> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/microsoft/graphrag/llms.txt
> Use this file to discover all available pages before exploring further.

# Search method comparison

> Compare global, local, and DRIFT search methods to choose the right approach for your queries

GraphRAG offers three distinct search methods, each optimized for different types of queries. This guide helps you understand when to use each method and how they compare.

## Search methods overview

<CardGroup cols={3}>
  <Card title="Global search" icon="globe">
    **Map-reduce over community reports**

    Best for high-level, dataset-wide questions and thematic analysis.
  </Card>

  <Card title="Local search" icon="magnifying-glass">
    **Entity-centric retrieval**

    Best for specific entity queries and detailed fact retrieval.
  </Card>

  <Card title="DRIFT search" icon="route">
    **Iterative graph traversal**

    Best for complex multi-hop reasoning and exploratory analysis.
  </Card>
</CardGroup>

## Quick comparison table

| Aspect               | Global Search                | Local Search              | DRIFT Search            |
| -------------------- | ---------------------------- | ------------------------- | ----------------------- |
| **Question Type**    | Broad, thematic              | Specific, factual         | Complex, multi-hop      |
| **Data Source**      | Community reports            | Entities + text + reports | Dynamic graph traversal |
| **Context Building** | Fixed (all reports at level) | Semantic retrieval        | Iterative expansion     |
| **LLM Calls**        | Many (map-reduce)            | Single                    | Multiple (iterative)    |
| **Cost**             | High                         | Low-Medium                | Medium-High             |
| **Response Time**    | Slow (2-10s)                 | Fast (\<2s)               | Medium (2-5s)           |
| **Token Usage**      | High                         | Low-Medium                | Medium-High             |
| **Coverage**         | Entire dataset               | Focused                   | Adaptive                |
| **Depth**            | Summarized                   | Detailed                  | Multi-level             |

## When to use each method

### Global search

<Tabs>
  <Tab title="Ideal for">
    * "What are the main themes in this dataset?"
    * "What are the top trends across all documents?"
    * "Summarize the key findings"
    * "What are the most significant events?"
    * "How do different topics relate globally?"
  </Tab>

  <Tab title="Strengths">
    * Comprehensive dataset coverage
    * Identifies high-level patterns
    * Synthesizes information across communities
    * Good for exploratory analysis
  </Tab>

  <Tab title="Limitations">
    * Expensive (many LLM calls)
    * Slower response times
    * Less detailed on specific entities
    * May miss fine-grained details
  </Tab>

  <Tab title="Example">
    ```bash theme={null}
    graphrag query \
      "What are the most important themes in this story?" \
      --method global
    ```
  </Tab>
</Tabs>

### Local search

<Tabs>
  <Tab title="Ideal for">
    * "Who is Dr. Jordan Hayes?"
    * "What is the relationship between X and Y?"
    * "What are the properties of this entity?"
    * "When did this specific event occur?"
    * "What does the document say about X?"
  </Tab>

  <Tab title="Strengths">
    * Fast response times
    * Cost-effective
    * Detailed entity information
    * Includes original text evidence
    * Good for fact-checking
  </Tab>

  <Tab title="Limitations">
    * Limited to nearby context
    * May miss global patterns
    * Less comprehensive for broad questions
    * Depends on entity retrieval quality
  </Tab>

  <Tab title="Example">
    ```bash theme={null}
    graphrag query \
      "Who is Scrooge and what are his relationships?" \
      --method local
    ```
  </Tab>
</Tabs>

### DRIFT search

<Tabs>
  <Tab title="Ideal for">
    * "How do organizations influence these events?"
    * "What chain of events connects A to B?"
    * "What patterns emerge from analyzing X, Y, and Z together?"
    * "How are these entities indirectly connected?"
    * "What are the multi-step dependencies?"
  </Tab>

  <Tab title="Strengths">
    * Handles complex reasoning
    * Multi-hop relationship traversal
    * Adaptive context gathering
    * Balances breadth and depth
    * Discovers indirect connections
  </Tab>

  <Tab title="Limitations">
    * Moderate to high cost
    * Slower than local search
    * Complexity in parameter tuning
    * May be overkill for simple queries
  </Tab>

  <Tab title="Example">
    ```python theme={null}
    # Currently via API/notebooks only
    result = await drift_search.search(
        "How do the different factions interact through intermediaries?"
    )
    ```
  </Tab>
</Tabs>

## Side-by-side examples

Here's how each method handles the same dataset but different query types:

### Dataset: "Operation Dulce" (sci-fi narrative)

<AccordionGroup>
  <Accordion title="Query 1: Broad thematic question">
    **Question**: "What are the main themes in this story?"

    <Tabs>
      <Tab title="Global Search (Best)">
        **Response**: Analyzes all community reports to identify overarching themes like government secrecy, alien contact, scientific ethics, and human cooperation.

        **Why it works**: Synthesizes information across the entire narrative structure.

        **Cost**: \~15,000 tokens
      </Tab>

      <Tab title="Local Search (Suboptimal)">
        **Response**: May identify themes related to specific entities it retrieves, but lacks comprehensive coverage.

        **Why it's limited**: Only sees themes connected to retrieved entities.

        **Cost**: \~3,000 tokens
      </Tab>

      <Tab title="DRIFT Search (Moderate)">
        **Response**: Explores themes through entity relationships but may not cover all communities.

        **Why it's moderate**: Can discover themes but less systematic than global.

        **Cost**: \~8,000 tokens
      </Tab>
    </Tabs>
  </Accordion>

  <Accordion title="Query 2: Specific entity question">
    **Question**: "Who is Agent Mercer and what is their role?"

    <Tabs>
      <Tab title="Local Search (Best)">
        **Response**: Retrieves Agent Mercer entity, related relationships, and relevant text chunks providing detailed information about their role, background, and actions.

        **Why it works**: Direct entity retrieval with supporting evidence.

        **Cost**: \~2,500 tokens
      </Tab>

      <Tab title="Global Search (Suboptimal)">
        **Response**: May mention Agent Mercer in community summaries but lacks specific details.

        **Why it's limited**: Community reports are high-level summaries.

        **Cost**: \~12,000 tokens
      </Tab>

      <Tab title="DRIFT Search (Good)">
        **Response**: Finds Agent Mercer and explores connected entities, providing both direct info and contextual relationships.

        **Why it's good**: Comprehensive but potentially unnecessary depth.

        **Cost**: \~5,000 tokens
      </Tab>
    </Tabs>
  </Accordion>

  <Accordion title="Query 3: Multi-hop reasoning">
    **Question**: "How do the government organizations indirectly influence the scientific research at the base?"

    <Tabs>
      <Tab title="DRIFT Search (Best)">
        **Response**: Traces connections from government entities through intermediary actors to research activities, revealing multi-step influence patterns.

        **Why it works**: Designed for graph traversal and multi-hop reasoning.

        **Cost**: \~7,000 tokens
      </Tab>

      <Tab title="Local Search (Limited)">
        **Response**: Shows direct connections between government and research but may miss indirect paths.

        **Why it's limited**: Doesn't traverse multiple hops effectively.

        **Cost**: \~3,500 tokens
      </Tab>

      <Tab title="Global Search (Moderate)">
        **Response**: Identifies that government influences research from community summaries but lacks specific pathways.

        **Why it's moderate**: Sees the pattern but not the mechanism.

        **Cost**: \~13,000 tokens
      </Tab>
    </Tabs>
  </Accordion>

  <Accordion title="Query 4: Relationship query">
    **Question**: "What is the relationship between Dr. Hayes and the Dulce facility?"

    <Tabs>
      <Tab title="Local Search (Best)">
        **Response**: Retrieves both entities and their connecting relationships with supporting text evidence.

        **Why it works**: Optimized for entity relationship queries.

        **Cost**: \~2,800 tokens
      </Tab>

      <Tab title="Global Search (Limited)">
        **Response**: May note connection in community summary but without specifics.

        **Why it's limited**: Relationships are abstracted in summaries.

        **Cost**: \~12,000 tokens
      </Tab>

      <Tab title="DRIFT Search (Good)">
        **Response**: Finds relationship and explores broader context around both entities.

        **Why it's good**: Comprehensive but may include extra context.

        **Cost**: \~4,500 tokens
      </Tab>
    </Tabs>
  </Accordion>
</AccordionGroup>

## Decision flowchart

<Steps>
  <Step title="Is your question about a specific entity or relationship?">
    **YES** → Use **Local Search**

    Examples:

    * "Who is X?"
    * "What is X's relationship to Y?"
    * "What are X's properties?"

    **NO** → Continue to next step
  </Step>

  <Step title="Does your question require understanding the entire dataset?">
    **YES** → Use **Global Search**

    Examples:

    * "What are the main themes?"
    * "What are the key trends?"
    * "Summarize this dataset"

    **NO** → Continue to next step
  </Step>

  <Step title="Does your question involve multi-hop reasoning or complex connections?">
    **YES** → Use **DRIFT Search**

    Examples:

    * "How does X indirectly affect Y?"
    * "What chain of events connects A to B?"
    * "What patterns involve X, Y, and Z?"

    **NO** → Try Local Search first, then escalate if needed
  </Step>
</Steps>

## Hybrid approaches

You can combine search methods for comprehensive analysis:

### Sequential querying

<Steps>
  <Step title="Start with global search">
    Get high-level themes and important entities:

    ```bash theme={null}
    graphrag query "What are the main entities in this dataset?" --method global
    ```
  </Step>

  <Step title="Follow up with local search">
    Deep dive into specific entities identified:

    ```bash theme={null}
    graphrag query "Tell me more about [entity from global results]" --method local
    ```
  </Step>

  <Step title="Use DRIFT for connections">
    Explore relationships between key entities:

    ```python theme={null}
    result = await drift_search.search(
        "How do [entity1] and [entity2] influence each other?"
    )
    ```
  </Step>
</Steps>

### Validation strategy

```python theme={null}
# Use multiple methods to validate findings

# Global: Identify main themes
global_result = await global_search.search(
    "What are the key themes?"
)

# Local: Verify themes with specific evidence
local_result = await local_search.search(
    "What evidence supports the theme of [X]?"
)

# DRIFT: Explore how themes connect
drift_result = await drift_search.search(
    "How do themes [X] and [Y] relate through entities?"
)
```

## Performance benchmarks

Based on typical "Operation Dulce" dataset queries:

| Metric              | Global | Local  | DRIFT     |
| ------------------- | ------ | ------ | --------- |
| Avg Response Time   | 8.5s   | 1.2s   | 3.8s      |
| Avg Tokens (Prompt) | 11,500 | 2,800  | 6,200     |
| Avg Tokens (Output) | 800    | 400    | 600       |
| Avg Cost (GPT-4)    | \$0.12 | \$0.03 | \$0.07    |
| Parallelizable      | Yes    | No     | Partially |

<Note>
  Actual performance varies based on dataset size, query complexity, and configuration parameters.
</Note>

## Configuration comparison

<Tabs>
  <Tab title="Global Search">
    ```python theme={null}
    # Key parameters
    context_builder_params = {
        "max_tokens": 12_000,
        "community_level": 2,  # 0=coarse, 3+=fine
        "use_community_summary": False,  # Full vs summary
    }

    map_llm_params = {
        "max_tokens": 1000,
        "temperature": 0.0,
    }

    reduce_llm_params = {
        "max_tokens": 2000,
        "temperature": 0.0,
    }
    ```
  </Tab>

  <Tab title="Local Search">
    ```python theme={null}
    # Key parameters
    local_context_params = {
        "text_unit_prop": 0.5,  # Text chunk allocation
        "community_prop": 0.1,  # Community report allocation
        "top_k_mapped_entities": 10,  # Entity retrieval count
        "top_k_relationships": 10,  # Relationship count
        "max_tokens": 12_000,
    }

    model_params = {
        "max_tokens": 2_000,
        "temperature": 0.0,
    }
    ```
  </Tab>

  <Tab title="DRIFT Search">
    ```python theme={null}
    # Key parameters
    drift_params = DRIFTSearchConfig(
        primer_folds=1,  # Initial retrieval rounds
        drift_k_followups=3,  # Expansions per iteration
        n_depth=3,  # Max traversal depth
    )
    ```
  </Tab>
</Tabs>

## Cost optimization strategies

<CardGroup cols={2}>
  <Card title="Use local search first" icon="dollar-sign">
    Start with low-cost local search; escalate only if needed
  </Card>

  <Card title="Adjust community level" icon="layer-group">
    Use level 1 for global search instead of level 2 to reduce tokens
  </Card>

  <Card title="Tune DRIFT conservatively" icon="sliders">
    Keep `n_depth=2` and `drift_k_followups=2` for most queries
  </Card>

  <Card title="Use summaries" icon="compress">
    Set `use_community_summary=True` in global search
  </Card>
</CardGroup>

## Common pitfalls

<AccordionGroup>
  <Accordion title="Using global search for specific entities">
    **Problem**: Expensive and provides less detail than local search.

    **Solution**: Use local search for entity-specific queries.
  </Accordion>

  <Accordion title="Using local search for broad themes">
    **Problem**: May miss important information not connected to retrieved entities.

    **Solution**: Use global search for dataset-wide questions.
  </Accordion>

  <Accordion title="Over-parameterizing DRIFT">
    **Problem**: Setting `n_depth=5` and `drift_k_followups=10` wastes tokens.

    **Solution**: Start with default parameters and increase only if needed.
  </Accordion>

  <Accordion title="Not checking context data">
    **Problem**: Accepting answers without verifying supporting evidence.

    **Solution**: Always inspect `result.context_data` to see what was used.
  </Accordion>
</AccordionGroup>

## Choosing based on use case

<Tabs>
  <Tab title="Research & analysis">
    **Primary**: Global search for themes and patterns

    **Secondary**: Local search for validating specific claims

    **Tertiary**: DRIFT search for exploring connections between findings
  </Tab>

  <Tab title="Q&A systems">
    **Primary**: Local search for most user questions

    **Secondary**: DRIFT search for complex follow-ups

    **Tertiary**: Global search for summarization requests
  </Tab>

  <Tab title="Investigation">
    **Primary**: DRIFT search for tracing connections

    **Secondary**: Local search for entity details

    **Tertiary**: Global search for context and patterns
  </Tab>

  <Tab title="Content summarization">
    **Primary**: Global search for high-level summaries

    **Secondary**: Local search for specific section details

    **Tertiary**: DRIFT search for understanding structure
  </Tab>
</Tabs>

## Next steps

<CardGroup cols={2}>
  <Card title="Global search" icon="globe" href="/examples/notebooks/global-search">
    Deep dive into global search
  </Card>

  <Card title="Local search" icon="magnifying-glass" href="/examples/notebooks/local-search">
    Master local search techniques
  </Card>

  <Card title="DRIFT search" icon="route" href="/examples/notebooks/drift-search">
    Learn DRIFT search methods
  </Card>

  <Card title="Query overview" icon="book" href="/query/overview">
    Complete query documentation
  </Card>
</CardGroup>