semantic_similarity indicates the semantic similarity between documents or data.
decay_rate represents the ratio at which the score decreases over time.
hours_passed is the number of hours elapsed since the object was last accessed.
The key feature of this approach is that it evaluates the “ freshness of information ” based on the last time the object was accessed.
In other words, objects that are accessed frequently maintain a high score over time, increasing the likelihood that frequently used or important information will appear near the top of search results. This allows the retriever to provide dynamic results that account for both recency and relevance.
Importantly, in this context, decay_rate is determined by the time since the object was last accessed , not since it was created.
Hence, any objects that are accessed frequently remain "fresh."
You can alternatively set API keys such as OPENAI_API_KEY in a .env file and load them.
[Note] This is not necessary if you've already set the required API keys in previous steps.
# Load API keys from .env file
from dotenv import load_dotenv
load_dotenv(override=True)
Low decay_rate
A low decay_rate (In this example, we'll set it to an extreme value close to 0) means that memories are retained for a longer period .
A decay_rate of 0 means that memories are never forgotten , which makes this retriever equivalent to a vector lookup.
Initializing the TimeWeightedVectorStoreRetriever with a very small decay_rate and k=1 (where k is the number of vectors to retrieve).
from datetime import datetime, timedelta
import faiss
from langchain.docstore import InMemoryDocstore
from langchain.retrievers import TimeWeightedVectorStoreRetriever
from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings
# Define the embedding model.
embeddings_model = OpenAIEmbeddings(model="text-embedding-3-small")
# Initialize vector store empty.
embedding_size = 1536
index = faiss.IndexFlatL2(embedding_size)
vectorstore = FAISS(embeddings_model, index, InMemoryDocstore({}), {})
# Initialize the time-weighted vector store retriever. (in here, we'll apply with a very small decay_rate)
retriever = TimeWeightedVectorStoreRetriever(
vectorstore=vectorstore, decay_rate=0.0000000000000000000000001, k=1
)
Let's add a simple example data.
# Calculate the date of yesterday
yesterday = datetime.now() - timedelta(days=1)
retriever.add_documents(
# Add a document with yesterday's date in the metadata
[
Document(
page_content="Please subscribe to LangChain Youtube.",
metadata={"last_accessed_at": yesterday},
)
]
)
# Add another document. No metadata is specified here.
retriever.add_documents(
[Document(page_content="Will you subscribe to LangChain Youtube? Please!")]
)
['58449575-d54f-47dc-9a76-806eccb850f3']
# Invoke the retriever to search
retriever.invoke("LangChain Youtube")
In this case, when you invoke the retriever, "Will you subscribe to LangChain Youtube? Please!" is returned first.
Because decay_rate is high (close to 1), older documents (like the one from yesterday) are nearly forgotten.
decay_rate overview
when decay_rate is set to a very small value, such as 0.000001:
The decay rate (i.e., the rate at which information is forgotten) is extremely low, so information is hardly forgotten.
As a result, there is almost no difference in time-based weights between recent and older information . In this case, similarity scores are given higher priority.
When decay_rate is set close to 1, such as 0.999:
The decay rate is very high, so most past information is almost completely forgotten.
As a result, in such cases, higher scores are given to more recent information.
Adjusting the decay_rate with Mocked Time
LangChain provides some utilities that allow you to test time-based components by mocking the current time.
The mock_now function is a utility function provided by LangChain, used to mock the current time.
[NOTE]
Inside the with statement, all datetime.now() calls return the mocked time . Once you exit the with block, it reverts back to the original time .
import datetime
from langchain_core.utils import mock_now
# Define a function that print current time
def print_current_time():
now = datetime.datetime.now()
print(f"now is: {now}\n")
# Print the current time
print("before mocking")
print_current_time()
# Set the current time to a specific point in time
with mock_now(datetime.datetime(2025, 1, 7, 00, 00)):
print("with mocking")
print_current_time()
# Print the new current time(without mock_now block)
print("without mock_now block")
print_current_time()
before mocking
now is: 2025-01-07 14:06:37.961348
with mocking
now is: 2025-01-07 00:00:00
without mock_now block
now is: 2025-01-07 14:06:37.961571
By using the mock_now function, you can shift the current time and see how the search results change.
This helps you find an appropriate decay_rate for your use case.
[Note]
If you set the time too far in the past, an error might occur during decay_rate calculations.
# Example usage changing the current time for testing.
with mock_now(datetime.datetime(2025, 1, 7, 00, 00)):
# Execute a search in this simulated timeline.
print(retriever.invoke("Langchain Youtube"))
