Create a Knowledge Base - 阿里云百炼 Agent Studio

A knowledge base is the core entity of the RAG platform, organizing a set of documents, chunks, embeddings, and index configurations to provide retrieval and question-answering capabilities.

Knowledge Base Types

You must select a knowledge base type during creation. Each type corresponds to different data formats and retrieval methods:

Type	Applicable Data	Retrieval Method	Description
Document Search	PDF / Word / Markdown / HTML / Excel	Hybrid Vector + Keyword	Builds a mixed-index for documents, files, images, and Excel sheets
Data Query	CSV / Excel / RDS Tables	Natural Language to SQL (NL2SQL)	Queries structured tables based on schema; supports NL2SQL for headers and column names
Image Q&A	Product Images / Design Mockups	Multimodal Embedding	Primarily indexes images; supports image search and multimodal (image + text) Q&A
Audio/Video Search	Screen Recordings / Training Videos	Transcription + Segment Localization	Integrates understanding and intelligent summarization of audio/video content

Use Cases

When selecting Document Search as the knowledge base type, you must further specify a use case. Different use cases determine distinct retrieval and response-generation strategies:

Use Case	Description
Basic Document Q&A	Retrieves document chunks and directly generates answers using the LLM
Rich-Media Responses	Automatically formats responses with layout and relevant images, generating rich-media replies
Visual Understanding (Rich Text Documents)	No chunking required; the LLM directly interprets visually rich documents and images
Ultra-Fast Q&A	Optimized for highly structured or simple documents, delivering ultra-low-latency Q&A

Other knowledge base types—Data Query, Image Q&A, and Audio/Video Search—do not offer use case options.

Knowledge Base Type and Use Case Selection

Above: Select the knowledge base type. When Document Search is selected, four use cases appear below. For other types, no use case selection is required.

Console-Based Creation

Navigate to Knowledge Management

Log in to the console and go to Data Ingestion → Knowledge Management, then click Create Knowledge Base in the top-right corner.

Enter Basic Information

Field	Required	Description
Name	Yes	1–20 characters; unique within the business space
Description	Yes	Up to 200 characters; aids filtering
Knowledge Base Type	Yes	Document Search / Data Query / Image Q&A / Audio/Video Search
Use Case	Yes	Basic Document Q&A / Rich-Media Responses / Visual Understanding / Ultra-Fast Q&A

Import Data

Choose a data source:

Data Source	Description
Upload Files	Upload files directly via the UI; up to 50 files per batch, each ≤ 150 MB
Select Category	Import from connector file categories; supports automatic synchronization
Select Files	Choose specific files under a connector category for import

Supported formats: PDF / DOCX / XLSX / PPTX / TXT / MD / HTML / CSV, etc.

Choose Parsing Method

By default, Default Settings are used—the system automatically selects an appropriate parser based on file type, requiring no manual adjustment in most scenarios.To customize parsing per format, click Custom Settings. Available parsing methods include:

Parsing Method	Description	Applicable Scenarios
Electronic Document Parsing	Standard text extraction	Well-formatted electronic documents
Intelligent Document Parsing	Layout-level OCR + structural recovery	Scanned documents, complex layouts
LLM-based Document Parsing	Uses large language models to understand document structure	Non-standard or irregular document formats
Qwen-VL Parsing	Vision-language model	Documents with mixed text and images
Audio/Video Parsing	Speech-to-text transcription + timestamped segment localization	Audio/video files

Configure Index Settings

Parameter	Default Value	Description
Chunking Method	Smart Chunking	Alternatives: by length / by page / by heading / by regex / by delimiter
Maximum Chunk Length	600	Range: 10–6000
Embedding Model	text-embedding-v4	Semantic vector model supporting both Chinese and English
Vector Storage	Platform Storage	Alternative: self-managed ADB-PG engine

See Chunking and Vectorization for details.

Retrieval parameters—including ranking model, TopK, and similarity threshold—are configured in the retrieval service. See Knowledge Retrieval for details.

Complete Creation

Click Create Knowledge Base. The system automatically performs parsing, chunking, vectorization, and index building. Once the knowledge base status changes to Ready, it becomes available for retrieval. Processing progress can be monitored in the document list.

API-Based Creation

Use the Create Knowledge Base and Import Data API to create a knowledge base and submit an import task in one step:

curl -X POST https://{workspaceId}.cn-beijing.maas.aliyuncs.com/api/v1/indices/rag/index/create_v2 \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-knowledge-base",
    "structureType": "unstructured",
    "sinkType": "BUILT_IN",
    "sourceType": "DATA_CENTER_CATEGORY",
    "embeddingModelName": "text-embedding-v4",
    "chunkSize": 600,
    "source_ids": ["cate_xxx"]
  }'

Chunking strategies can be adjusted after creation, but changing the embedding model triggers full index reconstruction. For complete capacity limits, see Capacity and Limits.

After creating your knowledge base, proceed to Upload Documents.

​Knowledge Base Types

​Use Cases

​Console-Based Creation

​API-Based Creation

Knowledge Base Types

Use Cases

Console-Based Creation

API-Based Creation