krzysztof/landing_page
1
0
Fork 0
mirror of https://github.com/qdrant/landing_page.git synced 2026-10-02 09:28:30 +02:00
Code Issues Packages Projects Releases Wiki Activity

Update case-study-pento.md

Browse Source
This commit is contained in:
daniel-azoulai
2025-07-14 13:14:52 -07:00
parent 042d69e882
commit 4e367811d7
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
qdrant-landing/content/blog/case-study-pento.md
+2 -2
View File
@@ -85,7 +85,7 @@ Let us clarify that clearly we don't have to use the entire user history, we can
Not all tastes carry the same weight, especially over time. A cluster of artworks a user connected with months ago may no longer reflect their current preferences. To account for this, we assign a recency-aware score to each cluster, emphasizing freshness without discarding history.
Each cluster is scored based on the timestamps of the interactions it contains, using an exponential decay function:
![exponential decay function](/blog/case-study-pento/exponential-decay-function.png)
![exponential decay function](/blog/case-study-pento/pento-exponential-decay-function.png)
Where:
* wi is the normalized rating calculated before
@@ -106,7 +106,7 @@ To maintain the agility of the system and to avoid an excess of likes per user,
#### User Representation as Multivectors
Once clusters are identified and scored, we distill a user’s taste into something compact, expressive, and ready for retrieval: a multivector representation.
Once clusters are identified and scored, we distill a user’s taste into something compact, expressive, and ready for retrieval: a multivector representation [learn more about multivectors representations in Qdrant](https://qdrant.tech/documentation/advanced-tutorials/using-multivector-representations/).
Each positively scored cluster contributes its medoid vector, a single embedding that represents the core of that aesthetic preference. We do the same for negatively scored clusters, treating them as regions the user tends to avoid.