開発者向けドキュメント > 開発者向け学習 > Python API client user guide > 生成AI > Vector databases

ベクターデータベース¶

Vector databases enable RAG (Retrieval-Augmented Generation) workflows by storing document embeddings and retrieving relevant context for LLM prompts. Vector databases allow you to create knowledge bases from your documents and use them to provide context-aware responses in your generative AI applications.

Validate a deployment as a vector database¶

Before using a deployment as a vector database, validate it using datarobot.CustomModelVectorDatabaseValidation:

import datarobot as dr
validation = dr.CustomModelVectorDatabaseValidation.create(
    deployment_id=deployment.id,
    name="My Vector Database",
    prompt_column_name="query",
    target_column_name="citations",
    wait_for_completion=True
)
if validation.validation_status == "PASSED":
    print("Vector database validation passed!")

Get supported embeddings¶

View available embedding models using datarobot.VectorDatabase.get_supported_embeddings():

supported = dr.genai.VectorDatabase.get_supported_embeddings()
print(f"Default embedding model: {supported.default_embedding_model}")
for model in supported.embedding_models:
    print(f"  {model.name}: {model.description}")

You can also get recommended embeddings for a specific dataset:

dataset = dr.Dataset.get(dataset_id)
supported = dr.genai.VectorDatabase.get_supported_embeddings(dataset_id=dataset.id)
print(f"Default embedding: {supported.default_embedding_model}")

Get supported text chunking configurations¶

To view available text chunking options:

chunking_configs = dr.genai.VectorDatabase.get_supported_text_chunkings()
for config in chunking_configs.text_chunking_configs:
    print(f"Chunking config: {config}")

ベクターデータベースの作成¶

Create a vector database from a dataset containing your documents. When creating a vector database, you should specify the following:

dataset_id: The ID of the dataset used for creation.
chunking_parameters: Parameters defining how documents are split and embedded, including embedding model, chunking method, chunk size, and overlap percentage.
name: An optional user-friendly name for the vector database.
use_case: An optional Use Case to link the vector database to.

dataset = dr.Dataset.upload("documents.csv")
supported = dr.genai.VectorDatabase.get_supported_embeddings(dataset_id=dataset.id)
from datarobot.models.genai.vector_database import ChunkingParameters
chunking_params = ChunkingParameters(
    embedding_model=supported.default_embedding_model,
    chunking_method="semantic",
    chunk_size=500,
    chunk_overlap_percentage=10
)
vector_db = dr.genai.VectorDatabase.create(
    dataset_id=dataset.id,
    name="Document Knowledge Base",
    chunking_parameters=chunking_params,
    use_case=use_case_id
)
vector_db

Update a vector database¶

Add more documents or update an existing vector database:

new_dataset = dr.Dataset.upload("updated_documents.csv")
updated_vector_db = dr.genai.VectorDatabase.create(
    dataset_id=new_dataset.id,
    parent_vector_database_id=vector_db.id,
    update_llm_blueprints=True
)

Link vector database to LLM blueprint¶

Associate a vector database with an LLM blueprint for RAG:

vector_db = dr.genai.VectorDatabase.get(vector_db_id)
blueprint = dr.genai.LLMBlueprint.get(blueprint_id)
blueprint.update(
    vector_database=vector_db.id,
    vector_database_settings={
        "max_documents_retrieved_per_prompt": 3
    }
)

ベクターデータベースの一覧表示と管理¶

List all vector databases:

all_dbs = dr.genai.VectorDatabase.list()
print(f"Found {len(all_dbs)} vector database(s):")
for db in all_dbs:
    print(f"  - {db.name} (ID: {db.id}, Status: {db.execution_status})")

Filter vector databases by Use Case:

use_case = dr.UseCase.get(use_case_id)
use_case_dbs = dr.genai.VectorDatabase.list(use_case=use_case)

Get vector database details:

vector_db = dr.genai.VectorDatabase.get(vector_db_id)
print(f"Name: {vector_db.name}")
print(f"Size: {vector_db.size} bytes")
print(f"Status: {vector_db.execution_status}")
print(f"Chunks count: {vector_db.chunks_count}")

Delete a vector database:

vector_db.delete()

Export a vector database as a dataset¶

To export a vector database as a dataset:

vector_db = dr.genai.VectorDatabase.get(vector_db_id)
export_job = vector_db.submit_export_dataset_job()
exported_dataset_id = export_job.dataset_id