LlamaCloud — Router & Onboarding

LlamaCloud is LlamaIndex's document-AI platform. This skill is the entry point: it picks the right product for a task, gets the user authenticated and configured, and hands off to the specialized skill. It holds the cross-product mental model and onboarding judgment — it does not teach any single product in depth.

When to route elsewhere

After identifying the product, defer to the owning skill (it auto-triggers on product verbs):

If the task spans products, stay here long enough to sequence them, then hand off in order.

The product mental model

Parse is the foundation most pipelines build on: document → clean text/markdown. Its output then feeds:

• Extract (structured fields), Classify (category routing), Split (segmentation), and
• Cloud Index (chunk → embed → retrieve for RAG).

Decision shortcuts:

• Need typed fields (invoice total, parties, dates) → Extract, not Parse. Extract parses internally; do not chain Parse→Extract manually unless you need the raw text too.
• Need semantic search / RAG over a corpus → Cloud Index (managed) or the OSS llamaindex-framework (self-managed). Choose managed when you want hosted ingestion + retrieval without running a vector DB; choose the framework when you need full control or local infra.
• Just need clean text/markdown for your own downstream LLM call → Parse.
• A pile of mixed document types arriving together → Classify (route) and/or Split (segment) before Parse/Extract.

See references/product-routing.md for the full decision tree, multi-product pipelines, and the managed-vs-OSS and Parse-vs-Extract boundary calls.

Version posture (authoritative — apply everywhere)

• Parse v2 and Extract v2 are the primary GA surface. Generate v2 code by default.
• Their /v1/ doc trees are legacy / migration reference only — never the default path.
• Cloud Index v2 is beta and deferred — teach the stable Cloud Index surface; flag v2 as beta.
• LiteParse is local/OSS with no API key.
• Classify, Split, and Sheets are beta — set expectations accordingly.

Never present a legacy (/v1/) or beta path as the default. This posture is the plugin's default-surface directive (durable); the GA/beta labels themselves drift — when a product's status matters to the answer, confirm the current GA/beta state live via the docs MCP before relying on it.

Onboarding sequence

Run this order when a user is setting up (the /llamacloud-setup command wraps it):

1. Account + key. Sign in at the LlamaCloud console and create an API key. The whole platform uses one credential, exposed to code as ${LLAMA_CLOUD_API_KEY}. Never hardcode it; never echo it.
2. Pick a region — this is a one-time, non-reversible choice:
   • NA (api.cloud.llamaindex.ai) — gets new features first; default.
   • EU (api.cloud.eu.llamaindex.ai) — data stored & processed in-region; pick for GDPR / EU data-residency. Usable from anywhere if residency demands it. Cannot migrate later. The SDK targets NA by default; point it at the EU base URL when EU is chosen.
3. Install the SDK. One package serves all products. Confirm the exact current package name and version live (it has shifted — see "Fetching facts" below) before pinning it.
4. First call. Verify the key with the smallest possible operation in the chosen product, then hand off to that product's skill.

See references/onboarding.md for the step-by-step setup checklist, the env-var/region wiring pattern (Python + TypeScript), key security/rotation guidance, and the pricing/credits mental model.

Fetching facts (do not hardcode)

Treat these as live lookups, never memorized: exact package names/versions, SDK class and method signatures, base-URL constants, pricing/credit numbers, and rate limits. They drift. Fetch them via:

• Docs MCP (wired, no auth): mcp__plugin_llamacloud_docs__search_docs for concepts, mcp__plugin_llamacloud_docs__read_doc for a known page, mcp__plugin_llamacloud_docs__grep_docs for an exact symbol.
• Clean-markdown fallback: append index.md to any https://developers.llamaindex.ai/... page URL and WebFetch it.

Anchor pages: /llamaparse/ (overview), /llamaparse/general/api_key/, /llamaparse/general/regions/, /llamaparse/general/pricing/.

SDK packages (verify live before pinning)

Two layers exist and the naming has changed across releases, so confirm before quoting:

• A high-level services SDK (commonly llama-cloud-services in Python) exposing LlamaParse, LlamaExtract, etc. — the recommended surface for v2.
• A lower-level generated client (llama-cloud / @llamaindex/llama-cloud).

Default to the high-level services SDK for Parse/Extract v2; fetch the exact current install command from the docs rather than asserting it from memory.

Additional resources

• references/product-routing.md — full routing decision tree, multi-product pipelines, the managed-vs-OSS and Parse-vs-Extract-vs-Index boundaries, and anti-patterns.
• references/onboarding.md — setup checklist, region/env wiring, key security, pricing/credits mental model, common first-run failures.