LLM keeps using code block formatting for non-code

[dvejsada]

I have an issue with Claude 3.5 Sonnet v1 used over Amazon Bedrock that keeps formatting parts of the answers (e.g. e-mail draft) in a code block formatting, even if I expressly instruct it in agent system message not to do it. Is there maybe any conflict with general agent system message from LibreChat?

The problem is that the text in code block is not wrapped, so it is very complicated to read the provided answer in code block formatting.

Any suggestions how to fix it?

[danny-avila]

If the content is indented in a certain way, markdown interprets this as a code block.

Is there maybe any conflict with general agent system message from LibreChat?

There is no general system message, but tools/capabilities can influence the LLM's context, if you have any enabled (artifacts, code interpreter, etc.).

In your prompt, you should be ultra-specific even with examples.