Configuration¶

Each installation keeps its settings in the ignored file obsidience/obsidience.toml. Start from the tracked obsidience.example.toml and never store credentials in it.

The loader is obsidience/harness/config.py. A key that matches a field of its Config dataclass sets that field (path fields become paths); every other key or table lands in CONFIG.extras, where the subsystem that owns it reads it with a default. A key you do not set keeps the default shown below.

How this reference was derived

The core keys are the Config dataclass fields. The other tables list every CONFIG.extras.get(...) (and CONFIG.extras[...]) read in the Harness, with the default written at that call site. Keep the TOML tables ([memory], [web], [graphs]) after all top-level keys, as TOML requires.

Device choices (microphone, speaker, camera, voice) and model residency are not set here: they are owner selections under Settings → AI & Voice and the Hardware assignment, stored in the installation's state.

Core¶

Key	Default	Meaning
`vault_dir`	`obsidience/vault`	The live Vault directory. Set it to match `obsidience init --vault-dir`.
`db_path`	`obsidience/state/obsidience.sqlite3`	The Harness SQLite ledger and index database.
`host`	`"127.0.0.1"`	API listen address. Keep it on loopback.
`port`	`8765`	API port. The `OBSIDIENCE_PORT` environment variable overrides it.
`git_commit`	`true`	Commit approved Article changes to the Vault's own local Git audit repository. Never touches the application repository.

Models and execution¶

Key	Default	Meaning
`llm_model`	`"obsidience-gemma"`	Default served model id from the model catalog (`obsidience/harness/models/runtime.py`). Agent and Task Articles that select `auto` resolve through it.
`task_reasoning_effort`	`"medium"`	Installation default reasoning effort for Tasks that do not set their own.
`llm_temperature`	`0.4`	Sampling temperature for model steps.
`max_steps`	`24`	Tool-loop step cap per activation.
`adk_model_client`	`"local"`	ADK model route. `"local"` streams from llama-server over the Harness's HTTP pool; `"litellm"` restores ADK's official `LiteLlm` route (same requests, more CPU).
`tick_seconds`	`20`	Longest scheduler wait in seconds; admissions, completions and foreground release also wake it.
`concurrency`	`1`	Concurrent autonomous Task executions the scheduler admits.

Retrieval¶

Key	Default	Meaning
`embed_url`	`"http://127.0.0.1:8791"`	The EmbeddingGemma 2 embedding service used by the Knowledge index.
`search_k`	`24`	Candidate count per retrieval lane before fusion.
`rrf_k`	`60`	Reciprocal-rank-fusion constant for combining the lexical and dense lanes.

Conversation¶

Key	Default	Meaning
`conversation_window`	`true`	Bound the provider's view of the Executive conversation to a window that rebases at idle edges (normally six exchanges / 4K estimated tokens, rebasing to three / 2K). `false` disables the window.
`memory_summary`	`true`	Include the Executive's cached Hindsight mental-model pages as one bounded, unverified prompt section. `false` keeps them out of the prompt while per-turn recall continues.

Voice¶

All voice keys are read by the speech worker or the Realtime runtime; restart the Harness (speech worker) to apply them.

Key	Default	Meaning
`realtime_wake_spotter`	`true`	A CPU keyword spotter hears the wake word about 0.2 s after it is spoken and plays the ready chirp before ASR text arrives. `false` keeps ASR-only wake detection.
`realtime_smart_turn`	`true`	Smart Turn v3.2 judges each pause on CPU. `false` restores the fixed 700 ms pause; a model that fails to load also falls back to it.
`realtime_smart_turn_model`	the v3.2 CPU ONNX path set in the speech worker	`"bundled"` selects Pipecat's bundled Smart Turn v3.1; otherwise a model path.
`realtime_short_replies`	`true`	Complete short replies ("yes", "no", "never mind") end the turn at the model's verdict even when scored incomplete.
`realtime_asr_silence_gate`	`true`	Skip ASR GPU steps 0.7 s after VAD speech or the last recognized word; each VAD onset replays its 320 ms pre-roll. `false` runs ASR continuously.
`realtime_asr_dtype`	`"bfloat16"`	Encoder precision for streaming ASR: `"bfloat16"` or `"float32"`.
`realtime_asr_phrases`	`obsidience/state/realtime-asr-phrases.txt`	Phrase-boosting file (one phrase per line, `#` comments; the wake word is always included). `false` disables phrase boosting.
`realtime_asr_early_final`	`false`	At the first 200 ms pause, decode the tail once more and send the words as a pause edge for speculative admission.
`realtime_semantic_turns`	`false`	Contextual speech admission: the Executive judges whether an utterance is complete in the same request that drafts its first response, speculated at pauses.
`realtime_early_speech`	`true`	Speak claim-checked sentences of a voice turn before completion acceptance. `false` speaks only the accepted reply.
`realtime_progressive_evidence`	`true`	Voice `computer.replay` returns an endpoint first while history is prepared, then a temporal overview and optional detail. Requires early speech; `false` uses one completed result.
`realtime_barge_in_ducking`	`true`	Speech over a reply ducks it, holds it once you keep talking and resumes it after a false alarm. `false` stops playback on speech onset.
`realtime_tts_quantize`	`false`	Pocket TTS uses dynamic int8 weights in its FlowLM (CPU, less memory, same voice file). Read at speech worker start.

See Realtime voice for how these fit together.

Optional provider tables¶

`[memory]`¶

Key	Default	Meaning
`hindsight_url`	unset (memory disabled)	Base URL of the local Hindsight API, for example `"http://127.0.0.1:8790"`. Leave unset until the service is configured. See Memory.

`[web]`¶

Key	Default	Meaning
`searxng_url`	unset	A local SearXNG instance for `web.search`. Without it, `web.search` uses its DDGS fallback.

`[graphs]`¶

Key	Default	Meaning
`codebase_url`	`"http://127.0.0.1:9749"`	codebase-memory-mcp's HTTP endpoint for the Code view.
`codebase_project`	the repository path with `/` replaced by `-`	The indexed project to display.

Environment variables¶

Variable	Meaning
`OBSIDIENCE_PORT`	Overrides `port`.
`OBSIDIENCE_OPTIMIZER_BACKEND`	`harness2` (default) or `autosaddler`, the retained rollback search backend for `harness.optimize`.
`OBSIDIENCE_MEDIA_CONFIG`	Path of the private Media configuration (default `obsidience/state/media.json`).

Example¶

llm_model = "obsidience-gemma"
git_commit = true
realtime_semantic_turns = true
realtime_asr_early_final = true

[memory]
hindsight_url = "http://127.0.0.1:8790"

[web]
searxng_url = "http://127.0.0.1:8888"