Chat
This chat
Welcome to Starpi
Ask questions about your knowledge base or about files you add to the on-device workspace. Answers cite their sources, and you can open every citation.
New here? Load three sample files of a fictitious project and ask about them. The files are read on this device; nothing is uploaded.
Preparing connection…
Knowledge base
Documents stored in Supabase that this session may read.
Knowledge graph
People, projects, budgets, technologies and regulations, and how they relate.
Entities (select one to see its details and relations)
Entity details
InfoSelect a node to see its details and connections.
Add knowledge
Search files privately on this device, or save text to the shared knowledge base.
On-device workspace
Private · on this deviceFiles are parsed, split into chunks and indexed with BM25 in a background worker on this device, and kept in this browser's storage unless you choose memory only. Nothing is uploaded or synced. Only when a model outside this device answers (the cloud assistant, NVIDIA NIM cloud through Starpi's server, or your own server) does file text go along with the question: in source mode the matching excerpts, in free chat the text of a file attached to that question (up to about 12,000 characters). Answers cite excerpts as [Doc: file, Chunk: n].
Drop files here or click to choose
PDF, DOCX, XLSX, PPTX, HTML, TXT, Markdown, JSON or CSV · up to 25 MB each
No files in the workspace yet.
Save to the knowledge base
SupabaseThe text is stored as a Markdown document owned by this browser session and protected by Row Level Security.
Saving to the knowledge base needs a browser session, which is not available right now. The on-device workspace above works as usual.
Diagnostics & benchmark
WebGPU adapter properties and a measured inference benchmark of the on-device model.
WebGPU adapter
Checking…- Vendor
- –
- Architecture
- –
- Device
- –
- shader-f16
- –
- System
- –
Limits and features
Features: –
Inference benchmark
A warm-up request, then a fixed prompt of about 64 tokens that generates exactly 128 tokens on the on-device model. Time to first token and decode throughput are measured on that run.
No model is loaded yet. Running the benchmark loads the model selected in Settings (download on first use).
- Time to first token
- –
- Decode throughput
- –
- Prefill throughput
- –
- Total time
- –
Settings
Assistant mode, providers and knowledge base connection.
Assistant mode
Which model Auto takes at a provider when you picked none of its models.
Local models are downloaded once, only after you confirm, and cached by this browser. GPUs without shader-f16 use the f32 variant; without WebGPU, processor (WASM) models run instead.
Models outside this device (the cloud assistant, NVIDIA NIM cloud, your own server) receive your question: in source mode with the matching excerpts, in free chat with earlier messages and an attached file. Models in this browser keep everything on the device. Without a model, Starpi quotes the knowledge base directly.
Knowledge base (Supabase)
Connecting- Project
- Supabase
- Access control
- Row Level Security · anonymous session
- Search
- BM25 and optional vectors on device · Postgres full-text and vectors (shared library)
- Chat history
- Checking…