Search

One search box over the whole workspace — papers, notes, report sections, experiments, milestones, log entries, PDF text and annotations. It runs in the browser, over an index built from a snapshot of your project, so it works offline and never sends a query anywhere.

Typing it wrong still finds it

Spelling is not a precondition. Measured against a corpus of real paper titles and seventeen plausible misspellings, search recovers sixteen:

you type	you get
attension	Attention Is All You Need
disentangelment	Disentanglement reading cluster
nerual	Neural Message Passing for Quantum Chemistry
abaltion	Graph-prior module ablation
gan	Generative Adversarial Networks
seuqence	Sequence to Sequence Learning…

Three mechanisms, in the order they apply.

1. Graduated fuzziness

Edit distance is allowed as a fraction of term length: none below four characters, 0.1 to five, 0.2 beyond. Short terms get none deliberately — at three characters most of the dictionary is one edit away, so tolerance there returns noise instead of the thing you meant.

2. One more edit, but only when nothing matched

A transposition — nerual for neural, attnetion for attention — is a single Damerau edit but two Levenshtein edits, and Levenshtein is what the index measures. A six-letter term is allowed one edit, so it can never reach its own transposition.

Raising the tolerance everywhere would fix that and make every working query worse, by pulling unrelated documents into results that were already correct. So the escalation is conditional: if a query returns nothing, it is retried once with one more edit per term. A query that found something is never touched, and the retry costs nothing on the path that matters — measured at 3,000 documents, matching queries are unchanged and a non-matching query costs 0.56 ms including the second pass.

Long queries are exempt. Five words that match nothing usually mean the thing is genuinely absent, and loosening every term returns a pile of weak matches rather than the one right answer.

3. Acronyms

gan is not a misspelling of Generative Adversarial Networks — it is three characters against ten, unreachable by any edit distance, and it is how people actually refer to papers. Initialisms are therefore derived from each title at index time and stored in aliases, which is already a boosted searchable field.

The derivation skips grammar words (so it is gan, not ganfi), stops at a colon (paper titles are overwhelmingly SHORTNAME: the long descriptive part, and the initials of the second half are not a handle anyone uses), and also picks up an initialism the title already contains — β-VAE, BERT, DDPM — which is one word and would never survive the word-count rule on its own.

What it does not do

• No "did you mean". The retry silently widens rather than suggesting a correction, so there is no second click.
• No stemming. training does not match trained. Fuzziness covers some of this by accident, not by design.
• A typo inside a three-character acronym is not recovered — vea will not find β-VAE. Tolerance at that length matches too much to be worth it.

Query syntax

kind:note, tag:vae, path:chapter-3, "exact phrase", and -excluded. Filters combine with terms; a query that is only filters enumerates rather than ranks.

Ranking

BM25 over boosted fields — title 8, aliases 6, headings 3, tags 3, path 2, body 1 — multiplied by a document signal: link degree first, recency only as a tiebreak. Field boosts and kind weighting are user settings.

An optional semantic arm can be turned on in Settings; when present its results are fused with the keyword arm rather than replacing them.

Where it lives

Ranking and tuning	search-tuning.ts
Acronyms	search-acronyms.ts
Retry policy	search-retry.ts
The one file that imports MiniSearch	minisearch-index.ts

The ranking rules live in @weaveforge/core deliberately, so they are testable without a search library and can be exposed as settings. SEARCH_SCHEMA_VERSION must be bumped whenever the indexed shape changes — a cached index from an older shape is discarded rather than migrated, because a subtly mismatched index returns wrong results indefinitely with no visible symptom.