Amazon Bedrock Knowledge Base Chunks Retrieval Issue

Question

We are currently using Amazon Bedrock knowledge Base (Embedding Model: titan text embedding V2) with DataSource being S3 Bucket which has weather data in CSV format.

When a prompt is asked for a specific date the knowledge base is not able to retrieve the correct chunk of the specified date. The same also happens when we convert csv to pdf/markdown.

What would be the best way that allows the bedrock knowledge base to retrieve the correct date chunk from vector DB. Do we have to change our text embedding model to multimodal embedding? or do we have to change our data to any other format?

Didier Durand · Answer

Hi,

You should try to add metadata with your date to the KB.

Announcement made in March: https://aws.amazon.com/about-aws/whats-new/2024/03/knowledge-bases-amazon-bedrock-metadata-filtering/

See documentation: https://docs.aws.amazon.com/bedrock/latest/userguide/kb-metadata.html

Best,

Didier

Alex T · Answer

Is this really a semantic vector search use-case? It sounds like you're trying to do retrieval over structured/tabular data, more like "question-to-SQL" than search... Which might be a use case for Bedrock Knowledge Base's new structured data retrieval capabilities, as detailed (alongside other features) in this blog.

If the knowledge base includes other information aside from the weather, you could also consider whether to treat the weather separately as one "tool", and use a Bedrock Agent to dynamically decide when to call it vs the un-structured knowledge base.

Amazon Bedrock Knowledge Base Chunks Retrieval Issue

Add your answer

Relevant content