RAG Pipelines (in LlamaIndex)

This is where it all comes together. A RAG pipeline in LlamaIndex is the full path from raw documents to a grounded answer: load → parse → index → store → query. LlamaIndex gives you this in five lines for the simple case — then lets you swap in better parsing, hybrid retrieval, and reranking without rewriting the pipeline.

💡 In one line: A LlamaIndex RAG pipeline runs load → parse → index → store → query, with each stage independently upgradable.

The Two Halves

Like any RAG system, a LlamaIndex pipeline has:

• Ingestion (offline) — load → parse into nodes → embed → store.
• Query (online) — question → retrieve → postprocess → synthesise → answer.

The index bridges them.

The Basic Pipeline

Five lines is the starting point, not the ceiling.

The Ingestion Pipeline

For production, IngestionPipeline makes ingestion explicit and reusable — a list of transformations applied to documents:

It also supports caching (skip re-processing unchanged docs) and document management (upserts via doc IDs, so re-running doesn't duplicate).

Upgrading Each Stage

The point of the abstraction: improve one stage without touching the rest.

Advanced Pipelines

Beyond the basics — patterns you've met in the RAG topic, available here as first-class features:

• Sentence-window retrieval — small match, wider context returned.
• Auto-merging / hierarchical — merge child nodes into parents.
• Sub-question decomposition — split a complex question.
• Recursive retrieval — follow node references.
• Router / agentic RAG — query engines as tools an agent chooses.

Whiteboard

Workflows (the Modern Path)

For anything non-trivial, LlamaIndex recommends Workflows — event-driven step composition with branching, loops, and parallelism. (Query Pipelines are deprecated in favour of it, and query engines are themselves Workflows under the hood.)

Evaluation

Measure both halves: retrieval (hit rate, MRR) and response (faithfulness, relevancy). LlamaIndex ships evaluators — FaithfulnessEvaluator, RelevancyEvaluator — and integrates with RAGAS.

Production Concerns

• Persist the index; cache ingestion.
• Manage updates with doc-ID upserts — don't rebuild blindly.
• Watch embedding cost on large corpora.
• Add observability (Langfuse, Phoenix, or LlamaCloud).

Best Practices

• Start with the 5-line default, then measure.
• Fix retrieval first — better parsing, chunking, and reranking beat prompt-tweaking.
• Use LlamaParse for messy documents.
• Persist and cache; evaluate before and after each change.

Summary

• A LlamaIndex RAG pipeline runs load → parse → index → store → query.
• IngestionPipeline makes ingestion explicit, with caching and upserts.
• Each stage is independently upgradable — parser, retriever, reranker, synthesiser.
• Advanced patterns (sentence-window, auto-merging, sub-question, agentic RAG) are first-class.
• Use Workflows for non-trivial apps, and evaluate retrieval and response.