add_nikud | Add full nikud (vowel points) to Hebrew text, for learners, vocalized display and text-to-speech, where unvocalized Hebrew is ambiguous. English terms and numbers pass through unchanged. The result is verified: `verified: true` means removing the points gives back the input; when false, `nikud` is the input unchanged. Up to 10,000 characters. Uses the monthly character quota. |
add_termsdestructive | Add protected terms to the account, or update existing ones (matched by exact term text). Every later protected translation on the account (translate, translate_batch and the clinical tools) keeps these terms exactly as written, or uses `translations.he` / `.en` / `.ru` for that target language when set. Sending a term that already exists replaces its translations and note. Up to 2,000 terms per account, each up to 200 characters. Changes apply to new requests within 30 seconds. Free. |
analyze_directionread-only | Resolve the right-to-left / left-to-right structure of mixed Hebrew text: base direction, whether it mixes directions, dominant language and script, and the runs to render, where every left-to-right unit (a Latin term, a price, a range such as 10–20%, a date, a model number) stays whole. Send one `text`, or `fields` (a map of name → string, up to 100) to analyze a whole record at once. `base`: `auto` (default), `ltr` or `rtl`. Deterministic and free: no model, no character quota. |
clinical_query | Turn a Hebrew clinical question (up to 2,000 characters) into an English query for a record search or an English model. English terms, numbers and dates are kept exactly, and Hebrew spellings of drug names are mapped back to English, preferring the patient's own terms in `known_terms` and the account terms. Counts 3× toward the monthly character quota. |
clinical_translate | Translate a complete English clinical text (a discharge summary, patient instructions, a visit note, a model's answer) into Israeli medical Hebrew for a clinician or a patient. Drug names, doses, lab values, units, ICD-10 codes, clinical abbreviations, identifiers and every number are kept exactly; every sentence is checked for numbers, negation and laterality and reviewed for meaning. A sentence that fails is retried, then returned in English with status `fallback`: unverified Hebrew is never returned. Treat fallback sentences as untranslated.
`audience`: `clinician` (default, keeps BID/PO shorthand) or `patient` (plain words). Pass the patient's medication and problem lists as `protected_terms`, names as `redact_terms` (hidden from the model), `reader_gender` and `patient_gender` when known, and `date_order` (US records are MDY). Counts 3× toward the monthly character quota. |
clinical_verify | Check Hebrew that was produced some other way (for example by your own model) against its English clinical source: every drug, dose, lab value, code and number must appear exactly, no number may be added, dates must be unambiguous, and negation and laterality must survive. `verification: "strict"` adds an independent model review for meaning changes; `standard` (default) is deterministic. Returns pass or fail with every issue. Counts 3× toward the monthly character quota. |
delete_termdestructive | Permanently delete one account term by its id (from list_terms or add_terms). Protected translations stop using it within 30 seconds. Free. |
detect_languageread-only | Detect the dominant language, script and direction of a string, with one segment per language. Built for short, code-switched input (chat messages, search queries, voice transcripts): a brand or acronym such as "PayPal" or "API" inside a Hebrew or Russian sentence does not change the dominant language. Use it to pick `source_lang` or the reply language before translating. Latin script is tagged `en`. Deterministic and free: no character quota. |
healthread-only | Check that the HebrewCore API is up and its database reachable. Needs no API key and uses no quota. |
list_languagesread-only | List every language HebrewCore translates to and from, with its code (use it as `source_lang` / `target_lang`), English and native name, and writing direction. Protected translation covers Hebrew, English and Russian. Free. |
list_termsread-only | List the account's protected terms: drug, product and brand names that every protected translation on the account keeps exactly as written (or renders with the term's per-language translation when one is set). Returns each term's id, which delete_term needs. Free. |
strip_nikudread-only | Remove nikud, dagesh and cantillation marks from Hebrew text, keeping punctuation such as the maqaf. Use it to normalize text for search keys, comparison and de-duplication, or to get display text back from vocalized text. Deterministic and free: no character quota. |
translate | Translate one text between Hebrew, English, Russian and the other languages from list_languages. Use it for product, support, interface, finance, legal and general clinical text. Returns the translation with its direction metadata.
Protected translation: set `protect: true`, a `domain` (`general`, `clinical`, `finance`, `legal`) or any protection field, and prices, quantities, numbers, dates, codes, identifiers, product names, interface variables ({name}, %s) and `protected_terms` are kept exactly as written, and every sentence is verified; a sentence that cannot be verified comes back in the source language (status `fallback`) instead of a wrong value. `finance` also protects IBANs, card and account numbers and checks credit/debit direction; `legal` keeps clause references, citations and defined terms and checks shall/may modality; `clinical` protects drugs, doses and lab values.
Gender: `listener_gender` is the reader ("you"), `speaker_gender` the speaker ("I", e.g. an assistant). `glossary` maps source terms to required renderings; `redact_terms` keeps names exact and hidden from the model. `message_format: "icu"` translates an ICU MessageFormat string and rewrites plural branches for Hebrew and Russian. `quality`: `fast`, `balanced` (default) or `smartest`.
Uses the monthly character quota. For many short strings use translate_batch; for full English clinical documents into Hebrew use clinical_translate. |
translate_batch | Translate 1 to 50 strings (each up to 5,000 characters) into one target language in a single call; results come back in input order. Use it for the interface strings of a screen, a string file, or the fields of a product.
Takes every option of translate: `protect`, `domain` (`general`, `clinical`, `finance`, `legal`), `protected_terms`, `glossary`, `listener_gender`, `speaker_gender`, `message_format: "icu"` and `quality`. With protection, interface variables such as {count}, {{name}} and %s, numbers, prices and dates stay exactly as written, and each item carries its own verification status. Uses the monthly character quota. |
transliterate | Romanize Hebrew into Latin letters (`direction: "he-to-latin"`, the default), or write Latin spelling in Hebrew letters (`direction: "latin-to-he"`) for name entry and search. Up to 5,000 characters. Uses the monthly character quota. |
wrap_for_displayread-only | Make one mixed Hebrew/English display line render correctly: every left-to-right unit (a Latin term, a price, a number with its unit, a range such as "10–20%", a date, a comparison such as "<5") is kept whole and in order, so nothing is reversed or mirrored inside a Hebrew line. `format`: `html` (dir spans, for web and React), `isolate` (Unicode isolates, for iOS, Android 7+ and browsers) or `marks` (LRM/RLM, for any renderer including older Android). Pass `base: "rtl"` for a Hebrew line that starts with a Latin term. Deterministic and free: no character quota. |
protogrid