#Retrieval

A retriever only answers “given a query string, return Nodes.” It does not call the LLM. That is the first stage inside a QueryEngine. Pulling it out lets you mix retrievers, wrap it as an agent tool, or call it from a Workflow step.

retriever = index.as_retriever(similarity_top_k=5)
nodes = retriever.retrieve("How do I request time off?")
for n in nodes:
    print(n.score, n.metadata, n.get_content()[:200])

Async: await retriever.aretrieve(query). Score scales differ by backend (cosine / inner product / distance)—do not compare raw scores across stores.

#VectorIndexRetriever

Explicit form of as_retriever:

from llama_index.core.retrievers import VectorIndexRetriever

retriever = VectorIndexRetriever(
    index=index,
    similarity_top_k=10,
)

as_query_engine(similarity_top_k=...) forwards that argument to the inner retriever. Debug recall with a fixed k and printed text before attaching a synthesizer.

#Postprocessors: filter and rerank

From the official querying guide:

from llama_index.core.postprocessor import (
    SimilarityPostprocessor,
    KeywordNodePostprocessor,
)
from llama_index.core.query_engine import RetrieverQueryEngine

engine = RetrieverQueryEngine.from_args(
    retriever,
    node_postprocessors=[
        KeywordNodePostprocessor(required_keywords=["policy"], exclude_keywords=["draft"]),
        SimilarityPostprocessor(similarity_cutoff=0.65),
    ],
)

Production often adds a reranker (cross-encoder) on the top-k. Class names and extra packages change—use current Node Postprocessor docs rather than stale imports.

#Metadata filters

metadata written at ingest time can filter at query time (tenant, year, file type). Syntax varies slightly by vector store; the idea is the same: shrink candidates, then score.

from llama_index.core.vector_stores import MetadataFilters, MetadataFilter, FilterOperator
from llama_index.core.vector_stores.types import FilterCondition

filters = MetadataFilters(
    filters=[
        MetadataFilter(key="department", value="hr", operator=FilterOperator.EQ),
    ],
    condition=FilterCondition.AND,
)
retriever = index.as_retriever(similarity_top_k=5, filters=filters)

Keys must match ingest. If your installed version’s MetadataFilters API differs, follow the official vector-store page.

#Better recall (same ideas as the RAG course)

The RAG tutorial explains why. Here we map where. When the model chooses query strings and hops, that is agentic RAG: expose the retriever or query engine as a FunctionAgent tool (Query Engines).

#Retrieve only, generate yourself

from llama_index.llms.openai import OpenAI

llm = OpenAI(model="gpt-4o-mini")
nodes = retriever.retrieve("How many days is the return window?")
context = "\n\n".join(n.get_content() for n in nodes)
prompt = f"Answer only from the notes. If unknown, say so.\n\n{context}\n\nQ: return window?"
print(llm.complete(prompt))

This is easy to unit-test (assert on retrieve) and easy to hand to LangChain or CrewAI: they receive a string tool result and do not need LlamaIndex’s synthesizer.

#Next steps

• Workflows
• Vector Stores
• Practical Examples