Observability

import synapsekit.observe enables SynapseKit's tracing hooks for LLMs, RAG pipelines, agents, and graphs.

Install

pip install synapsekit[observe,openai]

Quickstart

import synapsekit.observe as observe
from synapsekit import RAG

observe.configure(
    exporter="jaeger",
    endpoint="http://localhost:4317",
    service_name="my-rag-app",
    trace_llm_inputs=True,
    trace_llm_outputs=True,
    cost_tracking=True,
    sample_rate=1.0,
)

rag = RAG(model="gpt-4o-mini", api_key="sk-...")
answer = await rag.ask("What is the main topic?")

Export to Grafana in 5 minutes

1. Start Prometheus + Grafana (see the Helm chart below or your own stack).
2. Enable metrics and expose a /metrics endpoint:

from synapsekit.observability import PrometheusMetrics
import synapsekit.observe as observe

metrics = PrometheusMetrics(start_server=True, port=8000)
observe.configure(
    exporter="otlp",
    endpoint="http://localhost:4317",
    metrics=metrics,
)

3. In Grafana, add Prometheus as a data source.
4. Import assets/grafana/synapsekit-observe-dashboard.json.

Prometheus metrics

SynapseKit emits the following metrics when enabled:

• synapsekit_cost_usd_total (counter)
• synapsekit_tokens_total (counter)
• synapsekit_latency_seconds (histogram)

Exporters

Supported exporter names:

• console
• otlp
• jaeger
• langfuse
• honeycomb

You can also pass a custom exporter object with export() and clear() methods.

Privacy controls

observe.configure(
    trace_llm_inputs=False,
    trace_llm_outputs=False,
    redact_keys=["api_key", "password"],
)

Jaeger docker-compose snippet

services:
  jaeger:
    image: jaegertracing/all-in-one:1.57
    ports:
      - "16686:16686"
      - "4317:4317"

Grafana dashboard

Import assets/grafana/synapsekit-observe-dashboard.json into Grafana as a starting point for tracing dashboards.