Source: LangSmith observability documentation

LangSmith's trace data model (runs, traces, threads, feedback), run schema, OTLP ingestion, sampling, retention and plan pricing.

Contents

Summary

LangSmith is LangChain’s hosted tracing and evaluation platform. Its unit is the run (comparable to an OTel span); runs sharing a trace ID form a trace; traces sharing a thread_id metadata key form a thread. It ingests both its own SDK format and OTLP, mapping GenAI semantic-convention attributes onto runs.

Key claims

  • Run types: chain, llm, embedding, prompt, tool, retriever, parser. Runs carry dotted_order (sortable hierarchy path), inputs/outputs, tags, extra, error, token counts, total_cost, and first_token_time for streamed LLM runs.
  • A trace is capped at 25,000 runs; further runs are rejected.
  • Feedback (a key plus a continuous or categorical score) attaches to a run by run ID; tags and metadata are for filtering and grouping.
  • Newer “agent-based” workspaces organise traces as agents with Production/Staging/Development/Local environments, each backed by a tracing project.
  • OTLP endpoint is the API URL plus /otel (exporters append /v1/traces); auth via x-api-key. gen_ai.operation.name sets run type; gen_ai.request.* maps to invocation params; usage incl. cache and reasoning tokens maps to usage_metadata. OpenInference, Traceloop and Logfire attributes are also understood.
  • OTLP pitfalls: a span whose parent never arrives is dropped silently (endpoint returns 200 before processing); 8-byte OTLP span IDs cannot point at UUID run IDs.
  • Sampling: LANGSMITH_TRACING_SAMPLING_RATE (0 to 1, default all) or a per-client tracing_sampling_rate; applies to SDK/LangChain tracing.
  • SaaS keeps trace data at most 180 days from ingestion; datasets persist. Pricing page (2026-11): base traces have 14-day retention and extended traces (an upgrade at extra cost) 180-day; Developer plan free with 5k base traces/month and 1 seat, Plus $39/seat/month with 10k base traces/month and unlimited seats; self-hosting is Enterprise-only.

Feeds

LLM observability

Suggest changes →