AGENT SKILL

databricks-docs

Look up authoritative, up-to-date Databricks documentation covering the core platform, Unity Catalog data governance and security, data engineering (Lakeflow Jobs and Declarative Pipelines), SQL and analytics, machine learning and AI (agents, MLflow, model serving), developer tools (CLI, SDKs, asset bundles), compute, administration, integrations, migration, and troubleshooting. Use when the user asks how a Databricks feature works, what a config field or SQL clause does, how to set up jobs, pipelines, clusters, catalogs, or agents, when troubleshooting a Databricks error, or when you need current official docs rather than training-cutoff knowledge. Also trigger on Chinese phrasing such as 数砖 / Databricks 文档 / 官方文档 / 怎么配置 / 如何设置 / 报错 / 数据血缘 / 权限 / 作业调度 / 流水线.

License 许可
MIT
最近更新
2026-08-04
Stars
★ 0
Forks
0

01 前言 Front Matter

name
databricks-docs
description
Look up authoritative, up-to-date Databricks documentation covering the core platform, Unity Catalog data governance and security, data engineering (Lakeflow Jobs and Declarative Pipelines), SQL and analytics, machine learning and AI (agents, MLflow, model serving), developer tools (CLI, SDKs, asset bundles), compute, administration, integrations, migration, and troubleshooting. Use when the user asks how a Databricks feature works, what a config field or SQL clause does, how to set up jobs, pipelines, clusters, catalogs, or agents, when troubleshooting a Databricks error, or when you need current official docs rather than training-cutoff knowledge. Also trigger on Chinese phrasing such as 数砖 / Databricks 文档 / 官方文档 / 怎么配置 / 如何设置 / 报错 / 数据血缘 / 权限 / 作业调度 / 流水线.
allowed-tools
Bash(python3 *), WebFetch
argument-hint
[ "topic" ]

02 资源文件 9

03 SKILL.md 正文

Databricks Docs

Answers Databricks questions from the official docs on demand: searches a cached copy of the 47 KB llms.txt index by section or regex, then reads the matching pages with WebFetch. Always prefer this skill over recalling docs from memory — Databricks renames and reorganizes fast (dlt/ became ldp/, "Delta Live Tables" became "Lakeflow Declarative Pipelines", "Workflows" became "Lakeflow Jobs").

If the user passed an argument ($ARGUMENTS), treat it as the topic. Otherwise infer it.

When to use this skill

The index's own 15 sections, largest first:

  • Machine learning and AI — agents, MLflow, model serving, feature store, Genie
  • Core platform — Unity Catalog, catalogs, schemas, volumes, workspace, compute, notebooks
  • Developer tools — CLI, SDKs, asset bundles, VS Code extension, REST API, CI/CD
  • Data governance and security — permissions, row filters and column masks, lineage, audit
  • Data engineering — Lakeflow Jobs, Lakeflow Declarative Pipelines, ingestion, Auto Loader
  • SQL and analytics — SQL warehouses, queries, dashboards, alerts
  • Data sources and formats — Delta Lake, tables, external data, connectors
  • Overview and getting started, Administration, Reference and language-specific guides, Integrations and connectors, Migration and best practices, Specialized features, Additional resources, Troubleshooting and support

Out of scope:

  • Azure Databricks. docs.databricks.com/azure/… does not exist; Microsoft hosts those docs on learn.microsoft.com. This index covers AWS (default) and GCP.
  • DevHub (developers.databricks.com) — a separate site with its own 19 KB llms.txt, not covered here.
  • Your workspace's actual data. Reading tables, running SQL, or inspecting Unity Catalog objects needs the Databricks CLI or a Databricks MCP server, not this skill.

How this site works

Measured 2026-07-30; the numbers and the reasoning are in the top entry of references/mechanism.md.

  • Index: https://docs.databricks.com/llms.txt — 47,150 B (~11,787 tokens), 252 entries, 15 sections, 98% with prose descriptions.
  • Coverage: 252 index entries vs 5,645 sitemap URLs (4.5%) — hub-level. Many entries point at an area landing page, not a leaf. Expect a second hop for specific topics.
  • Content: HTML only. Every plain-text convention 404s (.md, /index.md, .txt, and an Accept: text/markdown header all fail). Pages measure 21 KB–51 KB of raw HTML — use WebFetch, which converts to markdown first. Never curl a page body.
  • Gotchas:
    • Index URLs are cloud/locale-neutral (/jobs/scheduled) and 301 to /aws/en/…. Follow the redirect; cite the URL that actually served the content.
    • www.databricks.com/llms.txt is the marketing index (36 entries, mostly product pages). It is not this index. Its "Databricks-owned LLM manifests" section is what points here.
    • The docs host publishes no llms-full.txt. The marketing site links one — never fetch it.
    • Docs exist in English, Japanese, and Portuguese only. There is no Chinese edition.

Procedure

1. Find candidate pages

python3 ${CLAUDE_SKILL_DIR}/scripts/docs_query.py search '<term>|<synonym>|<the word the docs would use>'

The index is fetched once and cached for 24 h, so repeat queries cost no requests. Only matching lines enter context — never load the whole index.

Routing commands, cheapest first:

Command Measured cost Use when
search '<regex>' ~60–1,300 tok always start here
sections ~253 tok search missed and you need the map
section '<name>' 222–1,617 tok full recall inside one area

Databricks renames things, so search the current and the former name as one alternation — 'Delta Live Tables|declarative pipeline|ldp', 'Workflows|Lakeflow Jobs'.

2. If the search comes back empty

Do not conclude the topic is undocumented. Escalate:

  1. Widen with synonyms — the docs' word is often not the user's. Measured on this site: cron returns 0 matches; 'schedul|orchestrat|trigger|recurring' finds Job scheduling. row-level security returns 0; 'row filter|column mask|fine-grained' finds Row and column filters.
  2. If the query was not in English, retry with English terms. This index is English-only, and there is no Chinese edition of the docs at all — 数据血缘 scores 0 matches, lineage finds the page. A non-English miss says nothing about coverage.
  3. sections, then section '<most plausible>' for full recall inside it.
  4. Descend a hub entry. The index covers 4.5% of the site, so the leaf you want often is not listed at all — only its area landing page is. WebFetch the landing page and read its own child links. Measured: vacuum|retention returns 0 index matches, but the /delta/ landing page links straight to /tables/operations/vacuum.
  5. Only then say it is not in the docs, and state what you searched.

3. Read the pages

WebFetch url=<url-from-search>
        prompt="<the user's actual question, not 'summarize this page'>"

Pages are HTML-only (21 KB–51 KB raw). WebFetch converts to markdown before it reaches context — do not curl them. docs_query.py get <url> exists but only prints this instruction, because piping raw Databricks HTML into context wastes tokens.

Fetch 1–3 pages per batch, then judge whether that answers the question. Loop if not, to a cap of 9 pages. At 9 and still short, stop and tell the user what you read and what is missing — do not silently continue or fill the gap with guesses.

Context budget

Step Measured cost Notes
search (narrow) ~58 tok row filter|column mask|fine-grained → 230 B
search (typical) ~160–200 tok lineage → 643 B; schedul|... → 781 B
search (broad) ~1,290 tok Unity Catalog → 5,147 B, 23 hits
sections ~253 tok 1,013 B
section 222–1,617 tok Troubleshooting 890 B … ML and AI 6,468 B
page via WebFetch ~0.3–1k tok from 21–51 KB of raw HTML

Typical question: search + 1–2 pages ≈ 1–2k tokens. Loading the index whole would be ~11,787 tokens, which is why nothing here does that. Read one section at a time; if you cannot tell which section to read, that is a signal to search again with better terms, not to load several.

Rules

  • Never invent a doc URL. Not in the index → search wider, or descend a hub. Databricks renames slugs (dlt/ → ldp/); an invented URL is a confident 404.
  • Never load the whole index, and never touch the marketing site's llms-full.txt.
  • Cite the URL that served the content — the post-redirect /aws/en/… one.
  • A 404 on a page URL means the index is stale — re-run /docs-skill-builder check .claude/skills/databricks/skills/databricks-docs.
  • Say which cloud you are answering for. The default is AWS; GCP pages live under /gcp/en/ and can differ. Azure is not on this host at all.
  • Pass through what the docs say. The user wants current authoritative behavior, not a synthesis with your prior knowledge.

04 同仓库更多 Skills