Knowledge-Density Eligibility Gate
Summry now finalizes video eligibility before chunking by combining metadata, transcript quality, and transcript sample signals.
Context
Summry's gating work is moving from passive signals toward a real product boundary. Gate 1 added durable fields. Gate 2 created a metadata prior. Gate 3 checked transcript quality. Gate 4 combines those signals into final eligibility before the expensive summarization path starts.
The goal is not to summarize every video. The goal is to be honest about whether Summry has enough evidence to produce a useful knowledge-video summary.
Problem
Before this gate, unsupported metadata and weak transcript signals could still flow into chunking and summarization. That made the product less reliable and treated clear scope failures like ordinary retryable processing failures.
Unsupported content should not keep retrying. It should stop once, explain why, and avoid spending more processing work.
Implementation
This gate adds final knowledge-density eligibility after transcript-quality analysis and before chunking.
The final decision combines:
- the metadata prior from title, description, category, channel, and tags
- the transcript-quality label
- a transcript sample from the beginning, middle, and end of the video
- lightweight knowledge-density and category signals from the transcript text
Supported and caution videos continue to chunking and summarization. Unsupported videos are marked failed with a product-scope reason, emit a terminal SSE event, and skip worker retries.
Tradeoffs
The main benefit is a cleaner product boundary. Summry can now stop obvious out-of-scope videos before generating a low-value summary.
The tradeoff is that the first knowledge-density check is still heuristic. It will need calibration with real examples, especially for videos whose metadata is vague but whose transcript is genuinely useful.
Lessons Learned
The worker needed to distinguish product-scope failures from infrastructure or processing failures. Retrying a video because it is outside the product promise does not improve the outcome. It only wastes processing time and creates confusing progress behavior.
Separating terminal eligibility failures from retryable failures makes the processing lifecycle more honest.
Next Steps
Gate 5 should use the persisted processing strategy to route summarization prompts by video shape. That is where supported tutorials, explainers, lectures, interviews, commentary, and reviews should start receiving different summary instructions.