Technicals.
Four systems make K.ai personal without making it complicated: a private path to your Mac, local voice processing, private indexing, and managed model routing.
K.ai Link
Your phone pairs to one home Mac with a QR-pinned public key. Face ID authorizes short sessions. Pear handles discovery, direct connection, and a blind relay when direct connectivity is unavailable.
The K.ai server is not in the Link path. It does not signal, discover, route, relay, or store Link traffic.
After local retrieval, the Mac may use K.ai’s ordinary no-logging inference to write the answer from bounded, redacted evidence. That API request is separate from Link connectivity.
- Encrypted, authenticated envelopes
- Replay rejection and transport-key pinning
- Per-folder authorization and instant revocation
Local Dictation
A global Mac shortcut records only while the session is active. Sticky mode uses one press to start and one to stop; hold-to-talk records until the keys are released.
On this Mac uses a pinned WhisperKit/Core ML model. After the one-time download, transcription works offline, creates no provider charge, and never silently falls back to Cloud.
Cloud is a separate route selected at the start of a recording. It needs internet and K.ai credit, but uses less battery and memory.
- About 500 MB one-time model download
- Temporary audio deleted after every outcome
- Idle model weights unload after three minutes
Private RAG
You choose read-only folders. The Mac extracts and chunks supported files, creates semantic vectors with Apple NLEmbedding, and stores the index and relationship graph in a separate local SQLite database.
The source library, absolute folder locations, vectors, and database stay on the Mac. When cloud generation is used, only the few redacted passages selected for that question and their relative source labels enter the transient no-logging request.
- FTS5 literal search plus local semantic search
- Links, tags, backlinks, and folders connect related notes
- Answers retain file and PDF-page citations
LLM routing
K.ai chooses a task lane, not a model brand. Request shape, language, attachments, recency, and complexity select the lane; a small classifier resolves ambiguous depth and search decisions.
Each lane points to a maintained model configuration for everyday answers, multilingual work, reasoning, coding, vision, low-latency voice, or deep research. Models can change without changing the app or asking you to follow the market.
This is K.ai’s maintained router. It is not the UIUC LLMRouter package. We use the same task routing idea, with a smaller implementation built around K.ai’s own lanes, accounting, source rules, and fallbacks.
Cloud inference is transient. K.ai keeps a delivery copy only until the app receives it or it expires. Usage accounting stores the lane, token counts, latency, and cost. It does not store the prompt or answer.