JEV Prefiltering | summarize

By default, the app sends the full transcript to the summary model. JEV prefiltering runs after the transcript cache and selects original passages before summarization, so the model sees a shorter transcript. The summary provider, prompt, and cached transcript stay unchanged.

How to enable it:

CLI: Add --use-jev-prefiltering to your command
Streamlit GUI: Check “Use JEV prefiltering” in the left sidebar

The scoring provider and selection fields appear only after it is enabled. Visual mode and grammar-correction prompts bypass it.

Why JEV Prefiltering Exists

Long transcripts are expensive to summarize and often contain filler, sponsorship, or material you do not want in the result. JEV scores passages before the summary model and keeps the original wording of the passages that match.

Use cases where JEV prefiltering shines:

Podcasts — keep only a particular concept about this video to filter

Sponsor-heavy videos — exclude sponsorship and self-promotion

Long lectures — send the summary model roughly the most relevant 35% of each chunk

Repeated runs — change the selection without downloading or transcribing again

Selection Fields

Field	Blank	Example
I want to include only…	Keep generally relevant passages	`a particular concept about this video to filter`
I want to exclude…	No additional exclusions	`Sponsorship and self-promotion`

An included subject can be selected even when it is not the video’s main topic. Exclusions take priority, including in mixed passages.

Example Config

JEV reuses an existing OpenRouter or TypeSafe provider and its API key. The default provider is openrouter; no separate provider entry is required. Scoring calls jev-latest on /systemone, not the provider’s chat model.

defaults:
  use-jev-prefiltering: true
  jev-provider: openrouter
  jev-include: "a particular concept about this video to filter"
  jev-exclude: "Sponsorship and self-promotion"
  jev-keep-ratio: 0.35

CLI Examples

# General compression
python -m summarizer --source "URL" --use-jev-prefiltering
# Keep one concept and drop sponsorship
python -m summarizer --source "URL" \
  --use-jev-prefiltering \
  --jev-include "a particular concept about this video to filter" \
  --jev-exclude "Sponsorship and self-promotion"
# Turn it off for one run
python -m summarizer --source "URL" --no-use-jev-prefiltering

HTTP JSON, upload, and batch requests accept use_jev_prefiltering, jev_provider, jev_include, and jev_exclude. An empty string clears a rule set in YAML.

Selection Limits

Default retention ceiling: about 35% of each original chunk
Any summary chunk size: JEV splits oversized scoring work into bounded requests without changing your summary chunk size
Scoring provider: existing OpenRouter or TypeSafe provider
One or more scoring requests per eligible chunk, depending on its size

jev-keep-ratio is a ceiling, not a quota. Irrelevant text is not added to fill unused space. Set jev-keep-ratio: 1.0 to keep every qualifying passage.

If nothing matches, the summary model is not called. When enabled, JEV scores every non-empty text chunk, including single-unit chunks, regardless of its configured summary chunk size.

Explicit include or exclude rules stop the run if scoring times out, returns an error, or a scoring unit cannot fit in a request. They do not send the unfiltered transcript to the summary model. With both fields blank, the original chunk is used after a warning.

Selected passages stay in their original order. A mixed passage is kept or dropped as a whole, so an exclusion can also remove neighboring useful text. Compare an important result with JEV disabled.