Amazon Bedrock
Collect Amazon Bedrock model invocation logs and runtime metrics with Elastic Agent.
Version | 0.11.0 (View all) |
Compatible Kibana version(s) | 8.13.0 or higher |
Supported Serverless project types | Security Observability |
Subscription level | Basic |
Level of support | Elastic |
Overview
Amazon Bedrock is a fully managed service that makes high-performing foundation models (FMs) from leading AI startups and Amazon available for your use through a unified API. You can choose from a wide range of foundation models to find the model that is best suited for your use case. Amazon Bedrock also offers a broad set of capabilities to build generative AI applications with security, privacy, and responsible AI. Using Amazon Bedrock, you can easily experiment with and evaluate top foundation models for your use cases, privately customize them with your data using techniques such as fine-tuning and Retrieval Augmented Generation (RAG), and build agents that execute tasks using your enterprise systems and data sources.
The Amazon Bedrock integration allows you to easily connect your Amazon Bedrock model invocation logging and runtime metrics to Elastic for seamless collection of invocation logs and runtime metrics to monitor usage.
Elastic Security can leverage this data for security analytics including correlation, visualization and incident response. With invocation logging, you can collect the full request and response data, and any metadata associated with use of your account.
IMPORTANT: Extra AWS charges on AWS API requests will be generated by this integration. Please refer to the AWS integration for more details.
Compatibility
This integration is compatible with the Amazon Bedrock ModelInvocationLog
schema,
version 1.0.
Data streams
The Amazon Bedrock integration collects model invocation logs and runtime metrics.
Data streams:
invocation
: Collects invocation logs, model input data, and model output data for all invocations in your AWS account used in Amazon Bedrock.runtime
: Collects Amazon Bedrock runtime metrics such as model invocation count, invocation latency, input token count, output token count and many more.
Requirements
You need Elasticsearch for storing and searching your data and Kibana for visualizing and managing it. You can use our hosted Elasticsearch Service on Elastic Cloud, which is recommended, or self-manage the Elastic Stack on your own hardware.
Before using any Amazon Bedrock integration you will need:
- AWS Credentials to connect with your AWS account.
- AWS Permissions to make sure the user you're using to connect has permission to share the relevant data.
For more details about these requirements, check the AWS integration documentation.
- Elastic Agent must be installed.
- You can install only one Elastic Agent per host.
- Elastic Agent is required to stream data from the S3 bucket and ship the data to Elastic, where the events will then be processed via the integration's ingest pipelines.
Installing and managing an Elastic Agent
To install and manage an Elastic Agent you have the following options:
Install a Fleet-managed Elastic Agent (recommended)
You install Elastic Agent and use Fleet in Kibana to define, configure, and manage your agents in a central location. We recommend using Fleet management because it makes the management and upgrade of your agents considerably easier.
Install Elastic Agent in standalone mode (advanced users)
You install Elastic Agent and manually configure the agent locally on the system where it is installed. You are responsible for managing and upgrading the agents. This approach is for advanced users only.
Install Elastic Agent in a containerized environment
You can run Elastic Agent inside a container, either with Fleet Server or standalone. Docker images for all versions of Elastic Agent are available from the Elastic Docker registry, and we provide deployment manifests for running on Kubernetes.
To run Elastic Agent, check these requirements.
Setup
To use the Amazon Bedrock model invocation logs, the logging model invocation logging must be enabled and be sent to a log store destination, either S3 or CloudWatch. For more details check the Amazon Bedrock User Guide.
- Set up an Amazon S3 or CloudWatch Logs destination.
- Enable logging. You can do it either through the Amazon Bedrock console or the Amazon Bedrock API.
Collecting Amazon Bedrock model invocation logs from S3 bucket
When collecting logs from S3 bucket is enabled, you can retrieve logs from S3 objects that are pointed to by S3 notification events read from an SQS queue or directly polling list of S3 objects in an S3 bucket.
The use of SQS notification is preferred: polling list of S3 objects is expensive in terms of performance and costs and should be preferably used only when no SQS notification can be attached to the S3 buckets. This input integration also supports S3 notification from SNS to SQS.
SQS notification method is enabled setting queue_url
configuration value. S3
bucket list polling method is enabled setting bucket_arn
configuration value
and number_of_workers
value. Both queue_url
and bucket_arn
cannot be set
at the same time and at least one of the two value must be set.
Collecting Amazon Bedrock model invocation logs from CloudWatch
When collecting logs from CloudWatch is enabled, you can retrieve logs from
all log streams in a specific log group. filterLogEvents
AWS API is used to
list log events from the specified log group.
Exported fields
Field | Description | Type |
---|---|---|
@timestamp | Date/time when the event originated. This is the date/time extracted from the event, typically representing when the event was generated by the source. If the event source has no original timestamp, this value is typically populated by the first time the event was received by the pipeline. Required field for all events. | date |
aws.cloudwatch.message | CloudWatch log message. | text |
aws.s3.bucket.arn | ARN of the S3 bucket that this log retrieved from. | keyword |
aws.s3.bucket.name | Name of the S3 bucket that this log retrieved from. | keyword |
aws.s3.object.key | Name of the S3 object that this log retrieved from. | keyword |
aws_bedrock.invocation.artifacts | flattened | |
aws_bedrock.invocation.error | keyword | |
aws_bedrock.invocation.error_code | keyword | |
aws_bedrock.invocation.image_generation_config.cfg_scale | double | |
aws_bedrock.invocation.image_generation_config.height | long | |
aws_bedrock.invocation.image_generation_config.number_of_images | long | |
aws_bedrock.invocation.image_generation_config.quality | keyword | |
aws_bedrock.invocation.image_generation_config.seed | long | |
aws_bedrock.invocation.image_generation_config.width | long | |
aws_bedrock.invocation.image_variation_params.images | keyword | |
aws_bedrock.invocation.image_variation_params.text | keyword | |
aws_bedrock.invocation.images | keyword | |
aws_bedrock.invocation.input.input_body_json | flattened | |
aws_bedrock.invocation.input.input_body_json_massive_hash | keyword | |
aws_bedrock.invocation.input.input_body_json_massive_length | long | |
aws_bedrock.invocation.input.input_body_s3_path | keyword | |
aws_bedrock.invocation.input.input_content_type | keyword | |
aws_bedrock.invocation.input.input_token_count | todo | long |
aws_bedrock.invocation.model_id | keyword | |
aws_bedrock.invocation.output.completion_text | The formatted LLM text model responses. Only a limited number of LLM text models are supported. | text |
aws_bedrock.invocation.output.output_body_json | flattened | |
aws_bedrock.invocation.output.output_body_s3_path | keyword | |
aws_bedrock.invocation.output.output_content_type | keyword | |
aws_bedrock.invocation.output.output_token_count | long | |
aws_bedrock.invocation.request_id | keyword | |
aws_bedrock.invocation.result | keyword | |
aws_bedrock.invocation.schema_type | keyword | |
aws_bedrock.invocation.schema_version | keyword | |
aws_bedrock.invocation.task_type | keyword | |
cloud.image.id | Image ID for the cloud instance. | keyword |
data_stream.dataset | The field can contain anything that makes sense to signify the source of the data. Examples include nginx.access , prometheus , endpoint etc. For data streams that otherwise fit, but that do not have dataset set we use the value "generic" for the dataset value. event.dataset should have the same value as data_stream.dataset . Beyond the Elasticsearch data stream naming criteria noted above, the dataset value has additional restrictions: * Must not contain - * No longer than 100 characters | constant_keyword |
data_stream.namespace | A user defined namespace. Namespaces are useful to allow grouping of data. Many users already organize their indices this way, and the data stream naming scheme now provides this best practice as a default. Many users will populate this field with default . If no value is used, it falls back to default . Beyond the Elasticsearch index naming criteria noted above, namespace value has the additional restrictions: * Must not contain - * No longer than 100 characters | constant_keyword |
data_stream.type | An overarching type for the data stream. Currently allowed values are "logs" and "metrics". We expect to also add "traces" and "synthetics" in the near future. | constant_keyword |
event.dataset | Event dataset | constant_keyword |
event.module | Name of the module this data is coming from. If your monitoring agent supports the concept of modules or plugins to process events of a given source (e.g. Apache logs), event.module should contain the name of this module. | constant_keyword |
gen_ai.analysis.action_recommended | Recommended actions based on the analysis. | keyword |
gen_ai.analysis.findings | Detailed findings from security tools. | nested |
gen_ai.analysis.function | Name of the security or analysis function used. | keyword |
gen_ai.analysis.tool_names | Name of the security or analysis tools used. | keyword |
gen_ai.completion | The full text of the LLM's response. | text |
gen_ai.compliance.request_triggered | Lists compliance-related filters that were triggered during the processing of the request, such as data privacy filters or regulatory compliance checks. | keyword |
gen_ai.compliance.response_triggered | Lists compliance-related filters that were triggered during the processing of the response, such as data privacy filters or regulatory compliance checks. | keyword |
gen_ai.compliance.violation_code | Code identifying the specific compliance rule that was violated. | keyword |
gen_ai.compliance.violation_detected | Indicates if any compliance violation was detected during the interaction. | boolean |
gen_ai.owasp.description | Description of the OWASP risk triggered. | text |
gen_ai.owasp.id | Identifier for the OWASP risk addressed. | keyword |
gen_ai.performance.request_size | Size of the request payload in bytes. | long |
gen_ai.performance.response_size | Size of the response payload in bytes. | long |
gen_ai.performance.response_time | Time taken by the LLM to generate a response in milliseconds. | long |
gen_ai.performance.start_response_time | Time taken by the LLM to send first response byte in milliseconds. | long |
gen_ai.policy.action | Action taken due to a policy violation, such as blocking, alerting, or modifying the content. | keyword |
gen_ai.policy.confidence | Confidence level in the policy match that triggered the action, quantifying how closely the identified content matched the policy criteria. | keyword |
gen_ai.policy.match_detail.* | object | |
gen_ai.policy.name | Name of the specific policy that was triggered. | keyword |
gen_ai.policy.violation | Specifies if a security policy was violated. | boolean |
gen_ai.prompt | The full text of the user's request to the gen_ai. | text |
gen_ai.request.id | Unique identifier for the LLM request. | keyword |
gen_ai.request.max_tokens | Maximum number of tokens the LLM generates for a request. | integer |
gen_ai.request.model.description | Description of the LLM model. | keyword |
gen_ai.request.model.id | Unique identifier for the LLM model. | keyword |
gen_ai.request.model.instructions | Custom instructions for the LLM model. | text |
gen_ai.request.model.role | Role of the LLM model in the interaction. | keyword |
gen_ai.request.model.type | Type of LLM model. | keyword |
gen_ai.request.model.version | Version of the LLM model used to generate the response. | keyword |
gen_ai.request.temperature | Temperature setting for the LLM request. | float |
gen_ai.request.timestamp | Timestamp when the request was made. | date |
gen_ai.request.top_k | The top_k sampling setting for the LLM request. | float |
gen_ai.request.top_p | The top_p sampling setting for the LLM request. | float |
gen_ai.response.error_code | Error code returned in the LLM response. | keyword |
gen_ai.response.finish_reasons | Reason the LLM response stopped. | keyword |
gen_ai.response.id | Unique identifier for the LLM response. | keyword |
gen_ai.response.model | Name of the LLM a response was generated from. | keyword |
gen_ai.response.timestamp | Timestamp when the response was received. | date |
gen_ai.security.hallucination_consistency | Consistency check between multiple responses. | float |
gen_ai.security.jailbreak_score | Measures similarity to known jailbreak attempts. | float |
gen_ai.security.prompt_injection_score | Measures similarity to known prompt injection attacks. | float |
gen_ai.security.refusal_score | Measures similarity to known LLM refusal responses. | float |
gen_ai.security.regex_pattern_count | Counts occurrences of strings matching user-defined regex patterns. | integer |
gen_ai.sentiment.content_categories | Categories of content identified as sensitive or requiring moderation. | keyword |
gen_ai.sentiment.content_inappropriate | Whether the content was flagged as inappropriate or sensitive. | boolean |
gen_ai.sentiment.score | Sentiment analysis score. | float |
gen_ai.sentiment.toxicity_score | Toxicity analysis score. | float |
gen_ai.system | Name of the LLM foundation model vendor. | keyword |
gen_ai.text.complexity_score | Evaluates the complexity of the text. | float |
gen_ai.text.readability_score | Measures the readability level of the text. | float |
gen_ai.text.similarity_score | Measures the similarity between the prompt and response. | float |
gen_ai.threat.action | Recommended action to mitigate the detected security threat. | keyword |
gen_ai.threat.category | Category of the detected security threat. | keyword |
gen_ai.threat.description | Description of the detected security threat. | text |
gen_ai.threat.detected | Whether a security threat was detected. | boolean |
gen_ai.threat.risk_score | Numerical score indicating the potential risk associated with the response. | float |
gen_ai.threat.signature | Signature of the detected security threat. | keyword |
gen_ai.threat.source | Source of the detected security threat. | keyword |
gen_ai.threat.type | Type of threat detected in the LLM interaction. | keyword |
gen_ai.threat.yara_matches | Stores results from YARA scans including rule matches and categories. | nested |
gen_ai.usage.completion_tokens | Number of tokens in the LLM's response. | integer |
gen_ai.usage.prompt_tokens | Number of tokens in the user's request. | integer |
gen_ai.user.id | Unique identifier for the user. | keyword |
gen_ai.user.rn | Unique resource name for the user. | keyword |
host.containerized | If the host is a container. | boolean |
host.os.build | OS build information. | keyword |
host.os.codename | OS codename, if any. | keyword |
input.type | Type of Filebeat input. | keyword |
log.offset | Log offset | long |
Metrics
Runtime Metrics
Amazon Bedrock runtime metrics include Invocations
, InvocationLatency
,
InvocationClientErrors
, InvocationServerErrors
, OutputTokenCount
,
OutputImageCount
, InvocationThrottles
. These metrics can be used for various use cases including:
- Comparing model latency
- Measuring input and output token counts
- Detecting the number of invocations that the system throttled
An example event for runtime
looks as following:
{
"@timestamp": "2024-07-15T07:35:00.000Z",
"agent": {
"ephemeral_id": "63673811-d18c-4209-8818-df8b346bcb28",
"id": "47a2173f-3f59-4a7c-a022-dee86802c2c1",
"name": "service-integration-dev-idc-1",
"type": "metricbeat",
"version": "8.13.4"
},
"aws": {
"cloudwatch": {
"namespace": "AWS/Bedrock"
}
},
"aws_bedrock": {
"runtime": {
"input_token_count": 848,
"invocation_latency": 2757,
"invocations": 5,
"output_token_count": 1775
}
},
"cloud": {
"account": {
"id": "00000000000000",
"name": "MonitoringAccount"
},
"provider": "aws",
"region": "ap-south-1"
},
"data_stream": {
"dataset": "aws_bedrock.runtime",
"namespace": "ep",
"type": "metrics"
},
"ecs": {
"version": "8.0.0"
},
"elastic_agent": {
"id": "47a2173f-3f59-4a7c-a022-dee86802c2c1",
"snapshot": false,
"version": "8.13.4"
},
"event": {
"agent_id_status": "verified",
"dataset": "aws_bedrock.runtime",
"duration": 174434808,
"ingested": "2024-07-15T07:44:02Z",
"module": "aws"
},
"host": {
"architecture": "x86_64",
"containerized": false,
"hostname": "service-integration-dev-idc-1",
"id": "1bfc9b2d8959f75a520a3cb94cf035c8",
"ip": [
"10.160.0.4",
"172.1.0.1",
"172.17.0.1",
"172.19.0.1",
"172.20.0.1",
"172.22.0.1",
"172.23.0.1",
"172.26.0.1",
"172.27.0.1",
"172.28.0.1",
"172.29.0.1",
"172.30.0.1",
"172.31.0.1",
"192.168.0.1",
"192.168.32.1",
"192.168.49.1",
"192.168.80.1",
"192.168.224.1",
"fe80::42:9cff:fe5b:79b4",
"fe80::42:a5ff:fe15:d63c",
"fe80::42:beff:fe39:f457",
"fe80::42a:f7ff:fe6c:421d",
"fe80::1818:53ff:fea8:3f38",
"fe80::4001:aff:fea0:4",
"fe80::8cfa:3aff:fedb:656a",
"fe80::c890:29ff:fe99:ac1b",
"fe80::fcfc:c2ff:feca:1e28"
],
"mac": [
"02-42-0D-A6-43-C0",
"02-42-23-32-CF-25",
"02-42-27-90-E6-54",
"02-42-34-10-CA-62",
"02-42-4F-1D-94-1B",
"02-42-50-2E-CB-58",
"02-42-5D-42-F3-1D",
"02-42-66-9B-25-B2",
"02-42-99-B7-1B-26",
"02-42-9C-5B-79-B4",
"02-42-A5-15-D6-3C",
"02-42-A6-68-F8-E9",
"02-42-BE-39-F4-57",
"02-42-CE-31-B7-A3",
"02-42-E8-F3-CF-7A",
"02-42-F1-35-B0-41",
"02-42-F4-2F-0F-22",
"06-2A-F7-6C-42-1D",
"1A-18-53-A8-3F-38",
"42-01-0A-A0-00-04",
"8E-FA-3A-DB-65-6A",
"CA-90-29-99-AC-1B",
"FE-FC-C2-CA-1E-28"
],
"name": "service-integration-dev-idc-1",
"os": {
"codename": "bionic",
"family": "debian",
"kernel": "5.4.0-1106-gcp",
"name": "Ubuntu",
"platform": "ubuntu",
"type": "linux",
"version": "18.04.6 LTS (Bionic Beaver)"
}
},
"metricset": {
"name": "cloudwatch",
"period": 300000
},
"service": {
"type": "aws"
}
}
Exported fields
Field | Description | Type | Unit | Metric Type |
---|---|---|---|---|
@timestamp | Event timestamp. | date | ||
agent.id | Unique identifier of this agent (if one exists). Example: For Beats this would be beat.id. | keyword | ||
aws.cloudwatch.namespace | The namespace specified when query cloudwatch api. | keyword | ||
aws.metrics_names_fingerprint | Autogenerated ID representing the fingerprint of the list of metrics names. | keyword | ||
aws_bedrock.runtime.bucketed_step_size | keyword | |||
aws_bedrock.runtime.image_size | keyword | |||
aws_bedrock.runtime.input_token_count | The number of text input tokens. | long | gauge | |
aws_bedrock.runtime.invocation_client_errors | The number of invocations that result in client-side errors. | long | gauge | |
aws_bedrock.runtime.invocation_latency | The average latency of the invocations. | long | ms | gauge |
aws_bedrock.runtime.invocation_server_errors | The number of invocations that result in AWS server-side errors. | long | gauge | |
aws_bedrock.runtime.invocation_throttles | The number of invocations that the system throttled. | long | gauge | |
aws_bedrock.runtime.invocations | The number of requests to the Converse , ConverseStream , InvokeModel , and InvokeModelWithResponseStream API operations. | long | gauge | |
aws_bedrock.runtime.legacymodel_invocations | The number of requests to the legacy models. | long | gauge | |
aws_bedrock.runtime.model_id | keyword | |||
aws_bedrock.runtime.output_image_count | The number of output images. | long | gauge | |
aws_bedrock.runtime.output_token_count | The number of text output tokens. | long | gauge | |
aws_bedrock.runtime.quality | keyword | |||
cloud.account.id | The cloud account or organization id used to identify different entities in a multi-tenant environment. Examples: AWS account id, Google Cloud ORG Id, or other unique identifier. | keyword | ||
cloud.region | Region in which this host, resource, or service is located. | keyword | ||
data_stream.dataset | Data stream dataset. | constant_keyword | ||
data_stream.namespace | Data stream namespace. | constant_keyword | ||
data_stream.type | Data stream type. | constant_keyword | ||
event.module | Name of the module this data is coming from. If your monitoring agent supports the concept of modules or plugins to process events of a given source (e.g. Apache logs), event.module should contain the name of this module. | constant_keyword |
Changelog
Version | Details | Kibana version(s) |
---|---|---|
0.11.0 | Enhancement View pull request | — |
0.10.1 | Bug fix View pull request | — |
0.10.0 | Enhancement View pull request | — |
0.9.1 | Bug fix View pull request | — |
0.9.0 | Enhancement View pull request | — |
0.8.0 | Bug fix View pull request | — |
0.7.0 | Enhancement View pull request | — |
0.6.0 | Enhancement View pull request | — |
0.5.0 | Enhancement View pull request | — |
0.4.0 | Enhancement View pull request | — |
0.3.0 | Enhancement View pull request | — |
0.2.0 | Enhancement View pull request | — |
0.1.3 | Bug fix View pull request | — |
0.1.2 | Bug fix View pull request | — |
0.1.1 | Bug fix View pull request | — |
0.1.0 | Enhancement View pull request | — |