Dear Michael, it's funny that no one dares to answer our questions to the point, or better, "humanly". My feeling is that Trados *might* send the preceding and the following sentence in a best-case scenario. I don't expect more than that. 

 Michael Schroeder Marco Simionato 

Marco Simionato said:

it's funny that no one dares to answer our questions to the point, or better, "humanly"

I'll give it a go... although someone from the AppStore team may do a better job of it!

The short answer is that the “contextually-aware” claim doesn’t mean the entire document is streamed to the LLM in one go, and it isn’t simply a hallucination... it's just more nuanced:

1. Per-segment processing by design
   Under the hood, Trados Studio (and therefore the OpenAI-Provider plugin) breaks your file into segments and sends each segment individually to the LLM for translation.  Even in a batch Pre-Translate task, segments are batched (typically 10-20 at a time) but still treated separately.  No full-document payload is sent in one API call .

2. What “contextually aware” actually means

   • The plugin builds its System Instructions dynamically based on project-level settings (source/target language, XML-tag rules, terminology preferences, etc.), and these instructions are submitted with each segment.

   • You can enable “Terminology-aware” translation, which injects your term lists into the prompt so the LLM honours them consistently.

   • In future iterations, the engineering team is exploring ways to include the previous and/or next segments in the prompt to give the LLM a little more context without sending the whole document .

3. Why results differ from the ChatGPT web interface

   • ChatGPT Web: Maintains a running history of your entire conversation, so it naturally “remembers” what came before.

   • Trados Plugin: Each segment request is stateless beyond the System Instructions and User Prompt you configure.  Unless you explicitly include extra context (e.g. domain subject-matter or prior segment text in your custom prompt), the LLM sees only that one sentence plus whatever prompt-level context you’ve provided.

4. Token-cost & performance trade-offs
   Sending full documents or multiple segments per request would dramatically increase token usage (and thus cost and latency).  That’s why the plugin adheres to segment-level calls by default, even though it limits context.  Overcoming this requires re-engineering the batch-task workflow and careful prompt engineering to balance quality, cost and speed.

In practice: if you need tighter consistency across neighbouring segments you can:

• Manually include the previous segment (or a short summary of it) in your User Prompt (although probably not too helpful for batch processing!).

• Define clear domain-specific system instructions (e.g. “This is a legal contract about employment law”) so the LLM applies a consistent style.

• Wait for future updates that may support “previous & next segments”.

But rest assured that the plugin doesn’t send your entire text to OpenAI in one shot, and “contextually aware” really refers to the prompt-engineering around each segment rather than a single, document-level API call.

I hope that helps?

Hi Paul 

Thank you very much for your extensive and "human" answer.

I suspected the conclusions as much, but it's good to have verification to understand exactly what we can expect from "contextually aware", and what we cannot. I found especially helpful the tip about providing domain-specific instructions directly in the prompt for each project, something that I hadn't considered before. All in all, I think the developers have done a good job balancing the different considerations into a useful tool. As the price/usage models evolve around AI, I am confident that RWS will keep pace. Thanks again!

Hi Paul 

Thank you very much for your extensive and "human" answer.

I suspected the conclusions as much, but it's good to have verification to understand exactly what we can expect from "contextually aware", and what we cannot. I found especially helpful the tip about providing domain-specific instructions directly in the prompt for each project, something that I hadn't considered before. All in all, I think the developers have done a good job balancing the different considerations into a useful tool. As the price/usage models evolve around AI, I am confident that RWS will keep pace. Thanks again!