OCR for Hebrew Text Images

I am trying to use an Open AI model to process Biblical Hebrew Text. I have success with Chat GPT providing the image file and a text file of skills but using the API and a number of models does not completely work. Most of the image is correctly analyzed including vowel points in addition to the base characters and punctuations. Howver thiere is one character the is never returned and I am told this is because it is not used in modern hebrew and the models are not trained to detect it. I have has some success in having the image report found shapes but cannot get the results to use that in creating the appropriate unicode to represent the detected text. It is as if these characters are just noise on the image. So how do we get Open AI to retrrain the models with better Biblical Hebrew character recogniztgion. This appars to be one character out of the possible suite of characters that is not detected and supprised that the entire Unicode set of Hebrew Characters is not in the training set. This seems like an ommision.

I would first inspect the underlying output, even requesting logprobs and seeing the bytes representation where there seems to be a missing glyph.

Reason: The AI can output just about anything that can be represented in UTF-8. It can go beyond the glyphs that are in an application’s fonts and even OS fallbacks, showing here for the case of Japanese:

Then: can the character even be written in Unicode? You would want to have a first-hand look and the Hebrew blocks and any accessory blocks, or even have ChatGPT do web research and see if it can identify and reproduce the writing, and then mirror that in your API application. Can the character be seen in web pages without it being embedded as an image?

Finally: in-context training may be an option. This can be some “example” turns, where you place a few small vision image inputs along with a user task about them, and then assistant message can have the desired output. Either of the turns might further discuss the particular characters, but input images can only be placed by a “user” message role. So: you would need to simulate a real conversation, not just a developer message. This can be the understanding needed when there is not sufficient corpus training in Hebrew text images and language.

Your unclear use of models: “an OpenAI model” is a very wide range of possibilities. The API has different quality settings and different default qualities depending on the model family you are using. For example, gpt-5.6 with “original” can take in very big images. This can either be used to your advantage when you upscale or scan at high resolution, or it can be wasteful, consuming self-attention.

So, there is plenty of self-help opportunity here besides asking for one very narrow “training” be done for you- where the tanakh task you are doing might be the only training and training would be on having the complete input-output on vision to text translated by human already, in a model where AI general-purpose vision is more “what color is the apple”.