Skip to content
Observability Stack

SDKs

The OpenSearch AI Observability SDKs instrument LLM applications using standard OpenTelemetry. They handle the gap that general-purpose OTel doesn’t cover: tracing your own agent logic — the workflows, agents, and tools that sit above raw LLM calls — and submitting evaluation scores back through the same pipeline.

The SDKs are thin wrappers. They do not replace OpenTelemetry, they configure it. Remove a decorator and your code still works unchanged.

What the SDK covers

Section titled “What the SDK covers”

Pipeline setup — one call (register()) creates a TracerProvider, wires up an OTLP exporter, and activates auto-instrumentation for any installed LLM library instrumentors (OpenAI, Anthropic, Bedrock, LangChain, and more).

Application tracing — decorators (Python) or wrapper functions (JavaScript) that produce OTEL spans with GenAI semantic convention attributes for four span types:

TypeUse for
workflowTop-level orchestration — the entry point of a pipeline run
taskA discrete unit of work inside a workflow
agentAutonomous decision-making logic that calls tools or LLMs
toolA function invoked by an agent

Evaluation scoring — score() emits evaluation metrics as OTEL spans at span, trace, or session level. No separate client or index needed — scores travel through the same OTLP pipeline as traces.

AWS support — built-in SigV4 signing for OpenSearch Ingestion (OSIS) and OpenSearch Service endpoints.

Architecture

Section titled “Architecture”
flowchart LR
    A["Your Application<br/>@workflow / @agent / @tool<br/>score()"] -->|"OTLP HTTP/gRPC"| B["OTel Collector<br/>or Data Prepper"]
    B --> C["OpenSearch<br/>traces + scores"]

The SDK configures a BatchSpanProcessor that exports in the background — your application code is never blocked waiting on network I/O.

Available SDKs

Section titled “Available SDKs”
Section titled “Related links”