Add ChatMessage placeholder on ChatPromptBuilder #7958
Description
Is your feature request related to a problem? Please describe.
I was working on issue #7868 using alternative ways for storing messages, one based on ChatMessageStack that is simply a stack that keeps the order of messages and returns the last N messages for short-term memory, and another based on the Indexable alternative proposed in #7830 for long-term memory. Once I worked out everything I needed memory-wise, I tried to group everything in the following manner:
- System message introducing the agent objective and introducing the long-term memory messages.
<The messages retrieved from the long-term memory>
- System message introducing the last N messages.
<The messages retrieved from the short-term memory>
- System message containing the context for RAG and introducing the last query.
- Chat message containing the user query.
But I needed to create another component just to stack chat messages, and hence take advantage of how chat models are trained. I find this should be implemented directly on the ChatPromptBuilder component, as it is good practice for prompting chat models.
Describe the solution you'd like
I want to implement some kind of placeholder on the ChatPromptBuilder template to be able to insert chat messages directly in there, as shown here:
template = [
ChatMessage.from_system("You are a helpful assistant. This is your chat history:"),
PlaceHolder("messages"),
ChatMessage.from_system("""Respond to the following query based on the following context.
Context:
{% for doc in documents %}
{{ doc.content }}
{% endfor %}
"""),
ChatMessage.from_user("{{ query }}")
]
prompt_builder = ChatPromptBuilder(template=template)
messages = [ # List of messages
ChatMessage.from_user("First query"),
ChatMessage.from_assistant("Your response")
]
documents = [ # List of retrieved documents
Document("Document Content")
]
result = prompt_builder(messages=messages, documents=documents, query="This is the last query")And should output the following inside the prompt key of result:
[
ChatMessage.from_system("You are a helpful assistant. This is your chat history:"),
ChatMessage.from_user("First query"),
ChatMessage.from_assistant("Your response"),
ChatMessage.from_system("""Respond to the following query based on the following context.
Context:
Document Content
"""),
ChatMessage.from_user("This is the last query")
]Describe alternatives you've considered
I considered writing my own component, but I believe this fits better inside the ChatPromptBuilder than in a new component.