LLM tracing

Trace LLM calls from prompt to final answer

Currai turns each LLM request into a trace with the prompt, model response, tool calls, retrieval steps, token usage, cost, latency, and metadata you need to debug it.

↗Start tracing Read the docs

Primary keyword

LLM tracing

Currai covers

Traces, generations, spans, evals, prompt A/B tests, token usage, cost, latency, sessions, users, and OpenTelemetry ingestion.

Replace print debugging with structured traces

Logs are useful, but they rarely preserve the full context of a model call. Currai traces the request tree so you can see what happened before and after each generation.

Trace a single chat completion, a multi-step agent run, or a RAG pipeline with nested spans for retrieval and tool execution.

Capture input, output, model parameters, usage, and status.
Nest spans for retrievers, tools, chains, and custom functions.
Group traces by session and user for multi-turn debugging.

Use SDKs or OpenTelemetry

Start with Currai's Python or TypeScript SDKs, migrate existing Langfuse instrumentation by changing the host, or send OTLP traces directly from any OpenTelemetry-compatible service.

Questions about LLM tracing

What should I trace in an LLM app?

Trace model calls, prompts, completions, retrieval, tools, retries, token usage, cost, latency, user/session identifiers, and quality signals.

Does tracing add latency?

Currai batches and flushes trace data in the background. In short-lived jobs, call flush before exit so queued traces are sent.

↗Sign In