Skip to content

Query System Guide

GraphRAG provides two search approaches for querying indexed knowledge graphs. Each approach is designed to handle a different category of question.

Overview: Two Query Types

GraphRAG provides two distinct search modes to handle different types of questions:

flowchart LR
    question["🔍 Your Question"] --> decision["What type of answer?"]

    decision --> local["🎯 Local Search<br/>Specific Facts"]
    decision --> global["🌍 Global Search<br/>Broad Themes"]

    local --> localout["Detailed entity info<br/>Precise relationships<br/>Source references"]
    global --> globalout["Thematic analysis<br/>Pattern analysis<br/>Content synthesis"]

Local Search: Specific Entity Questions

Purpose: Answer detailed questions about specific people, organizations, or concepts

How It Works

Local Search finds and analyzes entities and relationships that are most relevant to your question, then provides detailed answers with source references.

flowchart LR
    question["Your Question"] --> entities["Find Relevant Entities"]
    entities --> context["Get Related Context"]
    context --> answer["Build Answer with References"]

Question Types

Question Type Example Why Local Search
Specific People "What role did Ratan Tata play in Tata Group?" Finds entity-specific information
Relationships "How are Tata Consultancy Services and Tata Group connected?" Retrieves relationship details
Company Details "What businesses does Reliance Industries operate?" Gathers entity-focused facts
Event Details "What happened during the 2008 Mumbai attacks?" Finds specific event information

Example Queries

Local Search Questions: - "Who is the CEO of Tata Consultancy Services?" - "What products does Infosys offer?" - "How is Wipro related to the IT industry?" - "What awards has A.P.J. Abdul Kalam received?"

Local Search Response Format: - Detailed factual answers - Source references to original documents - Specific entity and relationship information - Precise, targeted responses

Global Search: Thematic Analysis Questions

Purpose: Analyze broad themes and patterns across your entire knowledge base

How It Works

Global Search analyzes community reports (thematic clusters) to identify key patterns and synthesize insights across the entire knowledge graph.

flowchart LR
    question["Your Question"] --> reports["Analyze Community Reports"]
    reports --> keypoints["Generate Key Points"]
    keypoints --> insights["Synthesize Analysis Results"]

Question Types

Question Type Example Why Global Search
Strategic Themes "What are the main challenges facing Indian technology companies?" Analyzes patterns across communities
Industry Trends "How has the Indian business landscape evolved?" Synthesizes broad insights
Comparative Analysis "What are the different leadership styles in major Indian corporations?" Compares across multiple entities
Impact Assessment "What has been the impact of digital transformation in India?" Evaluates system-wide effects

Example Queries

Global Search Questions: - "What are the major trends in Indian business leadership?" - "How do technology companies in India approach innovation?" - "What are the common challenges facing Indian startups?" - "What patterns emerge from successful Indian entrepreneurs?"

Global Search Response Format: - Comprehensive thematic analysis - Thematic analysis and patterns - Synthesized information from multiple sources - Broader perspective on topics

Search Type Selection

Use Local Search for:

Your Question Contains Examples
Specific entity names "What does TechCorp do?" "Who is Alice Smith?"
Direct relationships "How are CompanyA and CompanyB connected?"
Factual details "When did this event happen?" "What are the specifications?"
Entity attributes "What products does this company make?"

Use Global Search for:

Your Question Asks About Examples
Patterns across data "What trends appear in the technology sector?"
Themes and topics "What are the main challenges discussed?"
Comparative analysis "How do different approaches compare?"
System-wide insights "What patterns emerge across all documents?"

Implementation Approach

Local Search Setup

For precise, entity-focused questions about specific people, companies, or relationships:

from langchain_graphrag.query.local_search import LocalSearch, LocalSearchPromptBuilder, LocalSearchRetriever

# Configure with your LLM and indexed artifacts
local_search = LocalSearch(
    llm=your_llm,
    prompt_builder=LocalSearchPromptBuilder(),
    retriever=LocalSearchRetriever(
        context_selector=your_context_selector,
        context_builder=your_context_builder,
        artifacts=your_artifacts
    )
)

# Get the search chain and ask specific questions
search_chain = local_search()
response = search_chain.invoke("What role did Ratan Tata play in Tata Group?")

Sample Response:

Ratan Tata served as Chairman of Tata Group from 1991 to 2012, leading major international acquisitions including Jaguar Land Rover and Corus Steel [Data: Entities (12, 45), Relationships (156)]. He transformed the group into a global conglomerate operating in over 100 countries.

Global Search Setup

For thematic analysis across your entire knowledge base:

from langchain_graphrag.query.global_search import GlobalSearch

# Configure with community analysis components
global_search = GlobalSearch(
    kp_generator=your_kp_generator,
    kp_aggregator=your_kp_aggregator
)

# Ask thematic questions
response = global_search.invoke("What are the leadership patterns in Indian business?")

Sample Response:

Leadership Patterns in Indian Business

Visionary Innovation with Social Responsibility: Leaders like Ratan Tata and N.R. Narayana Murthy combine aggressive expansion with strong social consciousness [Data: Reports (12, 34, 67)].

Long-term Relationship Focus: Emphasis on multi-generational partnerships over short-term profits [Data: Reports (23, 45, 89)].

Best Practices

Optimal Query Design

Local Search Characteristics: - Specific entity names and timeframes - Direct relationships between known entities - Factual details and historical information - "Who," "what," "when," and "where" questions

Global Search Characteristics: - Questions without specific entity names mentioned - Asking about "trends," "patterns," "themes," or "strategies" - Comparing multiple companies/sectors without naming them - "What are the trends in...", "How do companies approach...", "What patterns emerge..." questions

Query Quality Guidelines

Factor Local Search Global Search
Scope Narrow, specific Broad, thematic
Entities Named entities preferred Entity-agnostic
Timeframe Specific periods Trend analysis
Response Factual details Thematic analysis

Data Flow Examples
Ready to see real data transformations?

Documentation Index
Return to documentation overview

You now understand the core GraphRAG concepts and implementation. The query system transforms your indexed knowledge into actionable insights through both precise fact-finding and strategic analysis capabilities.