Skip to content

Latest commit

 

History

History
108 lines (77 loc) · 2.54 KB

File metadata and controls

108 lines (77 loc) · 2.54 KB

Production Guide

Morning Signal is a reference implementation, not a complete production system. Use this checklist when hardening the pattern for real users.

Storage

The local app uses SQLite for simple persistence. On serverless platforms, use a durable managed store instead:

  • Postgres for generated briefs, users, and runs
  • Redis or a runtime cache for short-lived source and profile caching
  • object storage for raw evidence packets if they become large

In demo mode, the hosted version skips SQLite writes so it can run safely on serverless infrastructure.

Long-Running Work

Live web workflows can be slow or bursty. For production:

  • move long crawls into background jobs
  • persist run state
  • expose run IDs
  • support resume/retry
  • stream status events to the client
  • store partial results when a source fails

Reliability

Add source-level resilience:

  • retry transient network/provider errors
  • use exponential backoff
  • dedupe URLs before extraction
  • continue when non-critical extracts fail
  • record failed URLs and reasons
  • cap per-run searches and extracts

Cost Control

Every run should have an explicit retrieval budget.

Track:

  • number of searches
  • number of extracted URLs
  • selected driver mode
  • model tokens
  • total estimated cost
  • cache hit rate

Start with cheaper access modes and escalate only when the page requires it.

Source Quality

Add scoring before synthesis:

  • relevance to approved context
  • freshness
  • source authority
  • content depth
  • duplicate detection
  • source diversity
  • whether extract succeeded

Low-confidence sources should be excluded or labeled clearly.

Grounding And Citations

The model should only make claims supported by the evidence packet.

Recommended checks:

  • require source IDs in structured output
  • reject sections with missing citations
  • validate cited URLs exist in the evidence packet
  • flag unsupported recommendations
  • keep raw evidence available for audit

Security

For multi-user deployments:

  • add authentication
  • isolate tenants
  • encrypt secrets
  • never expose provider keys to the browser
  • rate-limit API routes
  • validate public URLs
  • reject local/private network targets
  • sanitize rendered content

Observability

Useful traces include:

  • approved context
  • planned queries
  • Nimble Search requests and result counts
  • extraction success/failure
  • source scoring decisions
  • synthesis prompt version
  • output validation status
  • latency and cost per run

LangSmith or another tracing system is useful because it shows the agent as a workflow rather than a hidden prompt.