CREATE_KNOWLEDGE_GRAPH

Use the DBMS_AI_GRAPHRAG.CREATE_KNOWLEDGE_GRAPH procedure to create a knowledge graph from a source and configures graph extraction and retrieval.

Purpose

This procedure processes the source text, extracts graph components, and creates the data used for retrieval.

• Graph creation processes source documents by chunking them, extracting entities and relationships, constructing the graph, and creating embeddings and vector indexes.
• For a textual graph, the procedure also generates descriptions or summaries for graph components. A non-textual graph does not include these generated descriptions.
• If DML maintenance is enabled, the procedure sets up asynchronous processing for source-data changes. It processes changed documents and updates affected graph data.

Syntax

DBMS_AI_GRAPHRAG.CREATE_KNOWLEDGE_GRAPH (
    kg_name               IN VARCHAR2,
    source_name           IN VARCHAR2,
    source_column         IN VARCHAR2,
    graph_type            IN VARCHAR2,
    params                IN JSON,
    dml_maintenance       IN BOOLEAN DEFAULT FALSE,
    model_failure_policy  IN VARCHAR2 DEFAULT 'ABORT',
    resume                IN BOOLEAN DEFAULT FALSE
);

Parameters

Parameter	Description
kg_name	Name of the knowledge graph.
source_name	Name of the source. The source can be a table or a view. DML maintenance is supported only for tables; for views, dml_maintenance must be FALSE.
source_column	Name of the column that contains the source text.
graph_type	Graph type. Valid values are TEXTUAL and NON_TEXTUAL. The default is NON_TEXTUAL. Textual graphs include generated descriptions for graph components; non-textual graphs do not.
params	JSON object that configures LLM generation, embeddings, chunking, vector indexing, and related options.
dml_maintenance	Enables DML-based graph maintenance. The default is FALSE.
model_failure_policy	Specifies how to handle model call failures during graph creation: - 'ABORT' (default): Stops graph creation and reports an error when a model call fails after applicable retries. Previously completed work is retained. After resolving the cause, call CREATE_KNOWLEDGE_GRAPH with resume => TRUE to resume creation. - 'CONTINUE': Records the failure and continues processing other graph items.
resume	Specifies whether to resume a previous knowledge graph creation attempt (TRUE) or create a new knowledge graph (FALSE, the default). The operation fails if a graph with the specified name already exists.

The params object supports the following fields:

Field	Description
llm_generate_params	(Required) LLM configuration for extraction and summarization, using the UTL_TO_GENERATE_TEXT parameter format.
embedding_params	(Required) Embedding configuration, using the UTL_TO_EMBEDDINGS parameter format.
chunk_params	(Optional) Chunking configuration, using the UTL_TO_CHUNKS parameter format.
vector_index_params	(Optional) Vector index configuration. If omitted, HNSW is used with default parameters.
language	(Optional) Language value included in the extraction prompt.
prompt_language	(Optional) Language for the prompt.
entity_type_list	(Optional) List of entity types to guide extraction.
llm_request_params	(Optional) LLM request settings: - max_concurrent_requests: Maximum parallel LLM calls. Default: 15. - max_retries: Maximum number of retries for failed LLM calls. Default: 5. - rate_per_minute: Maximum requests per minute. Default: NULL. - use_llm_cache: Whether to save outputs to an LLM cache table for later use. Default: FALSE.
dml_repeat_interval	(Optional) Interval for the asynchronous DML worker.
summary_character_limit	(Optional) Character limit for summaries. The default is 2000. For textual graphs, content that exceeds the limit is summarized with an LLM.
summary_prompt_character_limit	(Optional) Character limit for the prompt sent to the LLM for summarization. The default is 2000.

Example

This example creates a textual knowledge graph from the BOOKS table using OCI Generative AI for text generation and embeddings.

BEGIN
    DBMS_AI_GRAPHRAG.CREATE_KNOWLEDGE_GRAPH(
        kg_name       => 'BOOKS_KG',
        source_name   => 'BOOKS',
        source_column => 'CONTENT',
        graph_type    => 'TEXTUAL',
        params        => JSON('{
            "chunk_params": {
                "by": "words",
                "max": "300",
                "overlap": "0",
                "split": "recursively",
                "normalize": "all"
            },
            "llm_generate_params": {
                "provider": "ocigenai",
                "credential_name": "OCI_CRED",
                "url": "https://inference.generativeai.us-chicago-1.oci.oraclecloud.com/20231130/actions/chat",
                "model": "meta.llama-3.3-70b-instruct"
            },
            "embedding_params": {
                "provider": "ocigenai",
                "credential_name": "OCI_CRED",
                "url": "https://inference.generativeai.us-chicago-1.oci.oraclecloud.com/20231130/actions/embedText",
                "model": "cohere.embed-v4.0"
            }
        }')
    );
END;
/