LlamaIndex Cookbooks & Examples Finder

Route a user's described scenario to the most relevant existing recipes: the LlamaParse/LlamaCloud cookbooks and the LlamaIndex examples gallery. This skill is a router into the examples — it surfaces topics and URLs for inspiration and coverage checking. It does NOT re-teach any product. Once a recipe is found, hand the actual how-to back to the matching product skill.

When this skill runs

Trigger on discovery questions, not how-to questions:

• "Is there a cookbook/notebook/example for <scenario>?"
• "Find me a recipe/sample that does <thing>."
• "Show me a LlamaIndex example for <integration/pattern>."
• Coverage checks: "Does an official example already cover this, or do I build from scratch?"

If the user instead wants to do the task (write the schema, configure parse, build the index), that is a product-skill job — see the handoff table at the bottom. This skill points; the product skills teach.

Core method: find a recipe for any scenario

Do not rely on a memorized list — the gallery is large and changes. Work from the live catalog hubs plus search. Follow this sequence:

1. Classify the scenario into a category using the map below (e.g. "merged PDF of many invoices" → parse + extract cookbooks; "chat over my docs with citations" → index/RAG gallery section).
2. Search the live catalog with the doc tools (see "Fetching facts"). Query the scenario noun plus the pattern, e.g. search_docs "agentic RAG citations", grep_docs "SubQuestionQueryEngine".
3. Confirm against the hub pages. The two anchor index pages list the curated, currently-valid entry points; the gallery sidebar holds the long tail. Fetch the hub index.md to get fresh URLs rather than pasting remembered ones.
4. Evaluate fit (see references/recipe-finder.md): does the recipe match the user's data type, product (Parse vs Extract vs Index), and pattern — not just a keyword? Surface 1–3 best matches with their topic and URL, note any gap, and hand off to the product skill.
5. If nothing fits, say so plainly and route to the closest product skill to build from scratch. A weak keyword match is worse than an honest "no direct recipe; closest is X."

Keep the answer short: topic + URL + one line on why it fits. Do not paste notebook contents.

Worked example

Scenario: "Is there a notebook for pulling line items out of a stack of scanned invoices and then letting me chat over them?"

1. Decompose into axes (see references/recipe-finder.md): data type = scanned PDFs; layer = two legs — typed extract (line items) + managed index/RAG (chat); pattern = schema-driven extraction then retrieval.
2. Catalog: mixed → both. Extract leg → Cookbooks (extract); chat leg → Cookbooks (index). "Scanned" also touches parse robustness.
3. Search: search_docs "invoice extraction schema" and search_docs "index search chat over documents"; grep_docs "LlamaCloudIndex" to confirm the managed-index entry point.
4. Hub-confirm: fetch …/llamaparse/cookbooks/index.md for the current extract + index URLs.
5. Return, ranked:
   • Extract leg → extract Starter cookbook (resume_screening pattern, adapt schema to invoice fields) — URL — closest typed-extraction recipe; swap the schema.
   • Chat leg → index getting_started cookbook — URL — covers index + search + chat end to end.
   • Gap note: no single notebook chains scanned-invoice extract → chat; combine the two, and hand the extract schema design to llamacloud-extract, the RAG build to the index product skill.

That gap note is the deliverable's value: an honest "no one recipe does all of it; here are the two legs" beats forcing a single weak match.

Coverage checking (the inverse query)

This skill owns coverage checking as much as discovery — the question "is this already covered, or do I build from scratch?" Treat it as the inverse of recipe-finding:

• Run the same decompose + search + hub-confirm method, but the output is a verdict, not a recipe.
• Covered: a recipe matches all three axes (data type, layer, pattern). Return it and stop — route the user to adapt it, not rebuild it.
• Partially covered: recipes exist for some legs/axes but not their combination (the worked example above). Name what is covered, name the gap, and point each leg at its product skill.
• Not covered: no recipe matches the layer/pattern. Say so plainly and route to the nearest product skill to build from scratch. Do not invent a match to avoid an empty answer.
• Apply the fit rubric (references/recipe-finder.md) before declaring "covered" — a keyword hit on a different data type or layer is not coverage.

The two catalogs

LlamaParse / LlamaCloud cookbooks — product-centric, organized as Starter vs Advanced, across three services: LlamaParse, LlamaExtract, LlamaCloudIndex. Get current URLs from the cookbooks hub index.md (below) — it is the source of truth. (Underlying notebooks historically lived in the llama_cloud_services repo under examples/{parse,extract,index,batch,sheets,split,misc}/, but treat repo paths as rot-prone and confirm via the hub.) Use these for end-to-end LlamaCloud product scenarios (parse a contract, extract fund data, index + chat).

LlamaIndex examples gallery — framework-centric. Curated hub highlights Agents, Agentic Workflows, LLM Integrations, Embedding Models, Vector Stores; the full sidebar adds many more (query/chat engines, retrievers, postprocessors, structured outputs, evaluation, observability, finetuning, property graph, ingestion, multimodal, managed indexes). Use these for framework patterns and third-party integrations (which vector store, how to build a ReAct agent, text-to-SQL, evaluation).

Anchor doc paths (append index.md and fetch for fresh URLs):

• https://developers.llamaindex.ai/llamaparse/cookbooks/
• https://developers.llamaindex.ai/python/examples/

Scenario → category quick map

A small, durable starting map. It points to the right category, then search/hub-fetch from there for the exact current notebook. The full scenario→category table and fit rubric are in references/recipe-finder.md.

Fetching facts (live lookups — do not memorize)

URLs, notebook filenames, and gallery sections drift. Always confirm live:

• mcp__plugin_llamacloud_docs__search_docs — concept/BM25 search across the docs (best first pass: search the scenario noun + pattern).
• mcp__plugin_llamacloud_docs__read_doc — read a known doc/cookbook page once you have a path.
• mcp__plugin_llamacloud_docs__grep_docs — exact symbol/class/regex search (e.g. a specific engine or integration name).
• Fallback: append index.md to any https://developers.llamaindex.ai/... page URL and WebFetch it (e.g. the two anchor hubs above) to get current entry points and URLs.

Treat any notebook filename in this skill as illustrative, not authoritative — verify it still exists.

Handoffs — this skill points, product skills teach

After surfacing a recipe, hand the actual implementation to the matching product skill:

• Parsing a document → llamacloud-parse (cloud) or liteparse (local/no-key).
• Typed structured extraction / schema design → llamacloud-extract.
• Spreadsheet table extraction → llamacloud-sheets.
• Splitting a merged file into sections → llamacloud-split.
• Managed index / RAG / retrieval → llamacloud-index.
• Account, API key, region, "which product" → core llamacloud.

This skill does NOT teach parse/extract/index/split/sheets configuration, schema design, or RAG construction. It also backs the read-only cookbook-scout agent.