Meeting Transcript Processing with Fenic

View in Github

This example demonstrates how to use Fenic to automatically extract actionable insights from engineering meeting transcripts using semantic extraction and structured data processing.

Overview

Engineering teams generate valuable knowledge in meetings, but capturing and organizing this information is often manual and error-prone. This pipeline automates the extraction of:

• Action Items: Tasks, assignees, and deadlines
• Decisions: Key decisions and their rationale
• Technical Entities: Services, technologies, metrics, and incident references
• Team Analytics: Workload distribution and productivity metrics

Features

• Native transcript parsing with fenic's built-in functions
• Semantic extraction of technical entities, action items, and decisions
• Structured data processing on unstructured meeting content
• Automated knowledge capture for engineering teams
• Actionable insights for project management and team coordination

Sample Data

The example processes three types of engineering meetings:

1. Architecture Review - Technical discussions about system design and bottlenecks
2. Incident Post-Mortem - Analysis of outages and mitigation strategies
3. Sprint Planning - Task allocation and project prioritization

Pipeline Steps

Step 1: Transcript Parsing

# Parse transcripts into structured segments
parsed_transcripts_df = transcripts_df.with_column(
    "structured_transcript",
    fc.text.parse_transcript(fc.col("transcript"), 'generic')
)

Step 2: Segment Extraction

Break down transcripts into individual speaking segments with speaker, start_time, and content.

Step 3: Semantic Schema Definition

Define extraction schemas using both Fenic's ExtractSchema and Pydantic models:

# Technical entities using ExtractSchema
technical_entities_schema = fc.ExtractSchema([
    fc.ExtractSchemaField(
        name="services",
        data_type=fc.ExtractSchemaList(element_type=fc.StringType),
        description="Technical services or systems mentioned"
    ),
    # ... more fields
])

# Action items using Pydantic
class ActionItemSchema(BaseModel):
    has_action_item: str = Field(description="Whether this segment contains an action item (yes/no)")
    assignee: str = Field(default=None, description="Person assigned to the action item")
    task_description: str = Field(description="Description of the task or action")
    deadline: str = Field(default=None, description="When the task should be completed")

Step 4: Semantic Extraction

Apply AI-powered extraction to identify structured information from natural language:

enriched_df = segments_df.with_column(
    "technical_entities",
    fc.semantic.extract(fc.col("content"), technical_entities_schema)
).with_column(
    "action_items",
    fc.semantic.extract(fc.col("content"), ActionItemSchema)
).with_column(
    "decisions",
    fc.semantic.extract(fc.col("content"), DecisionSchema)
)

Step 5: Analytics and Aggregation

Generate meeting-level insights and team analytics:

• Action item workload by team member
• Technology and service mentions across meetings
• Decision summary and rationale tracking
• Meeting productivity metrics

Expected Output

The pipeline produces structured insights including:

Action Items Summary:

meeting_id	meeting_type	assignee	task_description	deadline
ARCH-2024-1	Architecture Review	Mike	investigate Redis impl	next Friday
INC-2024-12	Incident Post-Mortem	Sam	review batch processing	tomorrow EOD

Team Workload Distribution:

assignee	assigned_tasks
Mike	2
Sam	1
Lisa	1

Technology Mentions:

technologies	mention_count
Redis	3
PostgreSQL	2
JWT	2

Prerequisites

1. OpenAI API Key: Required for semantic extraction

bash export OPENAI_API_KEY="your-api-key-here"

1. Fenic Installation: bash uv sync uv run maturin develop --uv

Running the Example

uv run python examples/meeting_transcript_processing/transcript_processing.py

Use Cases

This pipeline is valuable for:

• Engineering Managers: Track team workload and action item distribution
• Technical Program Managers: Monitor project decisions and technical debt
• DevOps Teams: Analyze incident patterns and response procedures
• Architecture Teams: Identify technology adoption trends and system bottlenecks

Extensions

The example can be extended to:

• Integrate with calendar systems for automatic transcript ingestion
• Export to project management tools (Jira, Linear, etc.)
• Build dashboards for engineering metrics
• Create automated follow-up reminders
• Analyze team communication patterns

Technical Notes

• Uses gpt-4o-mini for fast and cost-effective semantic extraction
• Handles mixed transcript formats automatically
• Implements workarounds for current framework limitations
• Demonstrates both ExtractSchema and Pydantic model approaches