Settings — GratiaMesh v2

Runtime safety

Global agent-run kill switch and Bifrost failure protection.

Agent executions

Bifrost circuit breaker Open after repeated upstream 5xx/timeouts; one probe is admitted after the recovery window.

Failures Recovery seconds

General

Identity, branding and the dashboard summary window.

SSO providers live on Manage → Identity, not in this form. Incident kill switch: GRATIA_OIDC_ENABLED.

App display name Shown in the sidebar, login and browser title.

Log level DEBUG records verbose traces; INFO hides them.

Dashboard window Hours averaged for the home statistics.

Branding avatar Shown on login, sidebar and footer.

Upload

Accent color Chat button, highlights and accents on the end-user UI (hex #RRGGBB).

Corner radius Global shape for the end-user chat components.

Show "Powered by GratiaMesh" Display the attribution footer on the end-user chat.

Rate limit

Guard the API against bursts.

Enable rate limiting Applies globally to API requests.

Requests per minute Per-IP allowance once enabled.

Eval hard gate Blocks agent Save and prompt-improve Accept when the linked suite fails thresholds. Accept can override with force (audited).

Eval max cost per run (USD) Optional global ceiling for eval runs (instance-wide). Suite budgets are clamped to this cap.

Eval max concurrency Optional global cap on parallel eval cases (1–10).

Concurrency

Parallelism caps for shared resources.

Max concurrent LLM calls Upstream requests sent to Bifrost at once.

Max concurrent jobs Background workers running in parallel.

Delete jobs after (days) Background jobs older than this are removed automatically, regardless of status.

Delete chats after (days) 0 keeps chat sessions and their objects. Otherwise they are purged hourly.

Delete usage after (days) 0 keeps the usage log. A positive value deletes older token rows.

Delete audit events after (days) 0 keeps the audit trail. Confirm before shortening a window that is already in force.

Delete config exports after (days) 0 keeps export bundles. A positive value also deletes the stored object.

Autoscaling

Internal horizontal scaler for worker (queue) and sandboxes (tool invocations). Disable cluster HPA/KEDA when this is on.

Enable autoscaler Global kill-switch. When off, no reconcile jobs run.

Dry run Record decisions without changing Deployment replicas.

Worker min replicas Never scale the worker below this (phase 1 floor is 1).

Worker max replicas Hard cap for queue-driven scale-up.

Target jobs per replica Desired pending jobs each worker replica should absorb.

Sandbox defaults enabled When on, running sandboxes inherit autoscaling unless they opt out. Prefer per-sandbox enable in Deploy.

Sandbox default max replicas Default cap for MCP sandbox Deployments (override per sandbox).

API autoscaling Scale the API Deployment from HTTP p95/RPS/5xx. Disable Helm HPA when on.

API max replicas Hard cap for API scale-up (min is always ≥ 1).

Starts in dry-run. Turn dry-run off only after reviewing autoscaling_events and confirming HPA/KEDA are disabled for the worker.

Document extraction

How files are parsed into text and chunks.

Extraction engine Apache Tika runs as a sidecar; internal mode parses natively with OCR.

OCR on images Vision models transcribe scanned pages and images.

Max file size Bytes; 0 means the engine default.

Graph entity extraction LLM Default chat model for knowledge-graph entity extraction. Empty = fallback deepseek-v4-flash.

Chunking

Defaults used when a collection has no explicit value.

Chunk size Characters per retrieved text chunk.

Chunk overlap Shared characters between neighbor chunks.

HNSW m Graph degree used the next time the partial index is rebuilt (2–100).

HNSW ef_construction Build-time candidate list (4–400). Higher values cost more index time.

HNSW ef_search Search-time candidate list (1–1000). Queries raise it to at least twice k.

Rebuild HNSW index Enqueues a background job. Save the parameters first so the job reads them.

Secret seal How provider keys and connector secrets are sealed at rest.

Agent memory (ai-memory)

Sidecar MCP for shared agent memories. Disable to stop attaching memory tools (fail-closed).

Enable shared memory When off, agents do not receive memory_* MCP tools at runtime.

Backend URL Base URL of ai-memory (without /mcp). Loopback or internal DNS only.

Provision mode http_admin (Docker Compose / k8s) calls ai-memory /admin/*. docker_cli uses docker exec on the host (legacy VPS).

Docker container Only for docker_cli provision mode.

Root token

Knowledge sync

Limits for S3/Drive/Git/Web ingestion jobs.

Max objects / files Per sync run (S3, Drive, Git…).

Enqueue batch Files queued per job slice.

Max pending ingests Pause new syncs above this queue length.

Default token quotas

Fallback limits when a user, agent, or LLM has no explicit quota.

Soft-warn at % of quota FinOps: email when usage reaches this % (run continues). Hard block at 100% does not send email. Requires SMTP under Settings → SMTP.

Soft-warn recipients (roles) Who gets FinOps soft-warn mail. Comma-separated RBAC roles (e.g. admin). Eval suite recipients are configured on each suite in Operate.

Soft-warn recipients (emails) Optional extra addresses for FinOps soft-warn, comma-separated.

Fallback limits applied when a user, agent, or LLM has no explicit quota. Per-entity quotas still override these defaults.

SMTP

Instance email delivery channel. Alert rules live on Operate resources (eval suites) and under Token quotas (FinOps soft-warn).

SMTP is not configured. Product alerts will be recorded as skipped until host and from are set.

SMTP host e.g. smtp.example.com

Port

STARTTLS

From address

Username Optional; leave empty for open relay / IP allowlist.

Password

Test channel Sends a test email using current saved settings (save first if you changed host).

Recent deliveries Last emails the channel attempted (any product alert). Full inbox will live under Operate → Alerts later.

Observability

Bundled Prometheus + Grafana (opt-in). Admin password comes from Helm values; SSO multi-user is deferred.

Observability is off or GRATIA_GRAFANA_PUBLIC_URL is empty. Enable with make observability-up and set observability.grafana.adminPassword in Helm values.