This retriever uses a combination of semantic similarity and a time decay.
The algorithm for scoring them is:
semantic_similarity + (1.0 - decay_rate) ^ hours_passed
Notably, hours_passed refers to the hours passed since the object in the retriever was last accessed, not since it was created. This means that frequently accessed objects remain "fresh".
from datetime import datetime, timedelta
import faiss
from langchain.retrievers import TimeWeightedVectorStoreRetriever
from langchain_community.docstore import InMemoryDocstore
from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings
A low decay rate (in this, to be extreme, we will set it close to 0) means memories will be "remembered" for longer. A decay rate of 0 means memories never be forgotten, making this retriever equivalent to the vector lookup.
embeddings_model = OpenAIEmbeddings()
embedding_size = 1536
index = faiss.IndexFlatL2(embedding_size)
vectorstore = FAISS(embeddings_model, index, InMemoryDocstore({}), {})
retriever = TimeWeightedVectorStoreRetriever(
vectorstore=vectorstore, decay_rate=0.0000000000000000000000001, k=1
)
yesterday = datetime.now() - timedelta(days=1)
retriever.add_documents(
[Document(page_content="hello world", metadata={"last_accessed_at": yesterday})]
)
retriever.add_documents([Document(page_content="hello foo")])
['73679bc9-d425-49c2-9d74-de6356c73489']
retriever.invoke("hello world")
[Document(metadata={'last_accessed_at': datetime.datetime(2024, 10, 22, 16, 37, 40, 818583), 'created_at': datetime.datetime(2024, 10, 22, 16, 37, 37, 975074), 'buffer_idx': 0}, page_content='hello world')]
With a high decay rate (e.g., several 9's), the recency score quickly goes to 0! If you set this all the way to 1, recency is 0 for all objects, once again making this equivalent to a vector lookup.
embeddings_model = OpenAIEmbeddings()
embedding_size = 1536
index = faiss.IndexFlatL2(embedding_size)
vectorstore = FAISS(embeddings_model, index, InMemoryDocstore({}), {})
retriever = TimeWeightedVectorStoreRetriever(
vectorstore=vectorstore, decay_rate=0.999, k=1
)
yesterday = datetime.now() - timedelta(days=1)
retriever.add_documents(
[Document(page_content="hello world", metadata={"last_accessed_at": yesterday})]
)
retriever.add_documents([Document(page_content="hello foo")])
['379631f0-42c2-4773-8cc2-d36201e1e610']
retriever.invoke("hello world")
[Document(metadata={'last_accessed_at': datetime.datetime(2024, 10, 22, 16, 37, 46, 553633), 'created_at': datetime.datetime(2024, 10, 22, 16, 37, 43, 927429), 'buffer_idx': 1}, page_content='hello foo')]
Using some utils in LangChain, you can mock out the time component.
from langchain_core.utils import mock_now
tomorrow = datetime.now() + timedelta(days=1)
with mock_now(tomorrow):
print(retriever.invoke("hello world"))
[Document(metadata={'last_accessed_at': MockDateTime(2024, 10, 23, 16, 38, 19, 66711), 'created_at': datetime.datetime(2024, 10, 22, 16, 37, 43, 599877), 'buffer_idx': 0}, page_content='hello world')]
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4