Using Semantic Search

Assisted vs. Automatic search modes, relevance score display, and how ranking behaves inside every Writer Tool.

Where Semantic Search Appears

Semantic Search appears on pages that involve searching over text — anywhere it's useful to surface relevant material from your My Knowledge library as you type. Such a page debounces for about 600ms after you stop typing, then ranks your entire library by relevance to what you've written so far, surfacing the best matches in a picker.

Semantic Search in Writer Tools

Today, the screens that implement this pattern are the Writer Tools (Blog Writer, Essay Writer, Article Writer, Storyteller, Anime Writer, Scenario Writer, Smart Rewriter). Each has an Attach Knowledge button above its main text field, and as you type your topic, description, or source text, your sources are ranked in real time.

Search Modes

Assisted (default)

Opening Attach Knowledge shows your sources ranked by relevance, with a percentage badge on each. You still tick the checkboxes yourself — nothing is attached without your confirmation.

Automatic

The top-ranked sources (above a relevance threshold, capped at 3) are attached automatically as you type. A chip reading "N sources auto-attached" appears near the text field. You can still open Attach Knowledge at any time to review or override the selection — doing so switches that screen back to a manual selection for the current session.

Switch between modes in Settings → Agentic → Embedding Engine → My Knowledge Search Behavior. The setting applies globally across all Writer Tool screens.

Relevance Score Display

The percentage badge shown next to each ranked source in the Attach Knowledge picker can be displayed two ways, set in Settings → Agentic → Embedding Engine → Relevance Score Display:

Cosine Similarity (default)

The raw, unmodified similarity score between your topic and the source. Real-world scores tend to cluster in a narrow band — roughly 20-40% for unrelated content and 55-75% for genuinely related content, since even unrelated text shares common language structure in embedding space. Scores above 90% are rare.

Normalized (0-100% spread)

Min-max scales the current batch of scores so the most and least relevant sources in view span the full 0-100% range. This is purely a display transform for a more visually distinguishable spread — it does not change which sources rank highest, and it has no effect on the Automatic mode's attachment threshold (which always uses raw cosine similarity).

Reviewing or Overriding a Selection

Whether in Assisted or Automatic mode, clicking Attach Knowledge always opens the same picker: sources ranked by relevance (if an embedding model is configured), with checkboxes reflecting the currently attached set. Check or uncheck any source and click Confirm to update the attachment — this always takes precedence over whatever Automatic mode had selected.

When Ranking Is Unavailable

If no embedding model is configured (see Embedding Model), the Attach Knowledge picker shows a note explaining that ranking is unavailable and lists your sources in their original, unranked order. Generation and manual attachment are never blocked by this — only the ranking and Automatic auto-attach features require an embedding model.

FAQ

Does switching to Automatic mode delete my manual selections?

No. Opening Attach Knowledge and confirming a different selection always overrides whatever Automatic mode had auto-attached for that generation.

Why do unrelated documents still show a relevance score above 0%?

This is expected behavior for cosine similarity between text embeddings — even unrelated text shares general-language structure, so scores rarely approach 0%. See the tooltip next to Embedding Engine in Settings for the full explanation.

Does changing the Relevance Score Display setting re-rank my sources?

No. It only changes how the existing ranking is displayed (raw vs. normalized percentages) — the underlying order of sources by relevance is unaffected.