Workspace

Config

Manual ingestion default

…

Strategy Controls how manually ingested assets are segmented before evidence is attached and scored.

Evidence extractors

Chooses which evidence tracks are created for this default. These tracks determine what search, scoring, and comparison can use later.

Transcript Speech text Visual description Scene description Audio cue Music and sound events

Runtime tuning

Evidence and corpus

These values affect corpus readiness, comparison tie handling, and how much each evidence track contributes to evidence-aware scores.

Tier-1 n Default 20. Minimum scored assets needed in the same duration bucket before corpus percentile calibration is trusted. Example: at 30, a bucket waits for 30 comparable videos; lower values make percentiles appear sooner but noisier. Need states Default 6. Minimum distinct need states that must have data in a bucket before the corpus is considered broad enough. Example: if only 3 need states have scores, the bucket may be biased toward too narrow a taxonomy slice. Signal quantile Default 0.75. Cutoff used during corpus refit to count how many assets in a bucket have strong retrieval signal. Example: 0.75 counts upper-quartile matches as signal-bearing; 0.90 is stricter, 0.50 is more permissive. Tie epsilon Default 0.005. Cosine window used by Compare to find near-equal scores around the asset score. Example: 0.812 and 0.809 are tied. Increase if rankings look falsely precise; decrease if too many cells become tie bands. Tie spread Default 3. Minimum percentile span required before a near-equal cosine cluster renders as a tie band. Example: with 3, close scores spanning 3 percentile points show as a band; raising to 5 makes bands rarer.

Score shown in reports and compare Default weighted aggregation. This is the score source used for Compare immediately and for newly published insight reports. Existing report.json files keep the score they were published with until republished.

Weighted aggregation: scores segments with embedding similarity, then rolls them up by duration. Full-video score = longer scenes contribute more than shorter scenes.

Evidence-aware: scores each available evidence track, combines tracks by the weights below, then rolls up by duration. If only Transcript and Visual description exist, the score is normalized over those two weights. Use after evidence-aware scoring has run, otherwise pages may have no score for that source.

Evidence weights

Relative influence of each evidence track in evidence-aware scoring. Higher weights make a track count more in matching and ranking. Example: raising Transcript makes spoken claims matter more than visuals; raising OCR makes on-screen titles and captions matter more. Missing evidence is skipped and the remaining available weights are re-normalized.

Vision need-state scorer Runs direct 0-1 scene scoring against all active RMT need states. Max need states per batch Default 15. Batches are grouped by motivation and split deterministically when needed.

Updated: …

Config

Evidence and corpus

Visual extractor

Need-state scorer