Langfuse

What is HardGraph? HardGraph publishes curated, provenance-backed agent skills grounded in reproducible vendor documentation.

Langfuse is an open-source LLM engineering platform built around one data model — the trace — that three largely independent products consume: tracing/observability, evaluation, and prompt management. Treat them as separate concerns even when reaching for all three at once: a trace is captured regardless of whether anything evaluates or prompts against it, an evaluation can run against traces captured weeks earlier by a different SDK, and prompt management works even for an application that sends Langfuse no traces at all.

The decision that shapes everything else

Cloud versus self-hosted is a data-residency decision, not a convenience one. Langfuse Cloud means every trace — prompts, completions, and any metadata or user identifiers you attach — leaves your infrastructure and lands on Langfuse's servers, subject to their region and retention settings. Self-hosting keeps that data in infrastructure you control, but it is not a single container: a production deployment coordinates a relational store, an analytical store for trace data, a cache, and blob storage, each independently upgraded and scaled. Decide this before instrumenting anything — traces don't migrate between deployment modes for free, and masking/PII-scrubbing decisions are cheaper to make before the first trace ships than after.

Easily confused, worth getting right

• Tracing is not evaluation. Tracing answers "what happened"; evaluation (LLM-as-a-judge, human annotation, code-based scorers) answers "was it good," and runs as a separate pass — sometimes online against live traffic, sometimes offline against a fixed dataset. Conflating them leads to instrumenting for observability and then being surprised evaluation needs its own scorer configuration.
• Prompt management is not prompt engineering inside your codebase. Prompts pulled from Langfuse are versioned and can be updated without a deploy — which also means an application can silently start running a different prompt than the one committed in its repo.
• The ingestion protocol has moved across major versions, including a shift toward OpenTelemetry as the transport. An SDK and a server instance from mismatched eras can fail to talk to each other in ways that look like a networking bug.

What to verify rather than recall

Do not assert a remembered SDK version, self-hosted infrastructure requirement, or which features require an enterprise license key — these have changed across major versions and will again. Verify current guidance, SDK migration paths, and self-hosted component requirements against Langfuse's own documentation before committing to an approach.