Features

This reference explains how the major subsystems work and what to expect from each one.

Document Processing Pipeline

LegalEase processes documents through a multi-stage pipeline to make them searchable and reviewable.

Processing Flow

1. Upload - Files are uploaded to Firebase Storage under users/{userId}/documents/{documentId}/
2. Trigger - A Firestore trigger detects the new document and starts processing
3. Extraction - Docling extracts text, structure, tables, and bounding boxes. OCR runs automatically for scanned pages.
4. Chunking - Content is split into hierarchical chunks:
   • summary - Coarse overviews for jumping into large files
   • section - Medium sections (~500 tokens)
   • paragraph - Smaller blocks for precise matching
5. Embedding - Gemini generates dense vector embeddings for each chunk
6. Indexing - Vectors are stored in Qdrant with metadata for hybrid search
7. Page Renders - PDF pages are rendered as images for the viewer

Document Viewer

The built-in viewer provides:

• Page-by-page navigation with thumbnails
• Search hit highlighting using extracted bounding boxes
• Text layer overlay for copy/paste
• Entity sidebar showing extracted people, organizations, dates
• Chunk metadata for debugging

Supported Formats

Transcription Pipeline

Audio and video files are transcribed using frontier AI models with speaker diarization.

Providers

LegalEase supports two transcription providers:

Gemini 2.5 Flash (default)

• Works with Firebase Storage emulator
• Automatic speaker diarization via prompt engineering
• Speaker name inference from conversational context
• Supports files up to 9.5 hours
• Structured JSON output with timestamps

Google Speech-to-Text (Chirp 3)

• Requires production GCS (not compatible with emulator)
• Native speaker diarization
• Higher accuracy for some audio types
• Supports files up to 8 hours via BatchRecognize

Processing Flow

1. Upload - Audio/video uploaded to Firebase Storage
2. Trigger - Firestore trigger starts transcription job
3. Provider Selection - Based on TRANSCRIPTION_PROVIDER setting
4. Transcription - Provider generates segments with:
   • Start/end timestamps
   • Speaker identification
   • Transcript text
   • Confidence scores (Chirp only)
5. Speaker Inference - Names inferred from conversation
6. Summarization - Gemini generates summary, key moments, entities
7. Waveform - Audio peaks extracted for visual player (in progress)

Output Format

{
  fullText: string,
  segments: [{
    id: string,
    start: number,    // seconds
    end: number,
    text: string,
    speaker: string   // "Speaker 1", "Speaker 2", etc.
  }],
  speakers: [{
    id: string,
    inferredName?: string  // "John", "Jane", etc.
  }],
  duration: number,
  language: string,
  summarization: {
    summary: string,
    keyMoments: [...],
    actionItems: [...],
    topics: [...],
    entities: {...}
  }
}

Summarization

All transcripts are automatically analyzed using Gemini 2.5 Flash.

Output Components

Executive Summary

• 1-2 paragraph overview of the conversation
• Focus on key facts and outcomes

Key Moments

• Timestamped highlights with importance ratings (high/medium/low)
• Click to jump directly to that point in the audio

Action Items

• Follow-up tasks mentioned or implied
• Extracted automatically from conversation

Topics

• Main subjects discussed
• Useful for categorization and filtering

Entities

• People - Names mentioned in the conversation
• Organizations - Companies, agencies, firms
• Locations - Places referenced
• Dates - Dates, deadlines, timeframes mentioned

Hybrid Search

LegalEase combines semantic and keyword search using Qdrant vector database.

How It Works

1. Query Processing
   • Gemini generates dense vector embedding
   • BM25 generates sparse keyword vector
2. Search Execution
   • Both vectors query Qdrant simultaneously
   • Results from each method are retrieved
3. Fusion
   • Reciprocal Rank Fusion (RRF) combines results
   • Balances semantic understanding with keyword precision
4. Filtering
   • Results filtered by case, document type, date range
   • Permissions applied based on user/team

Search Modes

Indexed Content

• Document chunks (all granularities)
• Transcript segments
• Summaries and key moments
• Entity mentions

Firebase Integration

LegalEase leverages Firebase for a serverless, real-time architecture.

Services Used

Cloud Firestore

• Primary database for cases, documents, transcripts
• Real-time listeners for instant UI updates
• Automatic offline support

Firebase Storage

• File storage for uploads
• Secure, authenticated access
• Resumable uploads for large files

Cloud Functions

• Genkit-based AI flows
• Firestore triggers for background processing
• Scales automatically

Firebase Auth

• Google sign-in
• Email/password authentication
• Session management

Real-Time Updates

All data uses Firestore real-time listeners:

• Document processing status updates instantly
• Transcript completion triggers immediate UI refresh
• Team collaboration sees changes in real-time

AI Provider Architecture

LegalEase uses a provider abstraction for flexibility:

functions/src/transcription/
├── provider.ts      # Interface definition
├── types.ts         # Shared types
├── registry.ts      # Provider registration
├── index.ts         # Public API
└── providers/
    ├── gemini.ts    # Gemini 2.5 Flash
    └── chirp.ts     # Google Speech-to-Text

Adding New Providers

1. Implement TranscriptionProvider interface
2. Register in registry.ts
3. Provider automatically available via TRANSCRIPTION_PROVIDER env var

This pattern extends to future AI providers (OpenAI, Anthropic, local models).