Retriever chunking and reranking

When you’re building a skill prompt that uses a retriever you can use chunking and reranking to enhance the accuracy and relevance of your responses.

Before you begin

Role required: sn_skill_builder.admin

About this task

To set up the chunking and reranking options for a retriever, you must have a retriever tool added to your skill with the Hybrid or Semantic search criteria. The following steps come after the semantic configuration in Add a retriever.

Procedure

1. Navigate to All > Now Assist Skill Kit > Home.
2. Select the skill that you’re adding a retriever to.
3. Add a retriever.
4. On the form, fill in the fields.

   Table 1. Chunking and reranking form

   Field	Description
   Max number of chunks per document	The maximum number of chunks that you want returned per document. The default value is 10.
   Chunking strategy	Fixed size Fixed size breaks down the full text into smaller passages. You can choose the passage size limit in number of words or number of sentences. Small to big Small to big enables you to choose the top K best matched indexed chunks and expand them to include surrounding chunks. Then the expanded chunks are linked together into a large text and broken into smaller passages. Don’t use the small to big chunking strategy if you’re using full text or truncate index chunking configuration.
   Chunking unit	Words Sentences
   Chunk size	The size of the chunks that are returned. The default value for word size is 750. The default value for sentence size is 40.
   Expanded snippet size	The size of the expanded chunks that are included when you select the small to big chunking strategy.
   Top K results	The number of chunks the reranker returns. If you leave this field empty, the default number of chunks that are returned will be the limit you entered.

5. Select Next.
6. Select the type of condition that you want to be evaluated when the tool is executed.
7. Select Next.
8. Review the selections that you made for the tool and select Add tool.