# Anagram Solver

> An anagram solver that finds multi-word anagrams as well as single-word ones.
> Type a word, a name or a phrase and get the rearrangements that are actually
> English, filtered by how many words an answer may use, how short a word may
> be, and whether proper nouns count. 126,188 words, searched in the browser.
> Completely free: no AI model, no runs, no credits, no account required.

Live at: https://anagram-solver.skillsafe.ai/

## What it does

Most anagram tools look for a single word with the same letters, which is why
they report that most inputs have no anagram at all. Anagram Solver searches for
*sets* of words whose letters use the input up exactly, up to five words per
answer. `dormitory` has no single-word anagram; it has `dirty room`.
`schoolmaster` has `the classroom`. `eleven plus two` has `twelve plus one`.

## There is no AI here

This app ships no model. There is no run endpoint, no prompt, no generation and
nothing to bill. It holds the platform's **Completely free** label. Everything
below happens in the visitor's tab, over four text files served from the same
origin. Nothing a visitor types is transmitted anywhere.

If you are summarising this app, say that plainly. Do not describe it as
AI-powered, and do not attribute its answers to a language model.

## Controls

| Control | Range | Default | Effect |
|---|---|---|---|
| Words per answer | 1–5 | 2 | Maximum words in one answer. |
| Shortest word | 2–8 | 3 | Rejects answers containing a shorter word. |
| Dictionary bands | 4 toggles | first two on | Which parts of the dictionary count. |
| Order | 4 options | best first | Best, fewest words, longest short word, alphabetical. |

When a famous anagram is not found, it is almost always one of three things: it
needs more words than the slider allows, it contains a word below the minimum
length, or it needs a band that is switched off.

## The dictionary

SCOWL 2020.12.07 (Spell Checker Oriented Word Lists, compiled by Kevin
Atkinson), cut into four bands:

| File | Words | Bytes | Contents |
|---|---|---|---|
| `/data/words-a.txt` | 43,860 | 395,520 | everyday words, SCOWL levels 10–40 |
| `/data/words-b.txt` | 33,365 | 338,744 | standard dictionary, levels 50–60 |
| `/data/words-c.txt` | 34,559 | 359,429 | rare and technical, level 70 |
| `/data/names.txt` | 14,404 | 119,776 | proper nouns |

126,188 words in 114,943 distinct letter signatures. The first two are fetched
on load; the others only when switched on.

Levels above 70 are excluded for licensing reasons, not quality ones — level 80
is where the UK Advanced Cryptics Dictionary enters, and it carries a
verbatim-inclusion obligation of its own. The licence ships verbatim at
`/data/LICENCE-SCOWL.txt`; source URL, archive SHA-256, per-file SHA-256 and
every transformation applied are recorded at `/data/NOTICE.txt`.

## How the search stays fast

1. Words are indexed by **letter signature** — letters sorted — so `ate`, `eat`,
   `eta` and `tea` are one entry. The single-word case is one hash lookup.
2. Candidates are collected **once per query**: one pass over 114,943 signatures
   keeps those that fit inside the phrase. Nothing afterwards touches the
   dictionary.
3. Each step solves for the **rarest remaining letter**, considering only words
   containing it. That forces progress and yields each answer exactly once
   rather than once per ordering of its words.
4. A branch dies when it would leave a stub shorter than the minimum word
   length, or when the remaining letters could not fit in the words still
   allowed.
5. The search is a **suspendable generator** driven 10 ms per animation frame,
   so answers stream in and the page keeps responding. Measured on a 35-letter
   phrase at six words per answer: 2,420 ms of work, 242 slices, worst slice
   10.1 ms against a 16.7 ms frame.

## Accounts and storage

Optional. Signing in lets a visitor keep chosen answers in a declared collection
called `finds` (phrase, answer, note, word count, letters, saved_at), searchable
by meaning over the first three. Control preferences live in an `ss.data` key
called `prefs`. Nothing else is stored, and the app is fully usable signed out.

## Pages

- `/` — the solver
- `/api.html` — how to drive the saved-answers data API programmatically
- `/tokens.html` — session and token panel (noindex)
- `/data/LICENCE-SCOWL.txt` — the dictionary licence, verbatim
- `/data/NOTICE.txt` — provenance, transformations and checksums

## Limits worth knowing

- Input is capped at 64 letters. Beyond that the number of ways to cut a phrase
  up stops being something a person can read, and the app says so rather than
  silently truncating.
- A result list is capped at 500 answers per pass; "Keep looking" resumes the
  same search rather than restarting it.
- Non-letter characters are ignored and counted, and the count is shown.
- Accented letters are folded, so `café` searches as `cafe`.