Metadata Prior Video Gating

Context

Summry is narrowing its product promise toward knowledge-dense YouTube videos. Gate 1 added durable fields for eligibility, content category, transcript quality, processing strategy, and an explanatory reason. Gate 2 starts using those fields before the worker sees a transcript.

This step is intentionally metadata-only. It looks at the title, description, channel name, YouTube categories, and tags that Summry already fetches during video submission.

Problem

Before this gate, every new video still entered the system with generic unknown gating state. That was useful as schema groundwork, but it did not yet help the product distinguish likely tutorials, explainers, interviews, lectures, commentary, reviews, or obvious out-of-scope videos.

The challenge was to add a first prior without overclaiming. Metadata is useful, but it is noisy. A title can be misleading, a YouTube category can be broad, and many useful videos live in generic categories.

Implementation

This gate adds a service-level metadata classifier. It uses deterministic keyword and category heuristics to produce:

• eligibility status
• content category
• processing strategy
• eligibility reason

New videos persist this prior during creation. Existing video reuse, duration validation, chat room creation, processing enqueue behavior, and worker execution remain unchanged.

The classifier does not hard-reject anything at submission time. Even when metadata looks unsupported, the result is stored as a prior for later gates instead of blocking the current flow.

Tradeoffs

The main benefit is observability. Summry now has an early, inspectable guess about what kind of video a user submitted.

The tradeoff is precision. Metadata-only classification will be wrong sometimes, so this gate should not be treated as the final product decision. Transcript quality and knowledge-density checks still need to confirm or correct the metadata prior in later gates.

Lessons Learned

The right first classifier is modest. A deterministic prior is cheap, testable, and easy to reason about. It gives the next gates useful state without adding an LLM call or changing processing reliability.

Keeping this logic in a service module also preserves the route-service-repository shape: video creation still owns orchestration, the repository still owns persistence, and the gate service owns classification behavior.

Next Steps

Gate 3 should analyze transcript quality after captions or transcription are available. That will let Summry distinguish videos that look promising in metadata from videos whose transcript is too weak, visual-heavy, repetitive, or incomplete to summarize confidently.