Power Search

Power Search exposes the full query grammar of The Word Index in one form: positional patterns, a bounded regular expression, repeated-letter variables, letter budgets, length, sound, and source scopes, combined into a single query. Build it below or write raw JSON, watch the live match count, then run it against this build's compiled lexicon.

New to the grammar? Read the query syntax reference generated from the engine's query-language contract.

Spelling & structure
One token per position: letters, _ unknown, [abc] one-of.
Anchored full match, bounded grammar (max 120 chars, guarded quantifiers, no backslashes).
Uppercase letters repeat the same letter everywhere; lowercase are literals; _ is one unknown.
Subset matches: no letter used more often than supplied.
Letters
Repetition counts: eel needs two e and one l.
Length
Sound (needs pronunciation evidence)
One digit per syllable: 0 unstressed, 1 primary, 2 secondary.
Stress-stripped ARPABET symbols; one pronunciation variant must contain them all.
Scope & order
Tiers (build-specific bands)
Require source evidence

Add a constraint to see a live match count.

Result export follows the site's licence posture and is currently unavailable; the exact reason from the engine appears here after a query runs.

About Power Search

Every named finder on this site is a preset over one typed query engine. Power Search removes the presets: any combination of the engine's constraints runs as a single query, with the same normalization, the same deterministic ordering, and the same match counting the other tools use. The syntax reference is generated from the engine's query-language contract, so what it documents is what runs.

Every query reports its execution plan honestly. Constraints such as length, pattern shape, anagram sets, prefixes, sounds, and source membership can narrow the scan through compiled indexes; a query offering no such handle — a regular expression with unbounded length, for example — is answered by scanning every record in the selected pool, which is the slowest latency class. The plan and measured server time appear beside each count and result set.

Data sources and bounds

  • Matches are records in this build's compiled lexicon, assembled from pinned source lists. Membership means a source contributed evidence — not that a spelling is universally valid English or accepted by any particular game.
  • Commonness tiers are build-specific ranking bands derived from the pinned frequency artifact, not universal usage claims.
  • The regular-expression grammar is deliberately bounded: at most 120 characters, literals a–z with classes, groups, alternation, and quantifiers, repetition bounds up to 45, no backslash escapes or extension groups, and an ambiguity budget. Malformed or unsupported syntax is rejected with an error rather than silently broadened.
  • Letter-variable patterns support at most 10 distinct variables and imply an exact length.
  • Sound constraints (rhyme, syllables, stress, phonemes) match only records with retained CMUdict pronunciation evidence; words without pronunciation evidence never match them.
  • Contradictory constraints are rejected before execution; a valid query with zero matches returns an empty set with recovery suggestions, never a silently loosened query.
  • The server limits concurrent heavy queries; under load it can answer 503 with a retry hint rather than queueing unbounded work.
  • Results are windowed: 50 rows per page, at most 200 rows here for one query state, and API offsets up to 10,000.
  • One-letter-change neighbors come from a derived graph over words of 3–7 letters in tier bands 1–3 of this build; words outside those bounds report no neighbors.
  • CSV export is wired to the engine's result-export decision. While redistribution approval is pending for one or more contributing sources, the engine reports export as unavailable and the button stays off; visible results do not create a right to republish the compiled list.
  • The offensive-word display policy is applied to results and glosses by default.

Results reflect data version v1/2026.07-lexical.6; counts, tiers, and plans can change between builds.