Settings reference

Bedrova’s settings are deep but searchable. A search box jumps to any setting and shows its current value. Every section supports per-setting reset, section reset, and full import/export with secrets redacted; migrations repair settings across upgrades. Many controls are capability-gated — they appear only when your hardware or runtime supports them.

Screenshot: Settings with the search open — added at launch.

General

App updates (Sparkle channel — see Network calls), launch at login, default model and document directories, and offline mode. Point the model directory at an external drive here; see Data & file locations.

Appearance

Theme (dark-first, light opt-in), density, font sizes, reasoning-block display, and reduce motion.

Chat

Default system prompt, streaming, reasoning display, tool-call rendering (see Tool calling & MCP), and transcript retention.

Model Defaults

The global defaults that every model profile layers over: context, output budget, sampling (temperature, top-p, top-k, min-p, penalties), reasoning, MoE expert policy, speculative-decoding defaults, cache policy, and logprobs.

Model Loading Guardrails

The guardrail mode — strict, balanced, relaxed, off, or custom — plus max model size, warning thresholds, and load-anyway policy. This is the dial behind the fit-check; the modes are spelled out in Memory & loading.

Runtime

Runtime-pack versions, channel, and pin/rollback for MLX text, GGUF/llama.cpp text compatibility, Core ML perception, search/embedding, vision, voice, image, document parser, and vector-store packs. Hold a pack a model depends on, or roll back a regression. See Runtime & engines.

Engine Selection

Routing policy per workload, a preference such as prefer-native or prefer-no-helper, and a candidate comparison that shows the reasons other engines were rejected. For text models, MLX is recommended when both MLX and GGUF are viable; GGUF is labeled broad compatibility, usually slower.

Hardware

Your SoC, unified memory, and effective usable memory, a live monitor, and pressure thresholds. Watch this during long runs for thermal or memory-pressure warnings.

Server API

Port, auth, LAN/Tailscale exposure (off by default, requires an access key), CORS origins, just-in-time loading, idle TTL, auto-evict, and max parallel requests. Background: Server overview and Parallel requests & scheduling.

Integrations

Hugging Face token (stored in the Keychain, never echoed) and mirror base URL, plus MCP server config and per-server/tool permissions. See Discover & Hugging Face and Tool calling & MCP.

RAG Embeddings

Default embedding/search model, vector store, document parser, chunking, reranker, visual document retrieval provider, image/audio index retention, and citation policy. See Documents & RAG and Embeddings.

Multimodal

Default vision model, Core ML perception models per task, pixel budget, max images per message, perception artifact retention, and remote-media policy (off by default). See Vision.

Voice Audio

Default speech-to-text and text-to-speech models, microphone permission, transcription language, diarization, VAD, and voice selection. See Voice.

Image Generation

Default image model, gallery location, seed policy, dimensions, steps, guidance, and LoRA directory. See Images.

Privacy Storage

Local-only parsing, document and transcript retention, diagnostics redaction, and secret storage. The full stance is in Privacy & security.

Diagnostics

Log level, diagnostics-bundle creation, benchmark history, and the effective-config viewer. See Diagnostics & support bundles.

Developer

Runtime manifest, fork SHA, feature flags, and a read-only raw settings JSON view.

Search, reset, and import/export

The search box matches setting names and shows current values. Reset works at the setting or section level. Export writes your configuration with secrets stripped; import merges it back, and migrations bring older exports up to date on load.