Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

 Note: The text of the prompt is periodically revised, so it is not indicated for reference purposes so as not to misinform.

...

Why is the glossary so small?

OpenAI sets limits on the size of an incoming request. This size differs depending on the GPT version. For this reason, prompt cannot be infinite. We distribute the permitted volume into the source text, style guide, glossary and prompt text. Therefore, each of these partitions has size limitations.

How is a large file translated?

Typically, the user has two ways to complete the translation:

а) Segment by segment. As a rule, there are no issues here, except when the source contains long sentences, for example HTML with tags. In this case, it is recommended to use the Tags Encoding filter. It will collapse all tags into a short form and protect the tags from translation.

b) Translation of the entire file. Depending on the connector, the files are divided into batches of 10 - 30 rows. Batches are sent for http://translation.To work correctly with GPT, batches are divided into chunks of 10 rows for GPT3.5 and 5 rows for GPT4. After translation, the chunks are collected into batches and sent back.

How are costs calculated?

Each request is charged according to GPT rules. GPT counts not only the response, but also the request. The request is formed from: Source text, prompt, role, style guide and glossary. And the response includes what is returned in the form of a GPT translation (sometimes with comments). Thus, the volume of tokens for the request and for the response will be different.

The prompt will be applied every time a translation request is sent.

To reduce costs, minimise the size of details in Prompt engineering (style guide and glossary) and translate the document as a whole, not line by line. When translating an entire document, the prompt will be applied once per chunk (once per 10 lines for GPT3.5 and once per 5 lines for GPT4).

What happens if there is an error?

If GPT throws an error (unavailable, returns an empty response or an incomplete response), the console will make up to 5 attempts to receive the translation, after, if the issues still persist, it will return empty lines and continue working with another chunk.