Is Thiruttham rule-based or AI?
Both, and every suggestion says which one made it. Thiruttham has two layers. They do different jobs and have different limits.
Layer 1: fixed rules, in your browser
The rules layer is written code, not a language model. It runs in your browser, needs no account, works offline, and gives the same answer every time for the same text. A rule check sends your text nowhere.
- Consonant doubling (ottru) after அந்த, இந்த and எந்த; after ஆக and என; after the object ending ஐ and the dative ending கு, in the listed forms.
- A short list of standard spellings.
- A list of 45 common spoken words and endings and their written forms. This is on by default and you can turn it off.
Layer 2: AI-assisted checks, after you sign in
Some mistakes cannot be decided by a rule, because they depend on what the sentence means. For those, Thiruttham asks a language model, only when you are signed in and press Check.
- Similar-sounding letters: ர and ற, ல and ள and ழ, ந and ண and ன.
- Subject and verb agreement, in gender and in number.
- Spoken forms that have a written form, beyond the short list above.
The AI does not write for you. It proposes small edits, each one an exact piece of your text and its replacement. An edit that cannot be matched to your exact text is thrown away. An edit the model is not confident about is not shown: a missed mistake is better than a wrong correction. The explanation you read comes from a fixed Tamil reference for that rule, not from the model.
What is sent, and what is kept
- Rule checks: nothing is sent.
- AI checks: the sentences you check are sent to our server and to a language model provider whose terms exclude training on your text and keeping it. Each checked sentence is kept encrypted for up to 30 days so that a repeat check is fast; no person reads it.
- Your drafts stay in your browser. Your protected words are kept in your account.
- We count accepted and rejected suggestions and similar events. We never record the text in them.
The full statement is the Privacy page.
What is switched off or not offered today
- AI suggestions for the glide letters that join words (sandhi) are switched off because they were not reliable enough.
- Two fixed rules are switched off because they gave too many wrong suggestions in testing: doubling after எட்டு and பத்து, and the check for words that begin with ண or ன.
- AI suggestions that choose between the object ending ஐ and the dative ending கு are not offered.
- Google Docs is not supported yet.
- The extension is for Chrome on a computer. On a phone, use the web editor.
Where it can be wrong
Both layers can be wrong. A rule can flag a name or an unusual word. The AI can propose changing a word that is correct. That is why nothing changes until you accept it. You can dismiss any suggestion, add a word to your protected list so it is never changed, or report a wrong correction. We do not promise that either layer is always right.
Is Thiruttham rule-based or AI?
Both, and every suggestion says which one made it. The fixed rules run in your browser. The AI-assisted check handles mistakes that depend on meaning, and it needs a sign-in.
Does it send my text anywhere?
A rule check sends nothing. An AI-assisted check sends the sentences you check to our server and to a language model provider whose terms exclude training on your text and storing it. We do not train on your text. We keep each checked sentence encrypted for up to 30 days so that a repeat check is fast; no person reads it. If you choose to report a correction, that one sentence is kept longer, with your consent.
Does the AI write or rewrite my text?
No. It proposes small edits, each one an exact piece of your text and its replacement, and nothing changes until you accept it.
Can it be wrong?
Yes. Both layers can be wrong, and we do not promise that either is always right. You can dismiss any suggestion, protect a word, or report a wrong correction.