Transcript Summarization

Transcript Summarization

Table of Contents

• Overview
• BEDROCK Option (Default)
  • Prompt Template Storage
  • Attribute Naming Convention
  • Customizing Prompts
  • Admin UI
• LAMBDA Option
  • Lambda Function Requirements
  • Multi-Section Summaries
  • Minimal Python Example
• FetchTranscript Utility Lambda
• Related Documentation

Overview

LMA summarizes meeting transcripts automatically when a meeting ends, and on-demand via UI buttons during or after meetings. Two summarization options are available: BEDROCK (default) and LAMBDA.

BEDROCK Option (Default)

The BEDROCK option uses the selected Amazon Bedrock foundation model (default: Claude Haiku 4.5) to generate meeting summaries from prompt templates.

Prompt Template Storage

Prompt templates are stored in DynamoDB as two separate items:

1. Default prompt templates — Ship with LMA and may change with new releases. View the current defaults via the stack output LLMDefaultPromptSummaryTemplate.
2. Custom prompt templates — User-defined prompts that persist across stack updates and are never overwritten by LMA upgrades. View and edit via the stack output LLMCustomPromptSummaryTemplate or through the admin UI.

Custom prompts always take precedence over default prompts when both exist for the same attribute name.

Attribute Naming Convention

Summary attributes use the N#Label format, where:

• N is a sequence number that controls the display order in the UI.
• Label is the human-readable name shown as the section heading.

Examples:

• 1#Summary
• 2#Action Items
• 3#Key Decisions

Customizing Prompts

Each attribute’s value contains the prompt template text that instructs the model how to generate that section of the summary.

• Use the {transcript} placeholder in your prompt template where the meeting transcript should be inserted.
• Use <br> for newlines within the prompt template.

To add a new summary section: Create a new attribute in the Custom prompts DynamoDB item with a unique N#Label name (for example, 4#Follow-Up Questions).

To override a default section: Create an attribute in the Custom prompts item with the same name as the default attribute you want to override (for example, 1#Summary with your custom prompt text).

To remove a default section: Create an attribute in the Custom prompts item with the same name as the default attribute and set its value to an empty string or NONE.

Admin UI

Administrators can view and edit prompt templates directly in the LMA web interface at:

/#/configuration/transcript-summary

This provides a convenient way to manage prompts without directly editing DynamoDB items.

LAMBDA Option

The LAMBDA option lets you provide your own Lambda function to generate meeting summaries with complete control over the summarization logic.

Lambda Function Requirements

Configure the LAMBDA option by setting the EndOfCallLambdaHookFunctionArn CloudFormation parameter to the ARN of your Lambda function.

Your Lambda function:

• Receives the CallId in the event payload.
• Must return a JSON object with a summary key: {"summary": "..."}.
• The summary value supports Markdown formatting.

Multi-Section Summaries

To return a summary with multiple sections, return a JSON-encoded string as the summary value. The string should contain key-value pairs where keys are section labels and values are the section content. LMA will parse this and display each section separately in the UI.

Minimal Python Example

import json

def lambda_handler(event, context):
    print(json.dumps(event))
    call_id = event.get("CallId")
    # Implement your summarization logic here
    summary = "Placeholder for actual summary"
    return {"summary": summary}

FetchTranscript Utility Lambda

LMA provides a utility Lambda function for retrieving processed meeting transcripts. This is especially useful when building custom summarization Lambda functions.

The function ARN is available in the stack output FetchTranscriptLambdaArn.

Parameters:

Parameter	Type	Required	Description
CallId	String	Yes	The unique identifier of the meeting/call.
ProcessTranscript	Boolean	No	When true, condenses sequential utterances from the same speaker, removes filler words, and strips HTML tags.
TokenCount	Integer	No	Trims the returned transcript to approximately N tokens. Useful for staying within model context limits.

Example invocation payload:

{
  "CallId": "2359fb61-...",
  "TokenCount": 1024,
  "ProcessTranscript": true
}

Related Documentation

• Meeting Assistant
• Lambda Hook Functions