Enterprise Search
Took Johor Plantations’ internal search from a proof of concept on Amazon Kendra to production on Amazon Bedrock and Amazon OpenSearch Serverless.
The problem
Johor Plantations Group runs 23 oil palm estates across Johor and Pahang, and much of what its people need to know is written down somewhere in a large document archive. The trouble is that the answer to a specific question usually sits in one passage of one document. Keyword search returned every document that shared a word with the question, and left people to dig for the passage themselves.
The solution
I took it from a proof of concept to production:
- A proof of concept on Amazon Kendra and Amazon Bedrock, to show the archive could answer real questions before building any pipeline.
- A production search on Amazon Bedrock Knowledge Bases and Amazon OpenSearch Serverless. Ask it something and Amazon Bedrock writes an answer in plain words from the passages it found, with the source documents ranked beside it. An answer you cannot check is not one you can act on, so the sources always come with it.
Why production left Amazon Kendra
Amazon Kendra is a managed search service that indexes documents and ranks results on its own, which made it the quick way to a proof of concept: there was no pipeline to build first. Production needed two things it could not give us. Control over how documents are split into passages, which embedding model reads them, and whether search is semantic or hybrid, all of which are settings in Amazon Bedrock Knowledge Bases on Amazon OpenSearch Serverless. And cost: Kendra bills an index by the hour, whether anyone is searching or not.