Incident AI Summarization (public preview)

The Instana Incident AI Summarization uses IBM watsonx to summarize the data that is extracted from the related events, notes (including ServiceNow comments), and successful past actions of an incident. By using this feature, you can generate a short summary of events that occurred on an affected entity, the notes of incidents that include ServiceNow comments (bidirectional flow must be enabled) and a list of successful past actions that are taken on the incident in the past seven days. The affected entity in an incident is identified by using the topology of the resource (for example, k8s pod, application, service, and node) where the incident occurred.

The Instana Incident AI Summarization is offered as a public preview on Instana SaaS environments only.

Prerequisites

You can use Instana Incident AI Summarization only if the following prerequisite conditions are met:

  • Instana is deployed on SaaS environment.
  • At least one related event is present for the incident for which the summary is generated.

Generating the summary

To generate a summary of the related events to an incident, complete the following steps:

  1. On the Incidents tab, select the incident.
  2. Click Notes and activity to open the Notes and activity window.
  3. Click Generate a summary.
ServiceNow notes Figure 2. ServiceNow notes ServiceNow summary Figure 3. ServiceNow summary
Notes expanded example Figure 4. Full summary example (notes excluded)

The summary of all related events, notes, successful past actions that are taken on the incident is generated. Three sections are included in the generated summary:

Summary of the incident

Each bullet point in the summary represents an entity and the details of the related events.

Summary of notes

This section summarizes all notes that are recorded in the Notes & Activity section of an Incident. It includes the ServiceNow notes imported when bidirectional flow is enabled, up to the time when the summary is requested.

Successful past actions

This section lists unique actions from the Instana Automation catalog that you ran on similar incidents in the past. These incidents are with the same configuration and scope. Currently, the look back period is 7 days.

Notes:
  • The pre-processing logic that is configured in IBM watsonx makes sure that the language model does not make any assumptions or suggest fixes.
  • The generated output is checked and hallucinations (false or misleading information) are removed from the output.
  • Each generated summary is graded based on a hallucination score. Summaries are generated repeatedly until the score is 0.

Data fields that are used for summary generation

The following data is used from the incident and the related events to generate the summary:

  • Incident:
    • The problem description and details fields
    • Triggering event problem description and details fields
    • Event specification ID
  • Related events:
    • Entity name
    • Entity type
    • Problem description
    • Details
  • Notes:
    • Text content

Large language model - Granite Chat

The large language model (LLM) which is used for generating summary is ibm/granite-chat-13b-v2. This is a model that is used for chat completion tasks, summary generation. The specifications of the model are as follows:

  • The model was trained by using 4x Tensor Parallel + 4x Pipeline Parallel + Megatron distributed optimizer Megatron-LM.
  • Cluster: CCC
  • GPUs: 256x A100 80GB
  • Interconnect: 200 gigabit Infiniband
  • Dataset streamed over GPFS