copcrawler.com & AlertPage.ai

Some code snippets of the audio preprocessing, transcription, incident classification, and geocoding workflow for copcrawler.com & alertpage.ai

LLM classification & geocoding

flowchart TD
    A[Start: Radio Transmission] --> B[Split by Talkgroups]
    B --> C[For each transmission]
    C --> D{No Speech Found?}
    D -->|Yes| C
    D -->|No| E{Duplicate text?}
    E -->|Yes| C
    E -->|No| F[**JustifyIncident**<br/>LLM Function]
    
    F --> G{Extract incident?}
    G -->|No| H[Skip transmission]
    G -->|Yes| I[**ClassifyIncident**<br/>LLM Function]
    
    I --> J{Classification valid?}
    J -->|No - UNKNOWN| H
    J -->|Yes| K[**CheckForGeocodeableAddress**<br/>LLM Function]
    
    K --> L{Has address?}
    L -->|No| H
    L -->|Yes| M[**ExtractIncidentLocation**<br/>LLM Function]
    
    M --> N{Location extracted?}
    N -->|No| H
    N -->|Yes| O[**ValidateIncidentClassificationLocation**<br/>LLM Function]
    
    O --> P{Validation passed?}
    P -->|No| H
    P -->|Yes| Q[**IncidentKeywordSelector**<br/>LLM Function]
    
    Q --> R[Create LLM Transcript Node]
    R --> S[**Geocoding Process**]
    
    S --> T[Check Cache]
    T --> U{Cached result?}
    U -->|Yes| V[Use cached coordinates]
    U -->|No| W[LocationIQ Geocoding]
    
    W --> X{Geocoding successful?}
    X -->|No| Y[Skip incident]
    X -->|Yes| Z[**Validate Geocode**<br/>Based on address type]
    
    Z --> AA[**NormalGeocodeValidator**<br/>**PlaceGeocodeValidator**<br/>**IntersectionGeocodeValidator**<br/>LLM Functions]
    AA --> BB{Geocode valid?}
    BB -->|No| Y
    BB -->|Yes| CC[Cache result]
    CC --> DD[Create Simplified Incident]
    
    V --> DD
    DD --> EE[Return processed incidents]
    
    H --> C
    Y --> C
    
    %% style F fill:#e1f5fe
    %% style I fill:#e1f5fe
    %% style K fill:#e1f5fe
    %% style M fill:#e1f5fe
    %% style O fill:#e1f5fe
    %% style Q fill:#e1f5fe
    %% style AA fill:#e1f5fe
    
    classDef llmFunction fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    classDef decision fill:#fff3e0,stroke:#e65100,stroke-width:2px
    classDef process fill:#f3e5f5,stroke:#4a148c,stroke-width:2px
    classDef endpoint fill:#e8f5e8,stroke:#1b5e20,stroke-width:2px

Loading

LLM Functions Overview

1. JustifyIncident (gpt-4o-mini)

• Input: Transcript text
• Output: extract_incident (boolean)
• Purpose: Determines if transcript contains actionable police/fire/ems information

2. ClassifyIncident (gpt-4o-mini)

• Input: Transcript text
• Output: type_of_incident, meta_category, description
• Purpose: Classifies incident type and meta-category

3. CheckForGeocodeableAddress (gpt-4o-mini)

• Input: Transcript text
• Output: has_address (boolean)
• Purpose: Determines if transcript contains a geocodeable address

4. ExtractIncidentLocation (gpt-4o-mini)

• Input: Transcript text
• Output: extracted_location, address_type, confidence_score
• Purpose: Extracts location information and classifies address type

5. ValidateIncidentClassificationLocation (gpt-4o)

• Input: Transmission text, type of incident, extracted location
• Output: is_valid (boolean)
• Purpose: Cross-validates that extracted location matches incident classification

6. IncidentKeywordSelector (gpt-4o-mini)

• Input: Transcript text, valid keywords list
• Output: relevant_keywords (array)
• Purpose: Selects semantically relevant keywords from predefined list

7. Geocode Validators (gpt-4o-mini)

• NormalGeocodeValidator: Validates standard address geocoding
• PlaceGeocodeValidator: Validates place/landmark geocoding
• IntersectionGeocodeValidator: Validates intersection geocoding
• Input: LLM extracted address, geocoded address
• Output: is_valid (boolean)
• Purpose: Ensures geocoding service results match LLM extracted addresses

Workflow Summary

The system processes radio transmissions through a multi-stage LLM pipeline:

1. Pre-processing: Split transmissions by talkgroups, filter duplicates/empty
2. Justification: Determine if transmission is actionable
3. Classification: Categorize incident type and severity
4. Address Detection: Check for geocodeable addresses
5. Location Extraction: Extract specific location details
6. Cross-validation: Verify location matches classification
7. Keyword Selection: Extract relevant trigger keywords
8. Geocoding: Convert addresses to coordinates with validation
9. Output: Generate simplified incident records for dashboard display

The pipeline uses 7 different LLM functions with robust validation and caching to ensure high-quality incident data extraction from scanner audio transcripts.