Skip to content

METRICS.md

Latest commit

 

History

History
20 lines (16 loc) · 1.2 KB

METRICS.md

File metadata and controls

20 lines (16 loc) · 1.2 KB

Telemetry & Metrics Output

Dashboard Philosophy

ContextFlow operates exclusively as a headless middleware Python library. There are no integrated UI dashboards or web admin pages included in this repository.

The MetricsCollector

For agent runners to confidently use this library, measuring efficiency is mandatory. Through dependency injection, developers can provide a MetricsCollector (metrics.py) instance to track deterministic variables during the pipeline sequence.

Target Trace Metrics

  • tokens_before: Total baseline token count from raw sources.
  • tokens_after: The final token count mapped directly to the provider.
  • compression_ratio: Representing the literal percentage of fat trimmed (tokens_after / tokens_before).
  • latency_ms: How long the Mode + Compressor sequential execution takes. (Goal: <50ms standard)
  • cost_savings_estimate: A computed translation of tokens saved vs. the Provider's standard $/1M token pricing bracket.

Integration

Because the collector emits a standard dictionary state, teams are encouraged to sink this telemetry into backend enterprise platforms:

  • LangSmith / Datadog
  • Prometheus / Grafana stacks
  • ElasticSearch logs