llamacloud-index

LlamaCloud Cloud Index — Managed RAG

LlamaCloud Cloud Index is a fully managed RAG pipeline: connect a data source, and LlamaCloud parses, transforms, chunks, embeds, and stores documents into a sink, then exposes a hosted retrieval endpoint. Choose it to get production retrieval without operating ingestion workers or a vector database. This skill owns the build (source → sink) and retrieval tuning; it does not teach OSS RAG, raw parsing, or extraction.

When managed Cloud Index beats self-managed

Prefer Cloud Index when the value is retrieval that works, not infrastructure:

• No infra to run. Ingestion, embedding, and the vector store are hosted; no workers, no DB ops.
• Auto-sync sources. Connected data sources (cloud buckets, drives) re-ingest on a schedule — new/changed files flow in without custom pipelines.
• Parse quality built in. LlamaParse runs inside the pipeline, so complex PDFs/tables become clean chunks without a separate parse step.
• One retrieval endpoint. Tuning (hybrid/rerank/filters) is a request parameter, not a redeploy.

Prefer self-managed OSS (llamaindex-framework with VectorStoreIndex + your own vector DB) when: you need full control of chunking/embedding code, must keep data on existing local infra, want a vector store LlamaCloud doesn't host, or are doing rapid offline experimentation. See references/ingestion.md for the full decision table.

Build sequence (source → sink)

A Cloud Index is defined by an ordered pipeline. Decide each stage deliberately:

1. Data source — file upload for one-off corpora; a connected source (cloud bucket, drive) for corpora that grow and must stay in sync. Connectors and their exact config are live lookups.
2. Parse / transform — toggle LlamaParse and pick a tier. Default (empty params) is the cost preset: fast, fine for clean text, weak on complex layout. Set an agentic tier for tables/scans/multi-column. Then choose segmentation (none / page / element — element gives structure-aware units but is incompatible with fast parse) and chunking (character / token / sentence / semantic). See references/ingestion.md for which to pick.
3. Embedding model — pick the embedding model (provider key required for hosted providers). Optionally add a sparse model (splade / bm25 / auto) to enable hybrid search at the sink. Decide this now: hybrid must be enabled at ingestion, not bolted on at query time.
4. Sink — Fully Managed sink for the zero-ops path; an external vector store (BYO) when you must own the data plane. Hybrid search is only available on sinks that support sparse vectors.
5. Deploy — ingestion runs; then copy the retrieval endpoint / API key into the app.

Indexes are created and managed via the API/SDK as the durable default. A UI flow also exists, and the legacy "Index v1" UI is restricted to projects with pre-existing pipelines — treat API/SDK management as the path to teach.

Retrieval tuning

The retrieval endpoint accepts tuning per request — no re-ingestion needed (except enabling hybrid, which is set at build time). Three levers, applied in this order of impact:

• Retrieval mode — pick granularity for the query shape: chunk-level for pinpoint fact-finding; file-level when the right document matters (by metadata, or by holistic content for summarization); auto-routed when query types vary. Verify the current mode enum via MCP.
• Hybrid search — combine dense (semantic) + sparse (keyword) with an alpha blend. Shift toward keyword for exact terms/codes/names; toward semantic for paraphrased intent. Requires a sparse model configured at build time.
• Reranking — retrieve a wide candidate set (high top-k), then rerank and keep the top few. Turn on when precision of the final ordering matters more than latency.
• Metadata filtering — constrain by file/tag/date/category, or for access control. Use explicit filters when the app knows the constraint; use metadata filter inference (a schema the system reads to pull filters from natural-language queries) when end users phrase constraints in prose.

See references/retrieval-tuning.md for the mode-selection table, the hybrid/rerank/filter decision table, alpha-tuning intuition, and metadata-schema authoring tips.

Version posture

Teach the stable Cloud Index surface. Cloud Index v2 is beta and deferred — flag it as beta and do not make it the default path. Do not present v2 as GA.

SDK examples (minimal — confirm signatures via MCP)

Use the high-level services SDK. Confirm the exact package name and method signatures live before pinning — the Python package has shifted between llama-cloud and llama-cloud-services, and retrieval may run through llama-index client classes. These snippets show shape, not exact API:

# Python — query a deployed Cloud Index's retrieval endpoint (illustrative)
# Verify class/param names via the docs MCP before using.
retriever = index.as_retriever(
    dense_similarity_top_k=10,   # widen, then rerank
    alpha=0.5,                   # 0 = keyword, 1 = semantic
    enable_reranking=True,
    rerank_top_n=3,
)
nodes = retriever.retrieve("...query...")

// TypeScript — shape only; fetch the current @llamaindex client + params via MCP.
const retriever = index.asRetriever({ similarityTopK: 10, enableReranking: true });
const nodes = await retriever.retrieve("...query...");

For exact, current creation/retrieval signatures and connector/sink config, use the MCP — do not copy these verbatim.

Fetching facts (do not hardcode)

Treat package names/versions, class/method signatures, connector and sink names, parameter schemas, and credit numbers as live lookups. Fetch via:

• mcp__plugin_llamacloud_docs__search_docs — concept / BM25 search.
• mcp__plugin_llamacloud_docs__read_doc — read a known doc page.
• mcp__plugin_llamacloud_docs__grep_docs — exact symbol / regex search.
• Fallback: append index.md to any https://developers.llamaindex.ai/... page URL and WebFetch it.

Anchor doc paths:

• /llamaparse/cloud-index/getting_started/ — build flow, sources, sinks, deploy.
• /llamaparse/cloud-index/guides/parsing_transformation/ — parse tier, segmentation, chunking, embedding/sparse.
• /llamaparse/cloud-index/guides/retrieval/advanced/ — hybrid, rerank, metadata filtering & inference.

Handoffs

• Self-managed / OSS RAG (VectorStoreIndex, own vector DB) → llamaindex-framework.
• Raw document → clean text/markdown only → llamacloud-parse.
• Typed field/JSON extraction → llamacloud-extract.
• Account, API key, region, pricing/credits → core llamacloud.

Additional resources

• references/ingestion.md — managed-vs-self-managed decision table, source choice, parse-tier and segmentation/chunking decisions, embedding + sparse-model choices, sink selection, gotchas.
• references/retrieval-tuning.md — retrieval-mode selection table, hybrid/rerank/metadata-filter decision table, alpha intuition, metadata filter-inference schema tips, anti-patterns.