RAG SQL-PolarDB(PolarDB)-阿里云帮助中心

This topic describes the SQL functions for Retrieval-Augmented Generation (RAG) systems.

polar_ai.ai_rag_chunking

Splits the binary content of a document into Chunks.

Syntax

polar_ai.ai_rag_chunking(
    content BYTEA,
    doctype TEXT,
    deploymentName TEXT DEFAULT 'RAG',
    saveType TEXT DEFAULT 'binary',
    chunkoptions JSONB DEFAULT NULL
)
RETURNS SETOF record (chunk_id INTEGER, chunk_content TEXT)

Parameters

Parameter	Type	Required	Default	Description
`content`	`BYTEA`	Yes	-	The binary content of the document.
`doctype`	`TEXT`	Yes	-	Document types, such as `pdf`, `docx`, and `xlsx`.
`deploymentName`	`TEXT`	No	`RAG`	The deployment name of the RAG service.
`saveType`	`TEXT`	No	`binary`	Document type, such as `txt`, `url`, or `binary`.
`chunkoptions`	`JSONB`	No	`NULL`	Chunking options, provided in JSON format. For example, `'{ "chunkSize": 512, "chunkOverlap": 64 }'`. `chunkSize`: The maximum length of each chunk in bytes. The default value is `128`. `chunkOverlap`: The length of the overlap between adjacent chunks, in bytes. `separator`: The separator used to split text. The default separator is a space. `headersToSplitOn`: Chunks Markdown files based on heading levels. Valid values: an integer from -1 to 6. Default value: -1. If the value is -1, the file is split by all heading levels. For other values, such as `x`, the file is split into chunks based on headings with a level less than or equal to `x`.

Examples

-- Chunk with default options
SELECT * FROM polar_ai.ai_rag_chunking(content, 'pdf');

-- Customize chunkSize and chunkOverlap
SELECT * FROM polar_ai.ai_rag_chunking(content, 'pdf', chunkoptions := '{"chunkSize": 512, "chunkOverlap": 64}');

polar_ai.ai_rag_text_embedding

Converts text content into a Vector.

Note

You must first call polar_ai.ai_nl2sql_deployModel('RAG'); to deploy the model.

Syntax

polar_ai.ai_rag_text_embedding(
    content TEXT,
    model TEXT DEFAULT 'Qwen-8B-embedding',
    dimension INTEGER DEFAULT 1024,
    deploymentName TEXT DEFAULT 'RAG'
)
RETURNS float4[]

Parameters

Parameter	Type	Required	Default	Description
`content`	`TEXT`	Yes	-	The text content to be converted into a Vector.
`model`	`TEXT`	No	`Qwen-8B-embedding`	The embedding model to use. Valid values: `Qwen-8B-embedding` `GTE`
`dimension`	`INTEGER`	No	`1024`	The dimension of the generated vector must be strictly consistent with the dimension defined in the `VECTOR` column of the data table. A mismatch in dimensions will cause insertion or query errors. `Qwen-8B-embedding`: 2048, 1536, 1024 (default), 768, 512, 256, 128, 64. `GTE`: 768
`deploymentName`	`TEXT`	No	`RAG`	The deployment name of the RAG service.

Examples

-- Generate a 1024-dimensional vector
select polar_ai.ai_rag_text_embedding('It is a lovely film with lovely performances.');
-- Result:
{0.15527344,-0.09667969,-0.08984375,-0.13476562,0.087890625,-0.09423828, ....}

polar_ai.ai_rag_rerank

Reranks a list of retrieved Text Chunks based on their relevance to a question.

Note

You must first call polar_ai.ai_nl2sql_deployModel('RAG'); to deploy the model.

Syntax

polar_ai.ai_rag_rerank(
    question TEXT,
    contents TEXT[],
    model TEXT DEFAULT 'gte',
    deploymentName TEXT DEFAULT 'RAG'
)
RETURNS TEXT[]

Parameters

Parameter	Type	Required	Default	Description
`question`	`TEXT`	Yes	-	The original question.
`contents`	`TEXT[]`	Yes	-	An Array of Text Chunks to rerank.
`model`	`TEXT`	No	`gte`	The reranking model to use. Valid values: gte Qwen3-Reranker
`deploymentName`	`TEXT`	No	`RAG`	The deployment name of the RAG service.

Examples

-- Rerank an array of text
SELECT polar_ai.ai_rag_rerank(
    'What fruit do you like to eat?',
    ARRAY['Apple', 'Crystal', 'Soccer', 'Banana']
);
-- Result:
{'Apple','Banana','Crystal','Soccer'}