Minutes Generation Reference

Detailed documentation of all minutes generation options and settings.

Context Field

Purpose: Provide background information to improve AI understanding.

Best practices:

• Include meeting type (standup, review, planning)
• List key participants and their roles
• Define domain-specific terms
• Mention expected topics

Examples:

Meeting Type	Context Example
Engineering standup	"Daily standup for Backend team. Alice (Tech Lead), Bob (Senior Dev), Carol (Junior). Terms: PR, CI, k8s, hotfix."
Client meeting	"Project status with Acme Corp. Internal: John (PM), Sarah (Dev). Client: Mike (Product Owner). Project: E-commerce redesign Phase 2."
Board meeting	"Q4 Board of Directors meeting. Board: Dr. Smith (Chair), Johnson (CFO). Management: CEO, COO. Topics: Q4 financials, 2025 budget."
Interview	"Technical interview for Senior Developer. Interviewers: Sarah (Hiring Manager), Tom (Team Lead). Candidate: Alex Chen."

Correct Transcript First

Purpose: Improve transcript accuracy before generating minutes.

When to enable:

• Audio quality was poor
• Many technical terms were misrecognized
• Speaker names are incorrect
• Heavy accents or fast speech

Trade-offs:

Benefit	Cost
Improved word accuracy	Adds 30-60 seconds processing
Better speaker names	Uses additional AI tokens
Cleaner source text	Beta feature—may need review

When to skip:

• High-quality recording
• Already manually corrected
• Short meetings (<5 minutes)
• Time-sensitive delivery

Custom Templates

Purpose: Control the structure of generated minutes.

Default Sections

The AI pipeline can extract and fill these sections:

Section	Description
Summary / Overview	High-level meeting synopsis
Attendees / Participants	Who was present
Action Items	Tasks with owners and deadlines
Decisions	Choices made with rationale
Discussion Points	Key topics covered
Next Steps / Follow-ups	Planned future actions

Template Syntax

Use {{placeholder}} syntax for dynamic content:

# {{title}}

**Date:** {{date}}
**Attendees:** {{attendees}}

## Summary
{{summary}}

## Decisions
{{decisions}}

## Action Items
| Owner | Task | Due Date |
|-------|------|----------|
{{action_items}}

Template Examples

Executive Summary:

# {{title}} - Executive Summary

**Date:** {{date}}
**Participants:** {{attendees}}

## Key Takeaways
{{summary}}

## Decisions Made
{{decisions}}

## Action Items
{{action_items}}

Scrum/Standup:

# Daily Standup - {{date}}

**Team:** {{attendees}}

## Updates by Person
{{discussion}}

## Blockers
{{blockers}}

## Action Items
{{action_items}}

Client Meeting:

# Client Meeting Minutes

**Client:** [Client Name]
**Date:** {{date}}
**Attendees:** {{attendees}}

## Purpose
{{summary}}

## Discussion
{{discussion}}

## Agreed Actions
{{action_items}}

## Decisions
{{decisions}}

## Next Steps
{{next_steps}}

Template Limitations

Templates work best when they align with what the AI extracts. The pipeline extracts fixed elements from transcripts:

Supported (will be filled):

• Action items (with owners, deadlines, status)
• Decisions (with rationale)
• Key discussions and topics
• Attendees/participants
• Status updates
• Follow-ups and next steps
• Meeting metadata

Not supported (may be empty):

• Risk assessment
• Budget analysis
• Sentiment analysis
• Legal review
• Custom KPIs

WARNING

Custom sections requesting analysis beyond extraction scope will result in empty content or potential hallucinations. Stick to extractable elements for reliable results.

Output Format

Available formats:

Format	Extension	Use Case
Word	.docx	Sharing, printing, formal documentation
Markdown	.md	Wiki, technical documentation, further processing
Plain text	.txt	Simple archives, email pasting, universal compatibility
Auto	varies	System chooses based on template (recommended)

Format recommendations:

• Use Word for distribution to non-technical stakeholders
• Use Markdown for technical teams or wiki integration
• Use Auto when using custom templates (matches template format)

Verbosity

Levels:

Level	Description	Best For
Concise	Key points only, minimal detail	Executive summaries, quick updates
Normal	Balanced coverage (default)	Standard meeting notes
Detailed	Comprehensive, includes context	Legal records, thorough documentation

Examples:

Concise action item:

• John: Deploy to staging (Friday)

Normal action item:

• John will deploy the updated API to staging by Friday to enable QA testing

Detailed action item:

• John (Backend Lead) committed to deploying the updated REST API v2.1 to the staging environment by end of day Friday, January 31st. This deployment is necessary before QA can begin their testing cycle scheduled for the following week. Dependencies: Database migration must complete first.

Progress Steps

The 8-step agentic pipeline:

Step	Name	Description	Typical Duration
1	Segment	Identify topic boundaries in transcript	5-10s
2	Extract (Per-Segment)	Extract items from each topic (parallel)	15-30s
3	Extract (Global)	Find meeting-level items spanning topics	5-10s
4	Verify Extractions	Second pass for missed items	5-10s
5	Merge	Combine and associate items with topics	2-5s
6	Verify Claims	Check claims against source transcript	10-20s
7	Organize	Map extracted data to template structure	2-5s
8	Write	Generate polished final document	10-15s

Total typical time: 30-90 seconds depending on transcript length

Pipeline Features

• Source traceability: Each claim includes source indices and quotes
• Faithfulness scoring: Claims verified with score 0-1
• Speaker attribution confidence: Tracks confidence in owner assignments
• Anti-hallucination grounding: Explicit examples in prompts prevent fabrication
• Parallel extraction: Step 2 processes segments simultaneously (4 workers)
• Conditional verification skip: Step 6 skipped when extraction confidence ≥ 0.85

Performance Optimizations

• Parallel per-segment extraction: 4 workers process segments simultaneously
• Conditional claim verification: Skipped when aggregate confidence is high
• Per-segment limits: Long segments (>12K chars) get smart truncation
• No upfront truncation: Full transcript processed via per-segment extraction

Expected improvement: ~40% reduction in total pipeline time for typical meetings

Troubleshooting

Empty Sections

Causes:

• Meeting didn't contain that information type
• Template requested unsupported content
• Information wasn't explicitly stated

Solutions:

• Check if transcript actually contains action items, decisions, etc.
• Use templates that match the meeting content
• Add context to help AI recognize relevant content
• Try "Detailed" verbosity to capture more information

Inaccurate Speaker Names

Causes:

• Original transcript had misrecognized names
• Speakers not identified in audio
• Similar-sounding names confused

Solutions:

• Enable "Correct transcript first" option
• Include speaker names in context field: "Participants: John Smith (PM), Sarah Chen (Dev Lead)"
• Edit transcript before generating minutes

Missing Action Items

Causes:

• Actions weren't explicitly stated in meeting
• Owners not mentioned by name
• Implicit assignments not captured

Solutions:

• Add participant names and roles to context
• In future meetings, be explicit: "John will handle X by Friday"
• Use "Detailed" verbosity to capture implicit items

Processing Errors

Causes:

• Server under high load
• Very long transcript timeout
• Network interruption

Solutions:

• Wait a moment and retry
• For transcripts >2 hours, consider splitting
• Check transcript has actual content (not empty)
• Verify network connectivity

Wrong Language Output

Causes:

• System detected wrong source language
• Mixed-language transcript confused detection

Solutions:

• Specify language in context: "Meeting in German"
• Ensure transcript language setting is correct

Token Usage

Minutes generation tracks AI token consumption:

• Metrics: stenoris_ai_input_tokens_total{feature="minutes"}, stenoris_ai_output_tokens_total{feature="minutes"}
• Factors affecting usage: Transcript length, verbosity level, number of topics

Token-efficient practices:

• Use "Concise" verbosity when detailed notes aren't needed
• Split very long meetings (>2 hours) into separate transcripts
• Avoid "Correct transcript first" if transcript quality is already good

Related

• AI Minutes Guide - Getting started
• Workspaces - Save default settings