spellkit-general-medical

A biomedical dictionary pack for spellkit: drug, condition, gene and target vocabulary merged with general English, so one checker understands both acetaminophen and the.

# Gemfile
gem "spellkit"
gem "spellkit-general-medical"

# config/initializers/spellkit.rb
SpellKit.enable_dictionary(:general_medical, lazy: true)

SpellKit.correct("acetaminphen")   # => "acetaminophen"
SpellKit.correct("flexerol")       # => "flexeril"
SpellKit.correct("the")            # => "the"      (ordinary English is left alone)
SpellKit.correct("CDK10")          # => "CDK10"    (protected, never altered)

The gem registers itself when it loads, so naming it in your Gemfile is the whole setup. The data ships inside the gem — no network, no cache, works air-gapped and in CI.

An initializer runs in every process that boots the app — web, rails db:migrate, rake, console, sidecars — not just the one that searches. This pack at its default tuning costs ~2.1 GB resident and ~5.5s to index, and loading it eagerly has already OOM-killed a memory-constrained migrate init container in production.

Pass lazy: true, and warm it explicitly where you do want it:

# config/puma.rb
on_worker_boot { SpellKit.load_dictionary! }

stats and healthcheck deliberately don't trigger the load, so a liveness probe can't build the index in the process lazy protects. Full details in spellkit's README.

SymSpell's deletion index grows steeply with distance. If you are memory-constrained, this matters far more than lazy loading does:

SpellKit.enable_dictionary(:general_medical, lazy: true, edit_distance: 1)

Note edit_distance: 1 is less likely to corrupt a term — it trades recall, not safety.

Measured against a held-out corpus of real consumer misspellings (NLM's Consumer Health Vocabulary), scored over the same denominator as a generic English spellchecker:

403 misspellings are fixed only by this pack — brand names and newer drugs (flexeril, semaglutide, pembrolizumab) that no general word list carries. It also rewrites 0.0% of the thousand commonest English words.

Caveats, because the numbers flatter the pack by construction:

• A corpus of drug misspellings will always favour a drug dictionary. What makes the result meaningful is the pairing with the 0.0% English-mangling figure — gain on domain input, no measured cost on general input.
• Tuning was measured on drug misspellings only; the same constants are applied to condition and gene vocabulary, for which no held-out corpus exists.
• ~9.7% of corrections are wrong. Many are genuine ambiguity between real alternate spellings (bromocriptin vs bromocriptine), but a wrong correction is still wrong. Raise frequency_threshold if your surface prefers silence.
• Not yet validated against real query traffic, which is the only distribution that ultimately matters.

Built from RxNorm and MeSH (public domain, NLM), Open Targets/Ensembl gene symbols (CC0), and SymSpell's en-80k English frequency list (MIT, © Wolf Garbe; derived from Google Books Ngram data, CC BY 3.0). See ATTRIBUTION.md for the full table and the NLM disclaimer.

NLM's Consumer Health Vocabulary is not in this artifact — it ships only inside the gated UMLS Metathesaurus release and may not be redistributed. It is used locally as the held-out validation corpus; only the tuning constants and aggregate figures leave that process.

See RELEASING.md. Requires Substrate warehouse access for the source export and your own UMLS licence for the validation corpus.

bundle install
bundle exec rspec
bundle exec standardrb