Sapling

Grammar and spelling check

Check text for grammar, spelling, and style issues. Returns a list of suggested edits, each with the sentence, character span (start/end), replacement text, and an error type. Set auto_apply=true to also receive applied_text: the input with all suggested edits already applied (useful to return a corrected document in one call). Supports many languages via the lang parameter.

Spelling-only check

Check text for spelling errors only (no grammar or style suggestions). Returns edits in the same format as sapling_grammar_check. Faster and cheaper than a full grammar check when only misspellings matter.

AI-generated text detection

Score how likely a text is to be AI-generated. Returns an overall score between 0 (likely human-written) and 1 (likely AI-generated), plus per-sentence scores by default. Provide at least ~50 characters of text for a meaningful score.

Rephrase / paraphrase

Generate alternative phrasings of a text. mapping selects the transformation (default "paraphrase"); num_results controls how many alternatives are returned (default 5).

Summarize text

Produce a concise summary of a longer text or document.

Tone classification

Classify the emotional tone of a text (e.g. friendly, confident, sad, angry). Returns overall and per-sentence tone labels with confidence scores.

Sentiment analysis

Classify the sentiment of a text (positive / neutral / negative). Returns overall and per-sentence sentiment with confidence scores.

Language detection

Detect the language a text is written in. Returns an ISO 639-1 code and a confidence value.

PII detection and redaction

Detect personally identifiable information (emails, phone numbers, US SSNs, credit card numbers, IP addresses, IBANs, US bank routing numbers) in a text. Detection is deterministic (regex + checksum validation, no ML), and every entity comes with exact character offsets (start/end) into the submitted text. Set redact=true to also receive redacted_text: the input with each entity replaced by a placeholder like [EMAIL]. Person names and street addresses are not detected.

Readability statistics

Compute readability statistics for a text: grade-level readability scores, estimated reading time, and counts of characters, words, and sentences.

Writing quality score

Score the writing quality of a text from 1 (poor) to 5 (excellent). Returns score: a fluency/grammaticality score from a language model. Set sentence_scores=true to also get a per-sentence breakdown (sentences with start/end offsets and their own score) to find the weakest sentences. Set rubric=true for an LLM-judged rubric: overall score, per-dimension scores (clarity, coherence, correctness, concision), a summary, and concrete issues each with the quoted excerpt, offsets, a note, and a suggest

SEO analysis and metadata

Analyze a page or article for on-page SEO and generate metadata. Returns stats (words, sentences, paragraphs, reading time, Flesch reading ease and grade), per-target-keyword count/density/whether it appears in the first 100 words, top_terms (the most frequent content words and phrases), and suggestions: title-tag candidates (the model aims for <=60 chars; any over a 90-char hard cap are dropped), meta-description candidates (aims for 120-155 chars; hard cap 220), a URL slug (null if none could

Zero-shot text classification

Classify a text into a set of labels you supply — intent, topic, category, routing queue, sentiment buckets, anything — without training data. Returns label (the best match, always one of your labels), labels (single-label: just the best match; multi-label: every label whose score >= threshold, possibly empty), scores (one {label, score} per input label, descending; a distribution in single-label mode, independent 0-1 applicability in multi-label mode) and a one-sentence rationale. Labels can be

Structured data extraction

Pull structured fields out of unstructured text — invoice numbers, amounts, dates, names, addresses, order ids, anything the document actually states — by describing the fields you want. Returns data (a {field: value} map, null where the text does not state the field), fields (per-field value, the verbatim evidence span it came from, and whether it was found) and missing (the fields the text did not yield). Values are returned in the type you declare (string, number, integer, boolean, date as YY