Qdrant gives different and incorrect results with DatetimeRange filter

To my knowledge, semantic similarity score is absolute, not relative.

However, using the DatetimeRange filter, I get different results when the date range is expanded. The correct document disappears even though it should be within the range.

eg, given the question and searching for range > 2022:

from qdrant_client.http import models as qdrant_models
from langchain_community.vectorstores import Qdrant

# skip loading of vector store

vector_store.similarity_search(question,k=15, 
filter=qdrant_models.Filter(
must=qdrant_models.FieldCondition(key='meeting_date',
range=qdrant_models.DatetimeRange(gte='2022-01-01')))
)

This gives (doc id | doc title | score)

[('CR-ID-REPLACED', <title replaced>, 0.63587),
('CR-ID-REPLACED', <title replaced>, 0.62937194),
('CR-ID-REPLACED', <title replaced>, 0.629144),
('CR-ID-REPLACED', <title replaced>, 0.6280967),
('CR-ID-REPLACED',<title replaced>, 0.6271694)]

Which is incorrect. The relevant doc is not retrieved. However, if i include a max date, the result is correct.

vector_store.similarity_search(question,k=15,
filter=qdrant_models.Filter(
must=qdrant_models.FieldCondition(key='meeting_date',
range=qdrant_models.DatetimeRange(gte='2022-01-01', lte='2023-01-01'))))

[('CR-ID-REPLACED', <title replaced> 0.66140205), # <-- correct document!
('CR-ID-REPLACED', <title replaced>, 0.6271694),
('CR-ID-REPLACED', <title replaced>, 0.6211362),
('CR-ID-REPLACED', <title replaced>, 0.62014866),
('CR-ID-REPLACED', <title replaced>, 0.6195655)]

I also get the correct docs above if i dont set any date range filters. eg,

vector_store.similarity_search(question,k=15)

What I don't understand is why does this behavior happen?

• The correct doc has the highest score, but it completely disappear when we dont include a max date even though it should logically be in the same set?
• Furthermore, None of the documents retrieved in the 2nd query are the same! I expect at least some overlap?

This is boggling the mind. Would appreciate any help!

versions

Qdrant version v1.12.4

Using the langchain qdrant abstraction

langchain-community==0.2.17
qdrant-client==1.12.1

Start by reproducing the two DatetimeRange-filtered similarity_search calls and the unfiltered call with Qdrant v1.12.4, qdrant-client 1.12.1, and langchain-community 0.2.17. Compare the returned documents and scores, then trace the relevant Qdrant filtering and search behavior; done means the discrepancy is explained and a regression test demonstrates consistent results.