Retrieving passages - Amazon Kendra

Retrieving passages

You can use the Retrieve API as a retriever for retrieval augmented generation (RAG) systems.

RAG systems use generative artificial intelligence to build question-answering applications. RAG systems consist of a retriever and large language models (LLM). Given a query, the retriever identifies the most relevant chunks of text from a corpus of documents and feeds it to the LLM to provide the most useful answer. Then, the LLM analyzes the relevant text chunks or passages and generates a comprehensive response for the query.

The Retrieve API looks at chunks of text or excerpts that are referred to as passages and returns the top passages that are most relevant to the query.

Like the Query API, the Retrieve API also searches for relevant information using semantic search. Semantic search takes into account the search query's context, plus all the available information from the indexed documents. However, by default, the Query API only returns excerpt passages of up to 100 token words. With the Retrieve API, you can retrieve longer passages of up to 200 token words and up to 100 semantically relevant passages. This doesn't include question-answer or FAQ type responses from your index. The passages are text excerpts that can be semantically extracted from multiple documents and multiple parts of the same document. If in extreme cases your documents produce zero passages using the Retrieve API, you can alternatively use the Query API and its types of responses.

You can also do the following with the Retrieve API:

  • Override boosting at the index level

  • Filter based on document fields or attributes

  • Filter based on the user or their group access to documents

  • View the confidence score bucket for a retrieved passage result. The confidence bucket provides a relative ranking that indicates how confident Amazon Kendra is that the response is relevant to the query.

    Note

    Confidence score buckets are currently available only for English.

You can also include certain fields in the response that might provide useful additional information.

The Retrieve API currently doesn't support all features supported by the Query API. The following features are not supported: querying using advance query syntax, suggested spell corrections for queries, faceting, query suggestions to autocomplete search queries, and incremental learning. Note that not all features apply to the Retrieve API. Any future releases of the Retrieve API will be documented in this guide.

The Retrieve API shares the number of query capacity units that you set for your index. For more information on what's included in a single capacity unit and the default base capacity for an index, see Adjusting capacity.

Note

You can't add capacity if you are using the Amazon Kendra Developer Edition; you can only add capacity when using Amazon Kendra Enterprise Edition. For more information on what's included in the Developer and Enterprise Editions, see Amazon Kendra Editions.

The following is an example of using the Retrieve API to retrieve the top 100 most relevant passages from documents in an index for the query "how does amazon kendra work?"

Python
import boto3
import pprint

kendra = boto3.client("kendra")

# Provide the index ID
index_id = "index-id"
# Provide the query text
query = "how does amazon kendra work?"
# You can retrieve up to 100 relevant passages
# You can paginate 100 passages across 10 pages, for example
page_size = 10
page_number = 10

result = kendra.retrieve(
        IndexId = index_id,
        QueryText = query,
        PageSize = page_size,
        PageNumber = page_number)

print("\nRetrieved passage results for query: " + query + "\n")        

for retrieve_result in result["ResultItems"]:

    print("-------------------")
    print("Title: " + str(retrieve_result["DocumentTitle"]))
    print("URI: " + str(retrieve_result["DocumentURI"]))
    print("Passage content: " + str(retrieve_result["Content"]))
    print("------------------\n\n")
Java
package com.amazonaws.kendra;

import software.amazon.awssdk.services.kendra.KendraClient;
import software.amazon.awssdk.services.kendra.model.RetrieveRequest;
import software.amazon.awssdk.services.kendra.model.RetrieveResult;
import software.amazon.awssdk.services.kendra.model.RetrieveResultItem;

public class RetrievePassageExample {
    public static void main(String[] args) {
        KendraClient kendra = KendraClient.builder().build();
        
        String indxId = "index-id";
        String query = "how does amazon kendra work?";
        Integer pgSize = 10;
        Integer pgNumber = 10;

        RetrieveRequest retrieveRequest = retrieveRequest
            .builder()
            .indexId(indxId)
            .queryText(query)
            .pageSize(pgSize)
            .pageNumber(pgNumber)
            .build();

        RetrieveResult retrieveResult = kendra.retrieve(retrieveRequest);

        System.out.println(String.format("\nRetrieved passage results for query: %s", query));
        for(RetrieveResultItem item: retrieveResult.resultItems()) {
            System.out.println("----------------------");
            System.out.println(String.format("Title: %s", documentTitle));
            System.out.println(String.format("URI: %s", documentURI));
            System.out.println(String.format("Passage content: %s", content));
            System.out.println("-----------------------\n");
        }
    }
}