screencastgen

Pipeline Overview

How pipelines are structured, their shared steps, and orchestration patterns.

Source: screencastgen/pipelines/

Pipeline Architecture

The three document pipelines share a common extraction-to-synthesis flow, then diverge for their specific output format. The visualization pipeline is prompt-driven and bypasses document extraction/TTS.

flowchart LR
    Extract --> Preprocess --> Split --> Chunk --> Validate --> Synthesize
    Synthesize --> Audio[Concatenate audio]
    Synthesize --> Highlight[Align and render highlights]
    Synthesize --> LipSync[Align, animate face, and compose]
    Prompt[Prompt] --> Generate[Generate Manim source] --> Render[Render visualization]

Shared Steps

These are implemented in Pipeline Common and used by the document pipelines:

Step	Function	Module
1. Extract	`extract_and_chunk()`	Extractor
2. Preprocess	(within extract_and_chunk)	Text Processing
3. Split	(within extract_and_chunk)	Text Processing
4. Chunk	(within extract_and_chunk)	Text Processing
5. Validate	`validate_and_collect()`	Text Processing
6. Synthesize	`synthesize_chunks()`	TTS Registry

Pipeline-Specific Steps

Audio Pipeline

Concatenate audio chunks → single file

Highlight Pipeline

Align chunks (word-level timing)
Render frames / Build EPUB

Lipsync Pipeline

Align chunks
Generate lip-sync videos per chunk
Build reader bundle, EPUB, or composite MP4

Visualization Pipeline

Generate a Manim scene from a prompt
Render with a selected visualization provider
Persist MP4 output plus metadata/source artifacts

Request/Result Pattern

Each pipeline accepts a typed request dataclass and returns a PipelineRunResult:

def run_audio_pipeline(request: AudioPipelineRequest, reporter, backend_factory) -> PipelineRunResult
def run_highlight_pipeline(request: HighlightPipelineRequest, reporter, backend_factory) -> PipelineRunResult
def run_lipsync_pipeline(request: LipsyncPipelineRequest, reporter, backend_factory) -> PipelineRunResult
def run_visualization_pipeline(request: VisualizationPipelineRequest, reporter, renderer) -> PipelineRunResult

See Pipeline Types for the request hierarchy.

Event Reporting

All pipelines use PipelineReporter for progress output:

reporter.line(message) — Human-readable console output
reporter.emit(phase, current, total, message) — Structured event
reporter.phase_start(phase, message) — Phase transition
reporter.cancelled() — Host-requested early stop check used by long-running lip-sync work

The web app’s Progress Reporter subscribes to these events.

Module Structure

screencastgen/pipelines/
├── __init__.py          (public exports)
├── types.py             Pipeline Types
├── events.py            Pipeline Events
├── common.py            Pipeline Common
├── audio.py             Audio Pipeline
├── highlight.py         Highlight Pipeline
├── lipsync.py           Lipsync Pipeline
└── visualization.py     Visualization Pipeline