Akana. Make the structure of Turkish part of your application.
Akana processes Turkish text: it tokenizes, analyses morphology, corrects grammar and spelling, masks personal data and secrets, detects and humanizes AI-written text, and searches by meaning. Use it from Python, the command line or Rust.
- kitabıma: The root returns to its dictionary form: kitab becomes kitap.
- okullarımızdakilerden: Seven parts in one word: one root and six suffixes.
- çalıştırılamayacakmış: A verb with five suffixes. Each one changes what it means.
Akana splits three Turkish words into a root and suffixes. kitabıma: kitap (root, noun) + ım (possessive, my) + a (dative, to), “to my book”; the b turns back into the dictionary root kitap. okullarımızdakilerden: okul + lar + ımız + da + ki + ler + den, “from the ones in our schools”. çalıştırılamayacakmış: çalış + tır + ıl + ama + yacak + mış, “reportedly, it will not be possible to make it work”. Segmentation from Akana 0.6.0.
Eight tools for Turkish text.
Akana is a Turkish natural language processing toolkit built in Rust, with Python bindings and hardware SIMD acceleration. Each tool works on its own or in one pipeline with the others.
Tokenizer
Rule-based, zero-allocation tokenization and sentence segmentation that handles apostrophes (İstanbul’da), abbreviations, URLs, dates and currencies. About 950,000 tokens per second.
Morphological analyzer
Roots and suffixes from a 93,167-root lexicon, compound words, and words generated back from an analysis. About 56 times faster than Zeyrek.
Grammar checking and correction
Rule-based detection and correction across all 25 GECTurk categories (de/da, ki, mi, apostrophes, compounds), with explanations. 1,471 sentences per second on one CPU core.
Spell checker
SIMD-accelerated spell checking, Turkish diacritic restoration and normalization of informal writing (yapcam → yapacağım, çooook → çok).
PII and secret masking
More than 50 KVKK-aligned categories, from ID numbers, IBANs and phone numbers to API keys and passwords, with checksum validation and lossless restoration. About 3,570 documents per second.
AI-generated text detection
Finds the signatures of LLM-written Turkish: punctuation anomalies, repeated -mektedir predicates, monotonous rhythm, bureaucratic connectors and translationese.
Turkish humanizer
Turns those findings into concrete rewriting guidance in five registers, from legal and administrative to literary writing.
Semantic and hybrid search
A built-in 2.5 MB Turkish embedding model (20,000 sentences per second) and ALTAI’s own hybrid retrieval: 1-bit dense vectors fused with morphological BM25, reaching 96.7–100% Recall@10 on Turkish RAG benchmarks on a CPU alone.
From input to output.
- Input: A Turkish paragraph containing inflected words, abbreviations, and multiple sentences.
- Process: Tokenize the text with Turkish-aware rules, then apply morphology, grammar and spelling correction, PII masking or search as needed.
- Output: Text segments and linguistic analyses for search, document review, or natural language processing workflows.
In client projects.
ALTAI uses Akana to build the text-processing layer for Turkish document preparation and search workflows. We evaluate the relevant steps on your organization's text and integrate the selected operations into the application.
Run the first example.
Executed with Akana 0.6.0.
pip install akana==0.6.0
python - <<'PY'
import akana
print(akana.tokenize_words("Kitabıma yeni notlar ekledim."))
print(akana.Morphology().analyze("kitabıma"))
PY
Spec sheet
- Runtime: Rust + Python
- Interface: Python / CLI / Rust
- License: MIT / Apache-2.0
- Source: GitHub, Documentation
- Scope: Each tool is a separate job. Speed and accuracy figures are from the project’s own benchmarks (README).