Changelog

Superlinked through time.

[9.22.1] - 2024-08-28

Event support in batch: Clients can now batch calculate with events as well, rendering the online and batch systems in feature parity.
Support for more hard filter operations: Added support for LT, LTE, GT, GTE, AND, OR, CONTAINS, NOT_CONTAINS, IN, NOT_IN, unlocking more tabular data heavy use-cases.
Optimized GPU usage: Optimized GPU utilization for best performance, which below a certain size (~10k embeddings) is more optimal with CPU.
Integration tests of online executor with GPU: Added tests to measure if the frameworks correctly recognize the underlying GPU, if any.
Added scores to example notebooks: Expanded all notebooks to show how scores work, allowing users to better understand the underlying vector similarity scores.
Further NLQ support: Extended NLQ features with filter support and temperature tuning based on feedback.

Results were not ordered by score with Redis: Fixed issue where results were not properly ordered by their scores in ascending order.
Long categorical embeddings were dominating the results: Fixed an issue that resulted in all 0 vectors that skewed the results by breaking the normalization.
Source.put was behaving differently with different inputs: Fixed to work as expected with arrays and data frames alike.
Index temperature was not accepting integers: Fixed the index to accept integer temperatures.

Added better formatted logs in tests: Improved logging support for easier issue identification during development.
(breaking) Removed status endpoints for initial data loader: Necessary step to allow the executor to be stateless, preparing it for high availability hosting.

Benchmarked the batch performance: Ran batch in GCP for the first time and benchmarked the performance (700k items under 4 mins).

Return similarity scores: Now returning similarity scores along with results, allowing clients to better understand the distribution.
Improved feature notebooks: Extended with examples containing similarity scores, querying of recency space, and event parameters (max_count, max_age, temperature).
Code quality checks for server: Added automations to ensure server testing and unified code format.

Recency was not always using the same NOW from the context: Fixed to use the correct reference point.

Conducted retrospective meetings for Natural Language Queries and Distributing Superlinked Batch.

Bumped sentence-transformers to 3.0.1: Allows experimentation with highest scoring models from mteb leaderboard.
Support for empty list embedding: Enables ingestion of data with lists where not all rows have an item.
Default limits to vector database connectors: Set default return of 10 items for both Redis and Mongo.
Natural Language Queries: Users can describe parameters for prompting the underlying model.
Support logarithmic number embeddings: Captures non-linear preferences with large values.

Negative weights boosting recommendations: Fixed issues within the event handling system.
Negative weight now applied even for the first received event: Addressed bug where negative weights only worked after a positively weighted event.

Last updated 1 day ago