Paste or type in your Greek text with or without accents. Use input texts of at least 50 characters for optimal performance. The system automatically strips all diacritics (accents, breathings, iota subscripts, diaereses) and normalizes variant letter forms and punctuation to match what the model learned during training. This means you can paste text directly from a digital edition or database without any manual cleanup.

Use empty square brackets without any spaces `[]` wherever you want the model to generate a restoration:

κύριον [] ἐν δημ[]σίῳ

The model will propose completions for each [] ranked from most to least likely. You will see up to 10 suggestions per gap. If you see less than 10 suggestions, it means that fewer than 10 readings fall above the minimum probability threshold of the model.

Sometimes your text has a gap whose content is unknown and you do not want the model to attempt to fill it. Use a question mark ? for these:

κύριον ? ἐν δημ[]σίῳ

Here the model will fill the second gap but leave the first one alone. Internally, ? is converted to an ellipsis character, which is what the model saw during training for gaps of unknown length. This gives the model accurate context: it knows there is a gap in that position but is not asked to fill it.

If you know (or can estimate) how many characters are missing in a gap, you can give the model a hint. Place the number in square brackets immediately before the gap marker:

κύριον ἔστω [2][] δημοσίῳ

This tells the model "fill this gap, and it is approximately 2 characters long." The model uses this hint to produce completions closer to that length.

You can also provide a range:

κύριον [3-4][] ἐν δημοσίῳ

This says "the gap is between 3 and 4 characters." 

If you are unsure of the length, simply omit the hint and only use [] to mark the lacuna. The model was trained to work both with and without length hints.

[]           Fill this gap of unknown length

[5][]      Fill this gap of approximately 5 characters

[3-7][]  Fill this gap of 3 to 7 characters

?           Gap exists here, but do not fill it 

For each [] gap, the model returns a ranked list of unique candidate completions. Duplicate suggestions are automatically removed. The first candidate is the model's top prediction. You may want to review several candidates, as the correct restoration is not always the first one but is often in the top few.

Example input:

τοῖς ἐν ? [] προσκύνημα παρὰ τῶι [5][] Σαράπιδι

The model will:

1. Leave the ? as an unfilled gap (the model sees it as a gap of unknown extent).

2. Fill the first [] with up to 10 ranked suggestions.

3. Fill the [5][] with up to 10 ranked suggestions, biased toward 5-character completions.

* Apollo has not yet been taught to search and generate on the basis of parallels (Retrieval-Augmented Generation). This is the next step in our project.

* Input texts that begin with a gap (e.g. [] προσκυνημα παρα τωι [5][] σαραπιδι) may cause difficulties for the model.

* Occasionally, the model hallucinates by repeating sequences of letters. This usually happens when the model is unable to come up with restorations above a certain probability threshold.