World Search

World Search lets you ask questions about your own work and get answers grounded in what you’ve actually written. It indexes your manuscript passages, entity profile sections, and author memory entries so the chat can retrieve relevant material and cite specific sources back at you.

What gets indexed

Scene prose from your active book
Entity profile sections (characters, locations, props, lexicon — anything with a profile)
Author memory entries

AI tool outputs (scene analyses, arcs, findings) are not indexed directly. Anything you accept into an entity profile will be picked up the next time that profile is indexed.

How it works

When you ask a question, Wonderthrough turns it into an embedding vector and finds the most similar chunks from your project’s index.
Those chunks are passed to the AI model alongside your question.
The response cites the sources it used. Clicking a citation opens the relevant scene passage, entity section, or memory entry.

When should I use World Search instead of chat?

Use World Search when you want to locate information that is already in the project. It is strongest when the real question is “where is this?” rather than “what should I do?”

World Search is usually better when you want to:

find where a fact was mentioned before
check whether a character, place, or idea appears in specific scenes
confirm continuity details against existing text
gather cited evidence before making a creative decision

Chat is usually better when you want brainstorming, interpretation, planning, or editorial guidance. In practice, many good workflows use both: search first to recover the evidence, then chat to think through what it means.

Availability

World Search requires a Pro or Pro+ subscription, or a top-up credit balance. This applies regardless of which preset you use — the subscription is for access to the feature, not the AI cost.

Your search index builds automatically in the background as you write, even before you subscribe. If you subscribe later, search is ready immediately.

Privacy presets

World Search behaves differently across the three quick-set presets (configured in Settings → Privacy & Data). Each preset configures the four underlying atomic settings — where AI calls go, whether project content syncs to the cloud, and so on.

Standard

AI calls: embedding and generation use Wonderthrough’s platform keys, routed through our servers.
Search index: lives where the project’s content lives. With cloud sync on for the project: stored in our cloud database, scoped per-project with row-level security. With cloud sync off (desktop app only): stored locally on the device.
Diagnostic logging: a per-query diagnostic row (hash of the query — never raw text — plus shape metadata, latencies, model, status). The row honors Writing activity details: Off skips it entirely; Minimal drops the chunk identifiers (which entities were retrieved or cited); Robust keeps them. No query text, chunk text, or generated response is ever logged at any level. Sync activity between desktop and web then governs how the captured row flows — on the desktop app, it gates whether the row uploads at all (off keeps it in the local file); on the web, the row flows to our cloud regardless (web is cloud-by-construction) and this setting controls the cross-device view at read time instead. Chunk identifiers are derived from entity UUIDs already held by Wonderthrough; aggregated over time, they reveal which entities you searched most — which is why Minimal strips them.
Where it runs: web app and desktop app.

Private AI

AI calls: your own OpenAI key for embeddings, your own Anthropic key for generation. Calls go directly from the desktop app to each provider; your prompts, writing, and responses never touch Wonderthrough’s servers.
Search index: stored locally on the device in LanceDB, under Wonderthrough’s userData directory.
Diagnostic logging: none. Nothing about your queries is logged to Wonderthrough.
Setup: make sure BYO is checked under Available AI engines in Privacy & Data (the Private AI preset turns Wonderthrough’s keys off, so BYO needs to be available). Add both an OpenAI and an Anthropic API key in Settings → Keys & Models, then pick the Private AI preset in Privacy & Data.
Where it runs: desktop app only. On the web, even with your own provider keys configured, calls pass through our proxy — for AI calls that go directly from your device to your provider without us seeing the prompts or responses, use the desktop app.

Local-only

AI calls: local models via Ollama on your machine. No cloud provider is contacted.
Search index: stored locally in LanceDB, same as Private AI.
Diagnostic logging: none.
Setup: install Ollama, then pull an embedding model (ollama pull nomic-embed-text) and a generation model of your choice. Pick the Local-only preset in Settings → Privacy & Data.
Where it runs: desktop app only. Works best on Apple Silicon or a machine with a discrete GPU.

Cross-device tradeoff

The search index lives where the project’s content lives. With cloud sync on for the project (Pro or Pro+), both your manuscript and index are in our cloud database — available from any device where you’re signed in. With cloud sync off, or on the Private AI / Local-only presets, the index lives on the specific device where it was built. Opening the same project on another computer will trigger a fresh index build on that machine.

This is a deliberate consequence of the principle: your index never ends up somewhere your content isn’t.

Managing your index

Delete a project — when cloud sync is on for the project, this cascades and removes the cloud index rows; in all other configurations, the local index stays on disk until you remove it.
Clear a local index manually — remove the project’s folder under Wonderthrough’s userData directory at search-index/[projectId]/.

What’s never logged (any preset)

The raw text of your query
The text of any chunks returned to the model
The generated response
Any identifiers linking your queries to profiles outside Wonderthrough

See the Privacy Policy for the full picture.

For when to brainstorm or interpret instead of searching, see Multi-Lens Chat.
For the privacy presets that decide where your index lives, see Privacy & Data.
For structured, repeatable analysis of your manuscript, see Advanced Analysis.