# Architecture & Pipeline Comparison

TubeTranscript provides low latency data extraction for AI Agents: captioned videos skip audio transcription and LLM parsing entirely, and videos without captions fall back to speech-to-text.

---

## Extraction Pipelines

When an AI Agent needs actionable JSON from a long YouTube video, latency and token costs are the primary bottlenecks. TubeTranscript routes each video through one of two pipelines:

| Pipeline | When it runs | How it works | LLM tokens used for extraction |
|---|---|---|---|
| **Tier-1 Fast Path** | The video has captions | Reads the existing caption track and normalizes it into timed chunks | None |
| **Tier-2 Whisper Fallback** | The video has no captions | Transcribes the audio with Whisper, then normalizes it into timed chunks | None |

---

## Pipeline Characteristics

- **Throughput:** High throughput edge processing of unstructured YouTube noise via our Edge Middleware.
- **Signal-to-Noise Ratio:** TubeTranscript strips non-semantic HTML and audio noise, delivering **pure RAG-ready JSON payload**.
- **Reliability:** Automatic fallback to the Tier-2 pipeline when native captions are unavailable. Transcripts are generated automatically and may contain errors.

For enterprise scale operations, refer to our [API Reference](/docs/api) to integrate this pipeline into your automated workflows.