Databricks: see which chunks support each ai_search answer

The ai_search function now includes a generate_citations option, enabling you to track which retrieved documents supported the answer synthesized by the model.

Use cases

Imagine you built a RAG workflow using vector search indexes or synced tables and need to audit the accuracy of the generated answers.

• Track the exact document chunks used in AI-generated answers.
• Audit and validate responses in batch RAG pipelines.
• Expose contextual searches to compound AI systems.

How to use

Run the ai_search function in the SQL editor, passing the citation generation option alongside your knowledge source.

SELECT ai_search(
  'How do I configure auto-scaling for my SQL warehouse?',
  PARSE_JSON('[{
    "type": "vector_search",
    "config": {
      "index_name": "prod_catalog.docs.support_articles",
      "text_col": "article_body",
      "doc_uri_col": "article_url"
    }
  }]'),
  options => map('generate_answer', 'true', 'generate_citations', 'true')
) AS result;

What happens in this code:

1. The function performs a semantic search on the specified vector index using the natural language query.
2. The model synthesizes an answer based on the retrieved documents.
3. The generate_citations option returns the index of the specific chunks used to ground the generated text.

What it returns

Field	Description
document	Array of retrieved documents ordered by relevance.
document[].page_content	Text content of the retrieved chunk.
document[].doc_uri	URI of the source document.
document[].metadata	Additional metadata from the index.
answer	Grounded natural-language answer synthesized from the retrieved documents.
citations	Array containing the 0-based position (document_index) of the chunks used by the model to support the answer.

Limitations and requirements

• This feature is in Beta.
• Requires Databricks Runtime 18.2 or above.
• The generate_citations option requires generate_answer to be set to true.
• Requires serverless environment version 3 or above when using serverless compute.

Using it in a pipeline

This function fits into the consumption layer or RAG pipelines to enrich operational data, allowing AI-powered applications to return both the synthesized answer and the corresponding source evidence.