Arabic Pipeline API

Transcribe Arabic or English speech to plain text. Send the audio as base64 (webm, mp3, wav or mp4); the payload limit is 12 MB of base64. Returns the transcript plus the model that produced it.

Extract structured fields from an Arabic document image or PDF (receipts, invoices, official letters, national IDs, bank slips) as JSON: amounts, dates, names, identifiers and tax numbers, with null for fields the document does not contain. Send the file as base64 (jpeg, png or pdf); the limit is 12 MB of base64.

Extract structured fields from 1 to 5 Arabic documents in one paid call: $0.10 for the batch, i.e. $0.02 per document instead of $0.03. Documents are processed independently and in parallel, results come back in request order with an echoed id, and one unreadable image does not fail the rest (it appears with ok=false and an error). Use this for a stack of receipts, invoices or bank slips.

Split long text into chunks for a model context window. Cuts at sentence boundaries when possible, then at word boundaries, and never mid-word unless a single word is longer than max_chars. Returns each chunk with exact start/end offsets into the original string plus an optional overlap for context carry-over.

Identify whether text is Arabic, English, mixed or unknown from its script mix, with a confidence score and the count of Arabic, Latin and other letters. Deterministic and model-free, so it is cheap enough to run on every incoming message to route Arabic work.

Compare two texts word by word after Arabic normalisation (tashkeel, alef/ya/ta-marbuta and digit variants are treated as equal). Returns an equal flag, a similarity score and the added and removed words. Use it for deduplication, re-import checks and QA of edited Arabic copy. Each side is limited to 1500 words.

Measure text deterministically: characters, words, unique words, sentences, lines, paragraphs, digits, punctuation, average word and sentence length, reading time and an Arabic/Latin language mix. Works on any language, not only Arabic.

Canonicalise Arabic text for search, matching or indexing: remove tashkeel and tatweel, fold alef/ya/ta-marbuta, unify Arabic-Indic digits, strip bidi control characters and collapse whitespace. mode=light is lossless; search adds letter folding; index also strips punctuation so the result can be used as a dedup key.

Search Arabic (or any) text with normalisation: tashkeel, alef/ya/ta-marbuta, hamza and Arabic-Indic digit variants are ignored, so فَاطِمَة matches فاطمة and ١٢٣ matches 123. mode=words (default) matches whole words, phrase matches any substring, fuzzy tolerates up to max_distance edits per word. Returns exact start/end offsets into the original text plus a context snippet.