system check (RAM, whether @qvac/sdk is installed).
4. Pick a model profile:
   • Lite — smallest download
   • Recommended — best default (tool calling)
   • Strong — larger model if you have RAM to spare
5. Download & load (or continue in tools-only mode if the SDK is not installed).

Re-run setup anytime with Setup in the QVAC header.

Using chat

• Type a question or click a sample prompt.
• Answers use live tools (host_snapshot, charts, anomalies, processes).
• Tool chips under a reply show which RPCs ran.
• Unload frees model memory (weights stay on disk).
• New chat clears the conversation.

Good questions

• “Summarize host health right now”
• “Any open alerts or anomalies?”
• “List charts related to disk”
• “Why might CPU be high?” (uses CPU summary + snapshot)

Requires a connected agent

Connect from the Connect tab first. If offline, tools return a clear error instead of inventing numbers.

Tools-only vs full QVAC

Mode	When	Behavior
Full QVAC	@qvac/sdk present + model loaded	Local Qwen (or other profile) + tools
Tools-only	SDK missing or load failed	Grounded formatting of RPC results, no LLM

From source / pear run

Install dependencies in the PearData repo (CI does this with npm ci):

npm ci
pear run -d .          # or: npm run start:client

Do not rely on a global or parent-folder @qvac/sdk install — release builds only include packages declared in this project.

From CI / release zip

Rolling and tagged desktop builds package @qvac/sdk and a Bare worker when packaging succeeds. Open QVAC → Setup and choose a profile to download a model.

If a build was made with PEARDATA_SKIP_QVAC=1, the tab stays tools-only (metrics tools still work).

Settings

Settings → QVAC:

• Default model profile
• Inject local knowledge (guide + chart catalog into prompts)
• Idle unload (minutes) — free RAM when idle; weights stay on disk

Settings

Settings → QVAC:

Preference	Effect
Default model profile	Lite / Recommended / Strong / Tool-tiny (used on next Setup)
Inject local knowledge	Keyword RAG over user-guide snippets + chart catalog
Idle unload (minutes)	Unload model from RAM after inactivity (0 = never)
Open QVAC tab	Jump to the chat view

Tools available to chat

Tools are tiered (core always; deep on Recommended/Strong; write for operators).

Tool	Purpose
investigate_host	Best first call — findings, hot charts, top processes
host_snapshot	Lighter live health + KPIs
hot_metrics / related_charts / compare_chart_windows	Investigation drill-down
search_charts / summarize_chart / summarize_charts / query_metric	Catalog + time series
get_weights	Correlation / influence for a chart
list_anomalies / list_alerts / get_alert	Anomaly / alert state
list_processes	Top processes (when enabled on agent)
query_logs	Journal / anomaly / audit lines
fleet_health / list_child_peers / storage_info	Fleet + retention
agent_health / node_info / db_info / list_contexts / get_chart / list_jobs	Agent meta
local_knowledge	Offline product how-tos
open_chart / open_view	Navigate the desktop UI
silence_alert / ack_alert / run_job	Operator only, with confirmation

Privacy

• No cloud API keys for this feature.
• Inference stays on your machine.
• Agent never loads multi-GB models for the default design.

See also docs/QVAC.md.