Requirements: Windows 10/11 · offline mode requires 16GB RAM minimum (0.8B model) / 24GB (2B) / 32GB (4B) · set aside ~10GB disk (models included)
Step 2: First Launch
Sidemate auto-creates a data directory on your local machine:
%LOCALAPPDATA%\Sidemate\data\
Expanded: C:\Users\<you>\AppData\Local\Sidemate\data\. All conversations, settings, and KB live here. No network calls, no cloud account.
On startup, the app defaults to the Chat tab (the top has 3 tabs: Chat / Knowledge Base / Settings — there is no separate "Download" tab; downloads live inside Settings).
If any dependencies are missing (e.g. llama-server.exe or model weights), a notice appears at the top of the page — go to Settings → Environment Check for details and one-click repair.
"All features ready" card on first launch — click "Start browsing" to enter chat"Settings → Environment Check" page — all dependencies loaded
On first launch, a welcome card appears — click the "Local AI" card to go straight to Settings → Model Download. Click "Download recommended bundle": the app picks an LLM tier based on your RAM (16GB→0.8B / 24GB→2B / 32GB→4B) and then downloads the knowledge base models (~4.5GB) automatically after.
Once downloaded, the model auto-loads and warms up — no manual steps. Go back to the Chat tab and start typing. If you don't need the knowledge base yet, you can download just the LLM (0.5–2.7GB) and add KB models later from the same page; you can also pick a tier manually or install from a local .sidemate package.
"Settings → Model Download" page — 3 LLM tiers (0.8B/2B/4B) + KB model + Install from local
Step 4: Pick Default Mode
The top-right of the Chat tab has a three-segment mode selector: Offline / Online / Parallel. Pick Offline to start:
🔒 Offline (default): Chats and KB stay local. Works offline.
🔌 Parallel: Local column + cloud column answer side by side.
☁️ Online: Everything sent to cloud provider (e.g. OpenAI / DeepSeek / Qwen).
You can switch any time in Settings.
Step 5: Test Conversation
In the chat page, send anything, e.g. "Tell me a joke".
Local model streams token by token
If you get a sensible reply, the whole stack works
KB needs the embedding model (bge-m3) and reranker model (bge-reranker-v2-m3), 4.5GB total. If you used the "recommended bundle" in Step 3, they're already installed; if you skipped them, download separately here:
Go to Settings → Model Download
In the "Knowledge base retrieval model" card click "Download KB model"
Once downloaded, click the top "Knowledge Base" tab
Drag in a PDF or Word file; wait for background processing (watch the progress bar)
Once done, open the doc and ask "What's this about?"
If the answer cites document content → KB works
Fully local: bge-m3 runs on your machine; documents never leave your PC (see Privacy Policy § 4).