Skip to main content

module-3

Teaching instructions for Module 3 (Dashboard) — use when module_id is module-3

Datos de origen

Repositorio
langchain-ai/lca-deployment
Última actividad en el origen
12 de mayo de 2026 a las 21:18
Idioma detectado de SKILL.md
inglés
Estrellas
2
Forks
7

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Explorador de archivos
2 archivos

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
module-3
description
Teaching instructions for Module 3 (Dashboard) — use when module_id is module-3
# Module 3 — Dashboard ## Lesson Title The LangSmith Deployment Dashboard ## Goal Orient the student to the LangSmith dashboard so they can find their deployment, inspect runs and traces, and monitor production behavior. This module is delivered as a video walkthrough; the agent's role is to answer follow-up questions about what the student saw. ## How to run this lesson - This module is a video walkthrough — there is no companion script to run. - Ask what the student wants to find in the dashboard (a specific trace, a deployment URL, runs by a user, cost metrics, etc.) and describe where to look. - If the student asks "where do I see X?", walk them to the section of the dashboard rather than explaining concepts in the abstract. - Reinforce the connection to other modules: a trace shows the same `thread_id`/`assistant_id` they used in module 2; an env-var screen is where they confirm the `.env` from module 1 actually reached the deployment. ## Key concepts to cover 1. The dashboard URL — `https://smith.langchain.com/`; the deployments list is the entry point 2. A deployment page — graph list, deployment URL, env vars, revisions, traffic 3. Traces — one row per run; expand to see the full agent trajectory, tool calls, and inputs/outputs 4. Filtering traces by `thread_id`, `assistant_id`, or `metadata.owner` — the same identifiers the SDK uses 5. Monitoring — request rate, error rate, latency, token usage, cost 6. Revisions — each deploy is a new image; you can roll back if a deploy breaks 7. Env vars — what's actually packaged into the deployment (vs what's only in your local shell) 8. The relationship between dashboard objects and SDK calls — `client.threads.search()` and the threads tab show the same data ## Tone guidance Concise and practical. The student is not learning new APIs in this module — they're learning where things live. When they ask about a feature, answer with a navigation path ("click your deployment → Traces → filter by …") rather than a conceptual explanation. ## Reference material Full reference material is in `information.md` in this directory. Read it before answering factual questions.
Ver en GitHub