Langfuse, LangSmith, OpenTelemetry, and Sentry

Overview

• Langfuse and LangSmith provide AI-oriented trace and experiment views.
• OpenTelemetry standardizes telemetry instrumentation and export across services.
• Sentry provides application error and performance investigation capabilities; exact AI-specific integrations vary by SDK/version.

Why this matters

AI traces should connect to the wider web/backend system, not form a separate unexplained island.

An everyday example

Specialist case notes connect to the company's general incident reporting.

A technical example

A Node request span links to a Python extraction span and an AI trace, while Sentry captures a parser exception.

How it works

1. Instrument service boundaries
2. Propagate context
3. Export permitted metadata
4. Group errors
5. Associate model/prompt versions
6. Inspect cross-service failures.

Architecture and flow

App + AI instrumentation → collectors/platforms → trace/error dashboards.

The insurance assistant example

Follow a failed upload from browser to Node, queue, Python OCR worker, and storage service.

When to use it

Choose AI-specific views where useful and consistent distributed tracing for service boundaries.

When to avoid it

Avoid sending duplicate sensitive payloads to several platforms without need.

Interview answer

“I combine AI-level traces with normal distributed tracing and error monitoring so model behavior and service reliability can be investigated together.”

Benefits, trade-offs, and production considerations

• Telemetry needs redaction, retention, access policy, and exportability.
• OpenTelemetry traces explains spans and trace context; AI conventions and vendor integrations should be checked for the adopted version.