Profiles

A profile is a named, serializable preset that assembles a sequence of steps into a Pipeline. Picking a profile is how you choose the normalization trade-off: how much to repair versus how much to fold away.

araclean is non-destructive by default (ADR-0004). Each step declares a safety class, and a profile is only as lossy as its steps:

• encoding_repair — lossless. Repairs the encoding (Unicode form, presentation forms, tatweel, bidi/zero-width controls, look-alike letters, whitespace) without discarding any linguistic signal. The default LIGHT profile is only encoding repair.
• linguistic_folding — lossy. Discards a linguistic distinction within the Arabic text (dediacritization, alef/hamza/teh-marbuta/alef-maqsura folding, digit/punctuation mapping).
• cleaning — lossy. Removes non-linguistic noise around the text (URLs, mentions, hashtags, HTML, foreign-script spans).

Each profile page below enumerates exactly the steps it applies, in order, and labels each one lossless or lossy — so you can choose with full knowledge of the trade-off.

Profile	Lossless?	For
LIGHT	✓ lossless	the safe default — encoding repair for any corpus
CLASSICAL	✓ lossless	vocalized / Qur'anic text: repair while preserving every mark
ML	✗ lossy	model input: dediacritize + cap elongation, but keep letter distinctions
SEARCH	✗ lossy	search / IR recall: fold spelling & vocalization variants together
SOCIAL	✗ lossy	noisy user text: clean URLs/mentions/HTML, keep emoji

The lossless/lossy labels and the per-profile step lists are generated from the assembled pipeline itself, so they always match what the code actually does.

A profile is a starting point, not a take-it-or-leave-it bundle: per-knob overrides adjust one behavior with loud validation (see Tuning profiles), and drop/ select/explicit construction give you full control of the step sequence (see Composing pipelines).