Fragments, Scores, And Export

Fragments are the unit you review. Exports are the unit you reuse.

Knowlume is most useful when a source contains a lot of filler. Instead of reading or watching everything, use filters to find the few fragments that are reproducible, non-obvious, mention explicit external sources, or are otherwise worth keeping.

This page explains the core workflow after a source finishes processing. For the first complete run, see Quickstart.

Fragment Anatomy

A fragment usually contains:

  • title;
  • source reference;
  • Markdown body;
  • score values;
  • optional scoring rationale;
  • optional screenshot or image when available.

The Details panel is the authority for review. The list view is optimized for triage.

Score Axes

Each score is an integer from 0 to 10.

Reproducibility

High reproducibility means a reader can verify, repeat, or apply the fragment.

Examples of high reproducibility:

  • a procedure with clear steps;
  • a measurable claim with context;
  • a rule with conditions and exceptions.

Low reproducibility usually means the fragment is vague, motivational, biographical, opinion-heavy, or hard to test.

Use this score when you want material you can transfer into your own work: procedures, playbooks, checklists, implementation details, decision rules, and examples with clear conditions.

Originality

High originality means the fragment is not merely obvious background. Inside Knowlume, it is the public-facing inverse of the internal banality score: lower banality means higher originality.

Examples of high originality:

  • a surprising constraint;
  • a practical distinction;
  • a non-trivial synthesis;
  • a detail that changes implementation or decision-making;
  • rare professional experience;
  • an interdisciplinary connection;
  • a counterintuitive lesson or mistake.

Low originality does not mean wrong. It means the fragment may not be worth exporting.

How this works in epistack:

  1. It extracts a cause, context, and real effect from the fragment.
  2. It tries to keep the cause and context from revealing the real effect directly.
  3. The model sees only the cause and context and predicts five plausible effects.
  4. The real effect is compared with those predictions.
  5. If the real effect is easy to predict, the fragment is more banal. If the real effect is unexpected, specific, or absent from the predictions, the fragment gets higher originality.

If the cause and context already leak the effect too strongly, epistack does not reward that as originality. It conservatively raises banality so the fragment is not treated as novel just because the answer leaked into the prompt.

Sourcesness

High sourcesness means the fragment explicitly mentions external sources or entities you can return to: books, products, services, tools, brands, libraries, articles, websites, or URLs.

Examples of strong sourcesness:

  • the name of a book, article, library, or tool;
  • a website or URL;
  • a product, service, brand, or study;
  • a list of concrete resources worth opening separately.

Low sourcesness does not mean the fragment is unsupported. It means the fragment contains few explicit external mentions or links. A strong practical recommendation can have high reproducibility and originality while still having low sourcesness.

The current epistack sourcesness score is direct: it extracts unique explicit source mentions and URLs, then adds 2 points per mention up to a maximum of 10.

Filter Strategy

Start broad, then narrow.

Recommended first pass:

reproducibility >= 5
originality >= 5
sourcesness >= 5

Use sum mode when exploring. It is forgiving and helps find candidates.

Use product mode when you want strict quality gates. It penalizes fragments that are weak on any one axis.

Filter Combinations

Use different filter combinations for different goals:

Goal Useful filter pattern
Build a practical knowledge base High reproducibility + high originality
Learn a new area Reproducibility first, originality second
Find research leads Originality first, then inspect reproducibility manually
Collect links, tools, and primary sources Sourcesness first
Detect a low-signal source Look for sources with very few strong fragments

For a long video or podcast, a strong source may still have only a handful of excellent fragments. That is normal. The goal is to avoid spending an hour to discover that only three minutes were useful.

Selection Strategy

Select a fragment when it is:

  • useful outside the source;
  • specific enough to quote or act on;
  • clear enough to verify;
  • not a duplicate of a stronger fragment;
  • part of a useful sequence with neighboring fragments.

Skip a fragment when it is:

  • generic;
  • unsupported;
  • too narrow to reuse;
  • redundant;
  • only understandable after reading the full source.

When a fragment is useful but incomplete, inspect the neighboring fragments. Sometimes the best export is a small cluster: the main idea plus the setup, caveat, or example next to it.

Markdown Export

The web app supports this flow:

  1. Select fragments.
  2. Save the selection.
  3. Export Markdown.
  4. Download the artifact or send it through a webhook.

Export output is intended to be pasted into notes, shared with collaborators, or consumed by downstream systems.

Good exports are compact. They should contain enough context to reuse the idea later, but not the entire source transcript or article.

Webhook Delivery

Users can configure an Integration Webhook in settings.

Delivery modes:

Mode Behavior
File only Export creates a downloadable artifact.
Webhook only Export payload is sent to the configured URL.
Webhook and file Both delivery paths are used.

Webhook delivery is asynchronous. Export can succeed even if webhook delivery fails after retries.

Webhook payloads include selected Markdown text, source metadata, export metadata, and file URL when file delivery is enabled.

Public API Export

The public API supports source-scoped Markdown export. A client can:

  1. submit a URL source;
  2. poll source status;
  3. fetch filtered fragments;
  4. queue a Markdown export;
  5. poll export status;
  6. read the file URL or webhook-style payload.

See Integrations And API for request examples.

Common Pitfalls

Do not export before source status is succeeded.

Do not treat scores as absolute truth. Use them as a triage system.

Do not rely on browser folder export unless the browser supports directory selection.

Do not use webhook delivery as the only copy of important content unless the receiving system is monitored.