Avoiding ambiguity between formatted ChatPromptTemplate messages?

[chrispy-snps]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

from langchain_core.prompts import ChatPromptTemplate

#
# define system instructions
#
system_instructions = """
You are a helpful assistant. Use these documents to answer the human question below.

DOCUMENTS:
{md_docs}
END OF DOCUMENTS
""".strip()

#
# define Markdown document content
#
md_docs = """
# Document 1 Title

This is some text.

# Document 2 Title

## Topic Title

Here is an introduction.

All of this text forms a
single paragraph in Markdown
because adjacent text content
lines are combined into a
single paragraph.
""".strip()

#
# construct a PromptTemplate
#
prompt_template = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            system_instructions,
        ),
        (
            "human",
            "What is the capital of France?",
        ),
        (
            "assistant",
            "The capital of France is Paris.\n\nIt is a beautiful city.",
        ),
        (
            "human",
            "Can you tell me more about it?",
        ),
    ]
)

#
# render it to text
#
print(prompt_template.format(md_docs = md_docs))

Description

When a ChatPromptTemplate is rendered to text, all messages are string-concatenated directly adjacent to each other. But when newlines are significant within messages (such as to delineate Markdown block elements or assistant/user paragraphs), this causes ambiguity in the message boundaries because the role delimiters now occur in the middle of what the LLM might interpret as "paragraphs".

Here is the rendered ChatPromptTemplate output from the provided code, with markers added to show how the LLM might interpret paragraphs after processing the Markdown content:

System: You are a helpful assistant. Use these documents to answer the human question below.

DOCUMENTS:
# Document 1 Title

This is some text.

# Document 2 Title

## Topic Title

Here is an introduction.

All of this text forms a                 <-- 1
single paragraph in Markdown             <-- 1
because adjacent text content            <-- 1
lines are combined into a                <-- 1
single paragraph.                        <-- 1
END OF DOCUMENTS                         <-- 1
Human: What is the capital of France?    <-- 1
AI: The capital of France is Paris.      <-- 1

It is a beautiful city.                  <-- 2
Human: Can you tell me more about it?    <-- 2

We are working around this by post-processing all our messages to manually add trailing newlines, but is this something that LangChain should be handling natively?

System Info

langchain 0.3.9, using Python 3.10 on Linux