AMAZON.QnAIntent - Amazon Lex

AMAZON.QnAIntent

Note

Before you can take advantage of the generative AI features, you must fulfill the following prerequisites

  1. Navigate to the Amazon Bedrock console and sign up for access to the Anthropic Claude model you intend to use (for more information, see Model access). For information about pricing for using Amazon Bedrock, see Amazon Bedrock pricing.

  2. Turn on the generative AI capabilities for your bot locale. To do so, follow the steps at Optimize Lex V2 bot creation and performance by using generative AI.

Responds to customer questions by using an Amazon Bedrock FM to search and summarize FAQ responses. This intent is activated when an utterance is not classified into any of the other intents present in the bot. Note that this intent will not be activated for missed utterances when eliciting a slot value. Once recognized, the AMAZON.QnAIntent, uses the specified Amazon Bedrock model to search the configured knowledge base and respond to the customer question.

Warning

You can't use the AMAZON.QnAIntent and the AMAZON.KendraSearchIntent in the same bot locale.

The following knowledge store options are available. You must have already created the knowledge store and indexed the documents within it.

If you select this intent, you configure the following fields and then select Add to add the intent.

  • Bedrock model – Choose the provider and foundation model to use for this intent. Currently, Anthropic Claude V2, Anthropic Claude 3 Haiku, Anthropic Claude 3 Haiku, and Anthropic Claude Instant are supported.

  • Knowledge store – Choose the source from which you want the model pull information from to answer customer questions. The following sources are available.

    • OpenSearch – Configure the following fields.

      • Domain endpoint – Provide the domain endpoint that you made for the domain or that was provided to you after domain creation.

      • Index name – Provide the index to search. For more information, see Indexing data in Amazon OpenSearch Service.

      • Choose how you want to return the response to the customer.

        • Exact response – When this option is enabled, the value in the Answer field is used as is for the bot response. The configured Amazon Bedrock foundation model is used to select the exact answer content as-is, without any content synthesis or summarization. Specify the name of the question and answer fields that were configured in the OpenSearch database.

        • Include fields – Returns an answer generated by the model using the fields you specify. Specify the name of up to five fields that were configured in the OpenSearch database. Use a semicolon (;) to separate fields.

    • Amazon Kendra – Configure the following fields.

      • Amazon Kendra index – Select the Amazon Kendra index that you want your bot to search.

      • Amazon Kendra filter – To create a filter, select this checkbox. For more information on the Amazon Kendra search filter JSON format, see Using document attributes to filter search results.

      • Exact response – To let your bot return the exact response returned by Amazon Kendra, select this checkbox. Otherwise, the Amazon Bedrock model you select generates a response based on the results.

        Note

        To use this feature, you must first add FAQ questions to your index by following the steps at Adding frequently asked questions (FAQs) to an index.

    • Amazon Bedrock knowledge base – If you choose this option, specify the ID of the knowledge base. You can find the ID by checking the details page of the knowledge base in the console, or by sending a GetKnowledgeBase request.

      • Exact response – When this option is enabled, the value in the Answer field is used as is for the bot response. The configured Amazon Bedrock foundation model is used to select the exact answer content as-is, without any content synthesis or summarization. To use exact response for Amazon Bedrock Knowledge Base you need to do the following:

        • Create individual JSON files with each file containing an answer field that contains the exact response that needs to be returned to end-user.

        • When indexing these documents in Bedrock Knowledge Base, select Chunking strategy as No Chunking..

        • Define the answer field in Amazon Lex V2, as the Answer field in the Bedrock Knowledge Base.

The responses from the QnAIntent will be stored into the request attributes as shown below:

  • x-amz-lex:qnA-search-response – The response from the QnAIntent to the question or utterance.

  • x-amz-lex:qnA-search-response-source – Points to the document or list of documents used to generate the response.

Additional model configurations

When AMAZON.QnAIntent is invoked it uses a default prompt template that combines instructions and context with the user query to construct the prompt that’s sent to the model for response generation. You can also provide a custom prompt or update the default prompt to match your requirements.

You can engineer the prompt template with the following tools:

Prompt placeholders – Pre-defined variables in AMAZON.QnAIntent for Amazon Bedrock that are dynamically filled in at runtime during the bedrock call. In the system prompt, you can see these placeholders surrounded by the $ symbol. The following list describes the placeholders you can use:

Variable Replaced by Model Required?
$query_results$ The retrieved results for the user query from the Knowledge Store Anthropic Claude3 Haiku, Anthropic Claude3 Sonnet Yes
$output_instruction$ Underlying instructions for formatting the response generation and citations. Differs by model. If you define your own formatting instructions, we suggest that you remove this placeholder. Anthropic Claude3 Haiku, Anthropic Claude3 Sonnet No

Default prompt being used is:


$query_results$

Please only follow the instructions in <instruction> tags below.
<instruction>
Given the conversation history, and <Context>:
(1) first, identify the user query intent and classify it as one of the categories: FAQ_QUERY, OTHER_QUERY, GIBBERISH, GREETINGS, AFFIRMATION, CHITCHAT, or MISC;
(2) second, if the intent is FAQ_QUERY, predict the most relevant grounding passage(s) by providing the passage id(s) or output CANNOTANSWER;
(3) then, generate a concise, to-the-point FAQ-style response ONLY USING the grounding content in <Context>; or output CANNOTANSWER if the user query/request cannot be directly answered with the grounding content. DO NOT mention about the grounding passages such as ids or other meta data; do not create new content not presented in <Context>. Do NOT respond to query that is ill-intented or off-topic;
(4) lastly, provide the confidence level of the above prediction as LOW, MID or HIGH.
</instruction>

$output_instruction$

$output_instruction$ is replaced with:

Give your final response in the following form:
<answer>
<intent>FAQ_QUERY or OTHER_QUERY or GIBBERISH or GREETINGS or AFFIRMATION or CHITCHAT or MISC</intent>
<text>a concise FAQ-style response or CANNOTANSWER</text>
<passage_id>passage_id or CANNOTANSWER</passage_id>
<confidence>LOW or MID or HIGH</confidence>
</answer>
Note

If you decide not to use the default instructions, then whatever output the LLM provides will be returned as-is back to the end user.

The output instructions need to contain <text></text> and <passageId></passageId> tags and instructions for the LLM to return the passageIds to provide the response and source attribution.

Amazon Bedrock knowledge base metadata filtering support through session attributes

You can pass the Amazon Bedrock knowledge base metadata filters as part of session attribute x-amz-lex:bkb-retrieval-filter.


             {"sessionAttributes":{"x-amz-lex:bkb-retrieval-filter":"{\"equals\":{\"key\":\"insurancetype\",\"value\":\"farmers\"}}
Note

You need to use the Amazon Bedrock knowledge base as the Data store for the QnAIntent to use this filter. For more information, see Metadata filtering

Inference configurations

You can define the inference configurations that will be used when making the call to LLM using session attribute:

  • temperature: type Integer

  • topP

  • maxTokens

Example:


         {"sessionAttributes":{"x-amz-lex:llm-text-inference-config":"{\"temperature\":0,\"topP\":1,\"maxTokens\":200}"}}

Bedrock Guardrails support through build-time and session attributes

  • By using the Console at Buildtime – Provide the GuardrailsIdentifier and the GuardrailsVersion. Learn more under the Additional Model Configurations section.

  • By using Session attributes – You can also define the Guardrails configuration using the session attributes: x-amz-lex:bedrock-guardrails-identifier and x-amz-lex:bedrock-guardrails-version.

For more information on using Bedrock Guardrails, see Guardrails.