Skip to main content
Draft Thinker ships a pre-built Grafana dashboard that covers all Prometheus metrics exposed by the gateway. The dashboard and its Prometheus datasource are provisioned automatically when you start the stack with Docker Compose — no manual import or configuration is required.

​
Accessing the dashboard

1

Start the stack

Run Docker Compose from the repository root. This starts the gateway, Redis, Qdrant, Prometheus, and Grafana together.
docker compose up
2

Open Grafana

Navigate to http://localhost:3000 in your browser.
3

Sign in

Log in with the default credentials:
FieldValue
Usernameadmin
Passwordadmin
4

Open the dashboard

The Draft-Thinker Gateway dashboard is pre-provisioned and available immediately under Dashboards in the left sidebar. The dashboard auto-refreshes every 10 seconds and defaults to a 15-minute time window.
The default admin password is set via GF_SECURITY_ADMIN_PASSWORD=admin in docker-compose.yml. Change this value before deploying to any shared or production environment.

​
How provisioning works

Grafana is configured entirely through volume mounts defined in docker-compose.yml. No manual setup steps are needed.
docker-compose.yml (grafana service)
grafana:
  image: grafana/grafana:11.5.2
  ports:
    - "3000:3000"
  environment:
    - GF_SECURITY_ADMIN_PASSWORD=admin
  volumes:
    - ./grafana/provisioning:/etc/grafana/provisioning:ro
    - ./grafana/dashboards:/var/lib/grafana/dashboards:ro
  depends_on:
    - prometheus
Two provisioning directories are mounted:
  • grafana/provisioning/datasources/ — Configures the Prometheus datasource pointing at http://prometheus:9090, set as the default datasource.
  • grafana/provisioning/dashboards/ — Tells Grafana to load dashboard JSON files from /var/lib/grafana/dashboards, which maps to grafana/dashboards/ in the repository.
The dashboard JSON lives at grafana/dashboards/draft-thinker.json. To add or modify panels, edit that file directly and restart the Grafana container, or use the Grafana UI (the dashboard is marked editable: true).

​
Prometheus scrape configuration

Prometheus is configured to scrape the gateway’s /metrics endpoint every 15 seconds.
prometheus.yml
global:
  scrape_interval: 15s

scrape_configs:
  - job_name: draft-thinker
    static_configs:
      - targets:
          - gateway:8080
    metrics_path: /metrics
The target gateway:8080 resolves to the gateway container over the Docker Compose internal network.

​
Dashboard panels

The dashboard is organized into four row sections. Each row groups related panels.

Overview

Four summary stats shown at the top of the dashboard:
  • Request rate — sum(rate(draftthinker_requests_total[5m])) in requests per second
  • Draft acceptance rate — fraction of routing decisions where the drafter’s response was accepted, shown as a gauge with green/yellow/red thresholds
  • Cost reduction — calibrated value (91.6%) at threshold T=2.0 on 518 prompts
  • Cache hit rate — fraction of requests served from semantic cache, shown as a gauge

Latency

Three time-series panels showing P50, P95, and P99 percentiles:
  • Upstream latency by provider — draftthinker_upstream_latency_seconds broken out by drafter and heavyweight
  • Cache lookup latency — draftthinker_cache_lookup_latency_seconds end-to-end including embedding and vector search
  • Speculative latency saved — draftthinker_speculative_latency_saved_seconds P50 and P95

Routing

Two time-series panels showing decision flow and errors:
  • Routing decisions over time — draftthinker_routing_decisions_total stacked by accept, escalate, and cache_hit
  • Error rate by type — draftthinker_errors_total broken out by type label (invalid_request, routing_error, upstream_error, upstream_timeout, stream_error, internal_error)

Entropy and speculative

Three panels covering the routing engine internals:
  • Entropy distribution — draftthinker_entropy_distribution rendered as a heatmap over time, showing the per-token Shannon entropy spread in bits
  • Speculative trigger rate — rate(draftthinker_speculative_triggers_total[5m]) in triggers per second
  • Speculative cancellation ratio — draftthinker_speculative_cancellations_total / draftthinker_speculative_triggers_total as a gauge, green below 10%, yellow to 30%, red above

Build docs developers (and LLMs) love

Get started for freeTalk to us