English for LlamaIndex Developers

LlamaIndex conversations combine data-pipeline vocabulary with retrieval-augmented-generation terms, so a developer new to RAG needs both the general concepts and LlamaIndex’s specific naming for each pipeline stage.

Key Vocabulary

Document loader — a component that pulls raw content from a source, like a PDF, a website, or a database, and converts it into LlamaIndex’s internal Document objects for indexing. “Swap the generic file loader for the PDF-specific one — it’s preserving the table structure much better than plain text extraction.”

Node — a chunk of a document, split according to a chosen strategy, that becomes the actual unit stored and retrieved from an index. “These nodes are too large — the retriever keeps pulling back whole sections when we only need the one relevant paragraph.”

Index — a data structure, such as a vector store index or a summary index, that organizes nodes so they can be efficiently searched at query time. “We’re using a vector store index for semantic search, but a summary index would be a better fit for the ‘give me an overview’ queries.”

Retriever — the component responsible for fetching the most relevant nodes from an index given a query, before those nodes are passed to the language model for synthesis. “The retriever’s returning irrelevant nodes — try tuning the similarity threshold or switching to a hybrid retriever before touching the prompt.”

Query engine — the end-to-end pipeline that takes a natural-language question, retrieves relevant nodes, and synthesizes a final answer using an LLM. “Wrap the retriever and the response synthesizer into a single query engine so the API surface is just .query() for consumers.”

Common Phrases

• “Is the document loader preserving structure, or are we losing table and heading information at ingestion?”
• “Are these nodes chunked at the right size, or is the retriever pulling back too much context?”
• “Which index type fits this use case — vector search, or something more summary-oriented?”
• “Is the retriever the bottleneck here, or is the language model failing to synthesize a good answer from good context?”
• “Should we expose this as a query engine, or does the consumer need lower-level access to the retriever?”

Example Sentences

Debugging a retrieval quality issue: “The answers are vague because the nodes are too coarse — re-chunk with a smaller node size so the retriever can surface the specific passage instead of an entire chapter.”

Explaining an architecture choice: “We picked a vector store index for the FAQ search but kept a separate summary index for the ‘explain this whole document’ feature — one index type doesn’t fit both use cases.”

Reviewing a pull request: “This query engine skips reranking entirely — add a reranker step before synthesis, or we’re relying on raw similarity scores alone.”

Professional Tips

• Distinguish document from node precisely — a document is the whole source, a node is the chunked retrieval unit, and confusing them makes chunking discussions unclear.
• Name the index type explicitly when discussing performance — “the index is slow” is vague; “the vector store index” or “the summary index” tells the team exactly what to investigate.
• Say retriever rather than “search” when diagnosing RAG quality issues — it isolates whether the problem is retrieval or generation.
• Use query engine as the term for the full pipeline when describing the system to non-RAG-specialist teammates — it’s the natural unit of abstraction to expose.

Practice Exercise

1. Explain the difference between a document and a node in LlamaIndex’s pipeline.
2. Describe what a retriever does and how it differs from a full query engine.
3. Write a sentence explaining why chunk size affects retrieval quality.