LlamaIndex Querying and Retrieval

Control what is retrieved, how much is retrieved, whether results are reranked and how evidence reaches the answer.

#Query engine

pythonCopy
query_engine = index.as_query_engine()
response = query_engine.query("Summarize this document.")
print(response.response)
print(response.source_nodes)

The response text and its source nodes should be inspected together.

#Retrieval without answer generation

pythonCopy
retriever = index.as_retriever(similarity_top_k=3)
nodes = retriever.retrieve("What is a vector database?")

#Additional patterns

A chat engine can retain conversational context:

pythonCopy
chat_engine = index.as_chat_engine()
chat_engine.chat("My name is Alex.")
response = chat_engine.chat("What name did I give you?")

A router can select among indexes. A sub-question workflow can divide a comparison into smaller retrieval tasks and combine the results. Neither is necessary for every query.

#Retrieval improvements

Metadata filters restrict the relevant dates/categories. Reranking can narrow a larger initial set—for example twenty candidates to three reviewed passages. Hybrid search combines semantic and keyword signals. Measure basic recall before adding each stage.

#Structured responses

pythonCopy
from pydantic import BaseModel

class Info(BaseModel):
    name: str
    amount: float

query_engine = index.as_query_engine(output_cls=Info)

This illustrates a typed output target; validate the selected integration's support and the actual returned values.

If answers are only partly reliable, inspect top_k, metadata scope and candidate ordering before changing the generation prompt or model.