AutoRAG Search Methods

Understand the different search methods available in AutoRAG

Getting Started Path: Overview → Configure LLMs → Search Methods → Execute Queries → Verify

Overview

AutoRAG provides multiple search methods that leverage the structured knowledge graph created by the Importer to deliver accurate and contextually relevant responses to your natural language queries.

Method	`query_type`	`mode`	Best for	Latency
Global Search	`1` (GLOBAL)		Themes, patterns, high-level insights	Medium
Local Search	`2` (LOCAL)		Specific entities and relationships	Low
Deep Search	`2` (LOCAL) + `use_llm_planner: true`	`DEEP_SEARCH`	Detailed, multi-step research	Higher
Unified (Instant Search)	`3` (UNIFIED)	`INSTANT`	Fast answers with document references	Low
Custom Retriever	`4` (CUSTOM)	`DEEP_SEARCH`	Domain-specific search with custom logic	Varies

Instead of choosing a method from the table, you can send mode and let the service pick: INSTANT for a fast answer, or DEEP_SEARCH for a thorough one. DEEP_SEARCH uses your Custom Retriever tools if you have any, and Local Search if you do not. Set a query_type yourself when you want a specific method, such as Global Search.

VectorRAG and FullGraphRAG partitions

A partition is ingested either as VectorRAG or as FullGraphRAG, and the search methods you can run against it depend on that choice. AutoGraph assigns the strategy per cluster, and the Importer receives it as rag_mode.

VectorRAG stores documents and chunk embeddings. It does not build entities or communities.
FullGraphRAG builds the full knowledge graph (entities, relations, and communities) in addition to chunks.

Method	VectorRAG	FullGraphRAG
Unified (Instant Search)	Supported	Supported
Global Search	Not supported	Supported
Local Search	Not supported	Supported
Deep Search (Local Search plus planner)	Not supported	Supported
Custom Retriever, including Custom Deep Search	Supported when the tools search chunks	Supported

On a VectorRAG partition, use Instant Search. Global Search, Local Search, and Deep Search need entities and communities, which VectorRAG does not build, so they are not supported there. Deep Search does not fall back to chunk search when entities are missing.

If you need Deep Search on a VectorRAG partition, use the Custom Retriever with tools that search chunks.

For how AutoGraph decides which clusters become VectorRAG and which become FullGraphRAG, see the RAG strategizer. For how the resulting partitions are named and imported, see the Importer AutoGraph Integration.

Choosing a search method

Need high-level summaries across all documents? Use Global Search.
Need details about a specific entity? Use Local Search.
Need thorough, accurate analysis? Use Deep Search.
Need a quick answer with references? Start with Unified (Instant Search).
Need domain-specific search on custom collections? Use Custom Retriever.
Querying a VectorRAG partition? Use Unified (Instant Search), or a Custom Retriever with chunk-searching tools if you need Deep Search. See VectorRAG and FullGraphRAG partitions.

Next Steps

Execute queries: Learn how to call the search endpoints.
Parameters: Customize search behavior.