Describe the bug
In vertex_ai_rag_memory_service, the entire session conversation (the text parts) are being combined
and uploaded as a file.
|
async def add_session_to_memory(self, session: Session): |
Q1:
I would have thought that only the "user" prompt/input would be recoded as the memory and not the output/response of the AI.
Q2:
Assuming one vector embedding is being generated for the entire file, the quality of the feature vector would be low. I don't know the underlying behavior/implementation of VertexAiRag so it is possible that some chunking is happening. Even if it is the case you are storing event as json and chunking etc may not be optimal.
Q3:
Wouldn't it make sense to summarize the conversation or extract key elements and then store them as memory?
Regards & thanks
Kapil
Describe the bug
In vertex_ai_rag_memory_service, the entire session conversation (the text parts) are being combined
and uploaded as a file.
adk-python/src/google/adk/memory/vertex_ai_rag_memory_service.py
Line 67 in 62a543b
Q1:
I would have thought that only the "user" prompt/input would be recoded as the memory and not the output/response of the AI.
Q2:
Assuming one vector embedding is being generated for the entire file, the quality of the feature vector would be low. I don't know the underlying behavior/implementation of VertexAiRag so it is possible that some chunking is happening. Even if it is the case you are storing event as json and chunking etc may not be optimal.
Q3:
Wouldn't it make sense to summarize the conversation or extract key elements and then store them as memory?
Regards & thanks
Kapil