Skip to main content
The GraphRAG configuration schema defines all available settings for the indexing pipeline and search operations. Configuration can be provided via YAML or JSON files.

​
Root configuration

The root GraphRagConfig class contains all top-level configuration settings.
​
completion_models
dict[str, ModelConfig]
default:"{}"
Available completion model configurations. Maps model IDs to their respective configurations.
​
embedding_models
dict[str, ModelConfig]
default:"{}"
Available embedding model configurations. Maps model IDs to their respective configurations.
​
concurrent_requests
int
default:"25"
The default number of concurrent requests to make to language models.
​
async_mode
AsyncType
default:"threaded"
The default asynchronous mode to use for language model requests. See AsyncType enum.
​
input
InputConfig
default:"InputConfig()"
The input configuration for document sources.
​
input_storage
StorageConfig
default:"StorageConfig(base_dir='input')"
The input storage configuration. For file storage, base_dir defaults to input.
​
chunking
ChunkingConfig
The chunking configuration to use. See chunking configuration.
​
output_storage
StorageConfig
default:"StorageConfig(base_dir='output')"
The output storage configuration. For file storage, base_dir defaults to output.
​
update_output_storage
StorageConfig
default:"StorageConfig(base_dir='update_output')"
The output configuration for the updated index. For file storage, base_dir defaults to update_output.
​
table_provider
TableProviderConfig
default:"TableProviderConfig()"
The table provider configuration. By default, parquet files are read/written to disk. You can register custom output table storage.
​
cache
CacheConfig
The cache configuration for storing LLM responses and intermediate results.
​
reporting
ReportingConfig
default:"ReportingConfig()"
The reporting configuration. See reporting configuration.
​
vector_store
VectorStoreConfig
default:"VectorStoreConfig()"
The vector store configuration. Defaults to LanceDB with db_uri set to output/lancedb.
​
workflows
list[str] | None
default:"None"
List of workflows to run, in execution order. This always overrides any built-in workflow methods.

​
Indexing configuration

​
Text embedding

​
embed_text
EmbedTextConfig
default:"EmbedTextConfig()"
Text embedding configuration.Fields:
  • embedding_model_id (str): The model ID to use for text embeddings. Default: "default_embedding_model"
  • model_instance_name (str): The model singleton instance name. Default: "text_embedding"
  • batch_size (int): The batch size to use. Default: 16
  • batch_max_tokens (int): The batch max tokens to use. Default: 8191
  • names (list[str]): The specific embeddings to perform.

​
Graph extraction

​
extract_graph
ExtractGraphConfig
default:"ExtractGraphConfig()"
The entity extraction configuration to use.Fields:
  • completion_model_id (str): The model ID to use. Default: "default_completion_model"
  • model_instance_name (str): The model singleton instance name. Default: "extract_graph"
  • prompt (str | None): The entity extraction prompt to use. Default: None
  • entity_types (list[str]): The entity extraction entity types to use. Default: ["organization", "person", "geo", "event"]
  • max_gleanings (int): The maximum number of entity gleanings to use. Default: 1
​
extract_graph_nlp
ExtractGraphNLPConfig
default:"ExtractGraphNLPConfig()"
The NLP-based graph extraction configuration to use. Used for fast indexing mode.Fields:
  • normalize_edge_weights (bool): Whether to normalize edge weights. Default: True
  • text_analyzer (TextAnalyzerDefaults): Text analyzer configuration
  • concurrent_requests (int): Number of concurrent requests. Default: 25
  • async_mode (AsyncType): Async mode to use. Default: AsyncType.Threaded

​
Description summarization

​
summarize_descriptions
SummarizeDescriptionsConfig
default:"SummarizeDescriptionsConfig()"
The description summarization configuration to use.Fields:
  • prompt (str | None): The summarization prompt. Default: None
  • max_length (int): Maximum length in tokens. Default: 500
  • max_input_tokens (int): Maximum input tokens. Default: 4000
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • model_instance_name (str): Model instance name. Default: "summarize_descriptions"

​
Graph processing

​
prune_graph
PruneGraphConfig
default:"PruneGraphConfig()"
The graph pruning configuration to use.Fields:
  • min_node_freq (int): Minimum node frequency. Default: 2
  • max_node_freq_std (None): Maximum node frequency standard deviation. Default: None
  • min_node_degree (int): Minimum node degree. Default: 1
  • max_node_degree_std (None): Maximum node degree standard deviation. Default: None
  • min_edge_weight_pct (float): Minimum edge weight percentage. Default: 40.0
  • remove_ego_nodes (bool): Whether to remove ego nodes. Default: True
  • lcc_only (bool): Keep only largest connected component. Default: False
​
cluster_graph
ClusterGraphConfig
default:"ClusterGraphConfig()"
The cluster graph configuration to use.Fields:
  • max_cluster_size (int): The maximum cluster size to use. Default: 10
  • use_lcc (bool): Whether to use the largest connected component. Default: True
  • seed (int): The seed to use for the clustering. Default: 0xDEADBEEF

​
Claims extraction

​
extract_claims
ExtractClaimsConfig
default:"ExtractClaimsConfig(enabled=False)"
The claim extraction configuration to use.Fields:
  • enabled (bool): Whether claim extraction is enabled. Default: False
  • prompt (str | None): The extraction prompt. Default: None
  • description (str): Description of claims to extract. Default: "Any claims or facts that could be relevant to information discovery."
  • max_gleanings (int): Maximum number of gleanings. Default: 1
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • model_instance_name (str): Model instance name. Default: "extract_claims"

​
Community reports

​
community_reports
CommunityReportsConfig
default:"CommunityReportsConfig()"
The community reports configuration to use.Fields:
  • completion_model_id (str): The model ID to use. Default: "default_completion_model"
  • model_instance_name (str): The model instance name. Default: "community_reporting"
  • graph_prompt (str | None): Prompt for graph-based summarization. Default: None
  • text_prompt (str | None): Prompt for text-based summarization. Default: None
  • max_length (int): Maximum length in tokens. Default: 2000
  • max_input_length (int): Maximum input length in tokens. Default: 8000

​
Snapshots

​
snapshots
SnapshotsConfig
default:"SnapshotsConfig()"
The snapshots configuration to use.Fields:
  • embeddings (bool): Whether to save embedding snapshots. Default: False
  • graphml (bool): Whether to save GraphML snapshots. Default: False
  • raw_graph (bool): Whether to save raw graph snapshots. Default: False

​
Search configuration

​
local_search
LocalSearchConfig
default:"LocalSearchConfig()"
The local search configuration.Fields:
  • prompt (str | None): The local search prompt to use. Default: None
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • embedding_model_id (str): Model ID for embeddings. Default: "default_embedding_model"
  • text_unit_prop (float): The text unit proportion. Default: 0.5
  • community_prop (float): The community proportion. Default: 0.15
  • conversation_history_max_turns (int): Maximum conversation turns. Default: 5
  • top_k_entities (int): Top k mapped entities. Default: 10
  • top_k_relationships (int): Top k mapped relations. Default: 10
  • max_context_tokens (int): Maximum tokens. Default: 12000
​
global_search
GlobalSearchConfig
default:"GlobalSearchConfig()"
The global search configuration.Fields:
  • map_prompt (str | None): The global search mapper prompt. Default: None
  • reduce_prompt (str | None): The global search reducer prompt. Default: None
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • knowledge_prompt (str | None): The global search general prompt. Default: None
  • max_context_tokens (int): Maximum context size in tokens. Default: 12000
  • data_max_tokens (int): Data LLM maximum tokens. Default: 12000
  • map_max_length (int): Map LLM max response length in words. Default: 1000
  • reduce_max_length (int): Reduce LLM max response length in words. Default: 2000
  • dynamic_search_threshold (int): Rating threshold to include a community. Default: 1
  • dynamic_search_keep_parent (bool): Keep parent if child communities relevant. Default: False
  • dynamic_search_num_repeats (int): Number of times to rate same report. Default: 1
  • dynamic_search_use_summary (bool): Use community summary instead of full context. Default: False
  • dynamic_search_max_level (int): Maximum community hierarchy level. Default: 2
​
drift_search
DRIFTSearchConfig
default:"DRIFTSearchConfig()"
The DRIFT search configuration.Fields:
  • prompt (str | None): The DRIFT search prompt. Default: None
  • reduce_prompt (str | None): The reduce prompt. Default: None
  • data_max_tokens (int): Maximum data tokens. Default: 12000
  • reduce_max_tokens (None): Maximum reduce tokens. Default: None
  • reduce_temperature (float): Reduce temperature. Default: 0
  • reduce_max_completion_tokens (None): Max completion tokens. Default: None
  • concurrency (int): Concurrency level. Default: 32
  • drift_k_followups (int): Number of followup queries. Default: 20
  • primer_folds (int): Number of primer folds. Default: 5
  • primer_llm_max_tokens (int): Primer LLM max tokens. Default: 12000
  • n_depth (int): Search depth. Default: 3
  • Additional local search parameters for DRIFT’s local search component
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • embedding_model_id (str): Embedding model ID. Default: "default_embedding_model"
​
basic_search
BasicSearchConfig
default:"BasicSearchConfig()"
The basic search configuration.Fields:
  • prompt (None): The basic search prompt. Default: None
  • k (int): Number of results to return. Default: 10
  • max_context_tokens (int): Maximum context tokens. Default: 12000
  • completion_model_id (str): Model ID to use. Default: "default_completion_model"
  • embedding_model_id (str): Embedding model ID. Default: "default_embedding_model"

​
Helper methods

The GraphRagConfig class provides helper methods to retrieve model configurations:

​
get_completion_model_config

def get_completion_model_config(self, model_id: str) -> ModelConfig
Get a completion model configuration by ID. Parameters:
  • model_id (str): The ID of the model to get. Should match an ID in the completion_models list.
Returns:
  • ModelConfig: The model configuration if found.
Raises:
  • ValueError: If the model ID is not found in the configuration.

​
get_embedding_model_config

def get_embedding_model_config(self, model_id: str) -> ModelConfig
Get an embedding model configuration by ID. Parameters:
  • model_id (str): The ID of the model to get. Should match an ID in the embedding_models list.
Returns:
  • ModelConfig: The model configuration if found.
Raises:
  • ValueError: If the model ID is not found in the configuration.

Build docs developers (and LLMs) love

Get started for freeTalk to us