ELUVIXSovereign AI Router

Observability & Telemetry

Eluvix features comprehensive, enterprise-grade observability across every model invocation, agent tool call, and routing decision. Telemetry is available through both the visual canvas configuration and the real-time Observability & Activity Monitor (/observability).

Observability Canvas Node (observabilityNode)

In the Visual Routing Editor, the Observability node configures pipeline-wide telemetry collection and remote export pipelines:

Telemetry Exporter Options

  • Embedded SQLite (sqlite) (Default): Zero-overhead local persistence in ~/.local/share/eluvix/eluvix.db. Provides instant dashboard analytics without deploying external monitoring infrastructure.
  • OpenTelemetry Collector (otlp): Emits standardized OpenTelemetry traces and spans over OTLP/gRPC to collectors such as Jaeger, Datadog, or Grafana Tempo.
  • Langfuse Platform (langfuse): Streams detailed LLM execution graphs, token metrics, generation parameters, and tool call traces to self-hosted or cloud Langfuse instances.
  • Prometheus (prometheus): Exposes a /metrics scrape endpoint for Prometheus and Grafana dashboards.

Trace Sampling Rate

Configurable slider (1% to 100%) to control telemetry volume on high-throughput enterprise gateways, preventing trace storage exhaustion while maintaining statistical significance.

Privacy & PII Scrubbing

When enabled (scrubPii: true), telemetry recorders pass all inputs and outputs through the Privacy Guard before persisting traces, ensuring that sensitive user data is never stored in external log aggregators or APM systems.

Full I/O Capture

Toggle between lightweight metadata logging (latency, tokens, status codes) and complete payload capture (raw prompts, responses, tool arguments) for forensic debugging.

SLA Latency Alert Threshold

Defines a critical latency boundary (default: 5,000ms). Requests taking longer than this threshold trigger visual SLA warning badges on the dashboard and in exported trace spans.

Live Activity Monitor (/observability)

The web dashboard provides real-time operational intelligence with sub-second responsiveness:

| Dashboard Component | Description | | :--- | :--- | | Total Requests & Success Rate | Cumulative request volume with breakdown of HTTP 200 OK vs 4xx/5xx error states. | | Latency Distribution Tiers | Latency tracked across models and upstreams, classified into speed tiers (< 500ms Ultra-Fast, 500-2000ms Standard, > 2000ms Deep Thinking). | | Token Throughput | Total tokens processed, segmented into prompt tokens and completion tokens. | | Local Cost Savings | Running calculation of financial savings achieved by routing queries to local $0 hardware instead of commercial frontier APIs. | | Upstream Health Matrix | Real-time ping status for Lemonade, Ollama, vLLM, Antigravity AGY, and Claude Code. |

Interactive Audit Feed & Tooling

  • Live Streaming Polling (● LIVE 3s): Auto-refreshes every 3 seconds with a pause toggle for inspecting specific traces.
  • Audit Feed Filtering: Filter logs by status (All, 2xx Success, 4xx/5xx Errors) or search across model names and upstream endpoints.
  • Log Inspector Modal: Click 🔍 on any request row to review complete headers, latency timings, and token metrics.
  • JSON Export: Export recent request histories as structured JSON files for offline compliance reviews.
  • Send Test Ping: Dispatches a lightweight synthetic completion through the active pipeline to verify end-to-end health.
  • Purge Logs: Clear historical logs from request_logs to reset benchmarks.