If you are using Amazon Lex V2, refer to the Amazon Lex V2 guide instead.

 

If you are using Amazon Lex V1, we recommend upgrading your bots to Amazon Lex V2. We are no longer adding new features to V1 and strongly recommend using V2 for all new bots.

AMAZON.KendraSearchIntent

To search documents that you have indexed with Amazon Kendra, use the AMAZON.KendraSearchIntent intent. When Amazon Lex can't determine the next action in a conversation with the user, it triggers the search intent.

The AMAZON.KendraSearchIntent is available only in the English (US) (en-US) locale and in the US East (N. Virginia), US West (Oregon) and Europe (Ireland) Regions.

Amazon Kendra is a machine-learning-based search service that indexes natural language documents such as PDF documents or Microsoft Word files. It can search indexed documents and return the following types of responses to a question:

For an example of using the AMAZON.KendraSearchIntent, see Example: Creating a FAQ Bot for an Amazon Kendra Index.

If you configure an AMAZON.KendraSearchIntent intent for your bot, Amazon Lex calls the intent whenever it can't determine the user utterance for a slot or intent. For example, if your bot is eliciting a response for a slot type called "pizza topping" and the user says "What is a pizza?," Amazon Lex calls the AMAZON.KendraSearchIntent to handle the question. If there is no response from Amazon Kendra, the conversation continues as configured in the bot.

When you use both the AMAZON.KendraSearchIntent and the AMAZON.FallbackIntent in the same bot, Amazon Lex uses the intents as follows:

There are three ways to use the AMAZON.KendraSearchIntent to make a request to Amazon Kendra:

To specify the queryFilterString parameter when you create a bot, or to specify the kendraQueryFilterString field when you call the delegate action in a dialog Lambda function, you specify a string that is used as the attribute filter for the Amazon Kendra query. If the string isn't a valid attribute filter, you'll get an InvalidBotConfigException exception at runtime. For more information about attribute filters, see Using document attributes to filter queries in the Amazon Kendra Developer Guide.

To have control over the query that Amazon Lex sends to Amazon Kendra, you can specify a query in the kendraQueryRequestPayloadfield in your dialog Lambda function. If the query isn't valid, Amazon Lex returns an InvalidLambdaResponseException exception. For more information, see the Query operation in the Amazon Kendra Developer Guide.

For an example of how to use the AMAZON.KendraSearchIntent, see Example: Creating a FAQ Bot for an Amazon Kendra Index.

IAM Policy for Amazon Kendra Search

To use the AMAZON.KendraSearchIntent intent, you must use a role that provides AWS Identity and Access Management (IAM) policies that enable Amazon Lex to assume a runtime role that has permission to call the Amazon Kendra Query intent. The IAM settings that you use depend on whether you create the AMAZON.KendraSearchIntent using the Amazon Lex console, or using an AWS SDK or the AWS Command Line Interface (AWS CLI). When you use the console, you can choose between adding permission to call Amazon Kendra to the Amazon Lex service-linked role or using a role specifically for calling the Amazon Kendra Query operation. When you use the AWS CLI or an SDK to create the intent, you must use a role specifically for calling the Query operation.

Attaching Permissions

You can use the console to attach permissions to access the Amazon Kendra Query operation to the default Amazon Lex service-linked role. When you attach permissions to the service-linked role, you don't have to create and manage a runtime role specifically to connect to the Amazon Kendra index.

The user, role, or group that you use to access the Amazon Lex console must have permissions to manage role policies. Attach the following IAM policy to the console access role. When you grant these permissions, the role has permissions to change the existing service-linked role policy.

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "iam:AttachRolePolicy",
                "iam:PutRolePolicy",
                "iam:GetRolePolicy"
            ],
            "Resource": "arn:aws:iam::*:role/aws-service-role/lex.amazonaws.com/AWSServiceRoleForLexBots"
        },
        {
            "Effect": "Allow",
            "Action": "iam:ListRoles",
            "Resource": "*"
        }
    ]
}

Specifying a Role

You can use the console, the AWS CLI, or the API to specify a runtime role to use when calling the Amazon Kendra Query operation.

The user, role, or group that you use to specify the runtime role must have the iam:PassRole permission. The following policy defines the permission. You can use the iam:AssociatedResourceArn and iam:PassedToService condition context keys to further limit the scope of the permissions. For more information, see IAM and AWS STS Condition Context Keys in the AWS Identity and Access Management User Guide.

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": "iam:PassRole",
            "Resource": "arn:aws:iam::account:role/role"
        }
    ]
}

The runtime role that Amazon Lex needs to use to call Amazon Kendra must have the kendra:Query permissions. When you use an existing IAM role for permission to call the Amazon Kendra Query operation, the role must have the following policy attached.

You can use the IAM console, the IAM API, or the AWS CLI to create a policy and attach it to a role. These instructions use the AWS CLI to create the role and policies.

You can choose to update the Amazon Lex service-linked role or to use a role that you created when you create the AMAZON.KendraSearchIntent for your bot. The following procedure shows how to choose the IAM role to use.

Using Request and Session Attributes as Filters

To filter the response from Amazon Kendra to items related to current conversation, use session and request attributes as filters by adding the queryFilterString parameter when you create your bot. You specify a placeholder for the attribute when you create the intent, and then Amazon Lex V2 substitutes a value before it calls Amazon Kendra. For more information about request attributes, see Setting Request Attributes. For more information about session attributes, see Setting Session Attributes.

The following is a example of a queryFilterString parameter that uses a string to filter the Amazon Kendra query.

"{"equalsTo": {"key": "City", "value": {"stringValue": "Seattle"}}}"

The following is an example of a queryFilterString parameter that uses a session attribute called "SourceURI" to filter the Amazon Kendra query.

"{"equalsTo": {"key": "SourceURI","value": {"stringValue": "[FileURL]"}}}"

The following is an example of a queryFilterString parameter that uses a request attribute called "DepartmentName" to filter the Amazon Kendra query.

"{"equalsTo": {"key": "Department","value": {"stringValue": "((DepartmentName))"}}}"

The AMAZON.KendraSearchInteng filters use the same format as the Amazon Kendra search filters. For more information, see Using document attributes to filter search results in the Amazon Kendra developer guide.

The query filter string used with the AMAZON.KendraSearchIntent must use lower-case letters for the first letter of each filter. For example, the following is a valid query filter for the AMAZON.KendraSearchIntent.

{
    "andAllFilters": [
        {
            "equalsTo": {
                "key": "City",
                "value": {
                    "stringValue": "Seattle"
                }
            }
        },
        {
            "equalsTo": {
                "key": "State",
                "value": {
                    "stringValue": "Washington"
                }
            }
        }
    ]
}

Using the Search Response

Amazon Kendra returns the response to a search in the intent's conclusion statement. The intent must have a conclusion statement unless a fulfillment Lambda function produces a conclusion message.

Amazon Kendra has four types of responses.

The responses are returned in request attributes. There can be up to five responses for each attribute, numbered 1 through 5. For more information about responses, see Types of response in the Amazon Kendra Developer Guide.

The conclusion statement must have one or more message groups. Each message group contains one or more messages. Each message can contain one or more placeholder variables that are replaced by request attributes in the response from Amazon Kendra. There must be at least one message in the message group where all of the variables in the message are replaced by request attribute values in the runtime response, or there must be a message in the group with no placeholder variables. The request attributes are set off with double parentheses ("((" "))"). The following message group messages match any response from Amazon Kendra:

Using a Lambda Function to Manage the Request and Response

The AMAZON.KendraSearchIntent intent can use your dialog code hook and fulfillment code hook to manage the request to Amazon Kendra and the response. Use the dialog code hook Lambda function when you want to modify the query that you send to Amazon Kendra, and the fulfillment code hook Lambda function when you want to modify the response.

Creating a Query with the Dialog Code Hook

You can use the dialog code hook to create a query to send to Amazon Kendra. Using the dialog code hook is optional. If you don't specify a dialog code hook, Amazon Lex constructs a query from the user utterance and uses the queryFilterString that you provided when you configured the intent, if you provided one.

You can use two fields in the dialog code hook response to modify the request to Amazon Kendra:

After you have created the filter or query string, you send the response to Amazon Lex with the dialogAction field of the response set to delegate. Amazon Lex sends the query to Amazon Kendra and then returns the query response to the fulfillment code hook.

Using the Fulfillment Code Hook for the Response

After Amazon Lex sends a query to Amazon Kendra, the query response is returned to the AMAZON.KendraSearchIntent fulfillment Lambda function. The input event to the code hook contains the complete response from Amazon Kendra. The query data is in the same structure as the one returned by the Amazon Kendra Query operation. For more information, see Query response syntax in the Amazon Kendra Developer Guide.

The fulfillment code hook is optional. If one does not exist, or if the code hook doesn't return a message in the response, Amazon Lex uses the conclusion statement for responses.