LlamaIndex provides a comprehensive framework for Retrieval-Augmented Generation (RAG). From node parsers and index types to query engines and response synthesizers, understanding the pipeline components is key to building accurate, scalable knowledge retrieval systems.
0 / 5 completed
1 / 5
In LlamaIndex, what does a Node Parser do to ingested documents?
A Node Parser takes raw Document objects and splits them into TextNode chunks. It handles chunk size, overlap, and can preserve structural relationships (e.g., parent-child for hierarchical retrieval) and propagate metadata from document to nodes.
2 / 5
What is the difference between a VectorStoreIndex and a SummaryIndex in LlamaIndex?
VectorStoreIndex embeds nodes and retrieves the top-k most semantically similar ones for a query. SummaryIndex is designed for summarisation use cases: it iterates over all nodes in order and synthesises a response across the full document, without vector retrieval.
3 / 5
A Query Engine in LlamaIndex wraps a Retriever and a Response Synthesizer. What does the Response Synthesizer do?
The Response Synthesizer is the component that calls the LLM. It receives the original query plus the retrieved source nodes and applies a synthesis mode (e.g., compact, refine, tree_summarize) to produce the final answer, managing prompt construction and multi-chunk scenarios.
4 / 5
How do metadata filters improve retrieval in LlamaIndex?
Metadata filters are pre-retrieval constraints applied at the vector store query level. By filtering on fields like document date, author, or category, you narrow the candidate pool before similarity ranking, increasing precision without changing chunk embeddings.
5 / 5
In a LlamaIndex SubQuestion Query Engine, what happens when a complex question is submitted?
The SubQuestion Query Engine first prompts an LLM to break a complex question into simpler sub-questions, each matched to a relevant data source (index or tool). It retrieves answers for each sub-question independently, then synthesises a final, comprehensive answer.