Skip to content

Pipeline configuration

Pipeline settings live under memory.pipelineV2 in agent.yaml. Inference target choice belongs in Inference and routing. Do not add the retired memory.synthesis block: the current loader rejects it.

Restart the daemon after a pipeline configuration change. Long-running workers are started with a configuration snapshot, so editing YAML alone is not a reliable apply operation.

Terminal window
signet daemon restart
signet daemon status --json
memory:
pipelineV2:
telemetryEnabled: true
autonomous:
enabled: true
frozen: false
maintenanceIntervalMs: 1800000
maintenanceMode: execute
documents:
workerIntervalMs: 10000
chunkSize: 2000
chunkOverlap: 200
maxContentBytes: 10485760
continuity:
enabled: true
promptInterval: 10
timeIntervalMs: 900000
retentionDays: 7

Omit settings you do not need. The daemon supplies bounded defaults and clamps accepted numeric values while loading configuration.

telemetryEnabled defaults to true. Set it to false to opt out persistently; SIGNET_TELEMETRY_OPTOUT=1 disables telemetry for the process without changing YAML. See Analytics for the local audit path and privacy boundary.

autonomous.enabled controls autonomous maintenance. autonomous.frozen pauses autonomous writes without deleting configuration. autonomous.maintenanceMode is observe or execute. Use observe while evaluating a new deployment; use the repair endpoints only with an authenticated operator or admin principal where applicable.

The repair object bounds maintenance work with cooldowns and hourly budgets. Its current keys include re-embed, requeue, and deduplication cooldown and budget settings. Keep those limits conservative when a remote provider is involved.

documents controls the daemon document worker: poll interval, chunk target, overlap, and maximum accepted content bytes. These are processing limits, not a replacement for upload policy. See Sources for ingest behavior.

continuity controls checkpoint cadence, retention, and recovery-budget limits. It is independent of ordinary recall. Tune it only when you have a concrete recovery or storage requirement.

embeddingTracker controls the bounded background pass that detects missing or stale embeddings. guardrails bounds stored and injected text. subagents controls parent-context inheritance for supported harness flows. Leave defaults in place unless a measured issue requires a change.

worker.maxLlmConcurrency sets a shared cap for active LLM work. SIGNET_MAX_LLM_CONCURRENCY overrides it for the process when it is a positive integer. Do not set this from a guess: provider quota, local GPU memory, and latency are deployment-specific.

For canonical model selection, create or edit a routing target and bind the workload. See Inference and routing. A legacy provider/model field in an old workspace is not a safe substitute for a workload binding.

Terminal window
signet daemon status --json
curl -fsS http://127.0.0.1:3850/api/pipeline/status
curl -fsS http://127.0.0.1:3850/health/ready

Use Diagnostics when a queue, embedding, or maintenance problem persists. Do not infer worker health from a successful configuration write.