How to use TimeWeightedVectorStoreRetriever as VectorStoreRetrieverMemory's retriever?

[Percival-Chan]

Hi, I want to use a TimeWeightedVectorStoreRetriever for my chatbot's memory, but I faced AttributeError: 'TimeWeightedVectorStoreRetriever' object has no attribute 'get'. What would be the correct approach? My code is as follow:

from langchain.chains.llm_math.base import LLMMathChain
from langchain_community.utilities.serpapi import SerpAPIWrapper
from langchain.chains.conversation.base import ConversationChain
from langchain.memory import VectorStoreRetrieverMemory

decay_rate = 0.01
number_of_documents = 10

from langchain.retrievers import TimeWeightedVectorStoreRetriever
retriever = TimeWeightedVectorStoreRetriever(
vectorstore=vectorstore, decay_rate=decay_rate, k=number_of_documents
)

vector_memory = VectorStoreRetrieverMemory(retriever=retriever)
data = retrieve_data(user_name)
database_history = load_conversation_history_from_database(data)
memory_with_history = save_conversation_to_memory(database_history, vector_memory)
vectorstore_chain = ConversationChain(
llm=azure_llm_low_temp_model,
prompt=VECTORSTORE_PROMPT,
memory=memory_with_history,
verbose=True
)

vectorstore_chain.invoke

[dosubot[bot]]

Hey there, @Percival-Chan! I'm here to help you out with any bugs, questions, or contributions you have. Let's tackle this together!

The TimeWeightedVectorStoreRetriever does not have a get method, which is likely causing the AttributeError when you try to use it with VectorStoreRetrieverMemory. The VectorStoreRetrieverMemory expects the retriever to have a get method to retrieve documents, but TimeWeightedVectorStoreRetriever is designed differently, focusing on time-weighted retrieval.

To resolve this, you might need to implement a custom method or wrapper that provides the expected interface for VectorStoreRetrieverMemory. Alternatively, you could modify your approach to use a retriever that is compatible with VectorStoreRetrieverMemory or adjust the memory implementation to work with the TimeWeightedVectorStoreRetriever.

If you want to stick with TimeWeightedVectorStoreRetriever, consider creating a custom memory class that can handle the retrieval logic specific to this retriever, ensuring it aligns with the expected methods and attributes.

To continue talking to Dosu, mention @dosu.

[Percival-Chan]

@dosu
do you have an example showing me how to create a custom memory class for this case or how to add decay time to FAISS?