Changelog
Qyver through time
06-11-2024
Framework/Server/Batch: 12.2.0/12.2.0/1.13.1 → 12.23.0/12.23.0/1.15.2
New Features
Streamlined Imports: Importing Qyver is now simpler—just use
import qyver as qv.Expanded Image Embedding Options: OpenCLIP models are now supported alongside sentence-transformers for embeddings.
Fixes
File Detection Issues: Fixed issues affecting file discovery.
Sentence-Transformers Dynamic Cache: Ensured smooth operation when working with batch processes.
Zero-Division in Event Handling: Resolved cases where events with no age led to division errors.
Handling of Unreferenced Fields: Corrected storage behavior for fields that weren’t indexed.
Similarity Defaults: Parameters now default to a value of one if not explicitly provided.
StringLists in NLQ: Natural Language Queries can now dynamically populate StringLists.
Categorical Similarity Nodes: Fixed bugs affecting event handling in categorical similarity nodes.
Error Handling for Unknown IDs: Queries now return an error instead of similarity results when an invalid ID is used.
Other Enhancements
NLQ Improvements: Natural language querying has been enhanced for better accuracy.
NLQ Example Updates: New examples now showcase product searches instead of reviews.
Logging Upgrades: Expanded logging coverage for external message bus interactions.
Test Performance Optimizations: Reduced execution time from 4.6s to 0.6s.
23-10-2024
Framework/Server/Batch: 10.1.0/10.1.0/1.11.1 → 12.2.0/12.2.0/1.13.1
New Features
Debugging Enhancements: JSON logs now include stack traces for better troubleshooting.
Fixes
Schema Dependencies Handling: Now supports scenarios where Schema A & B influence Schema C, improving accuracy in behavioral models.
Registry Stability: Fixed an internal bug in schema registration.
09-10-2024
Framework/Server/Batch: 9.43.0/9.42.1/1.8.0 → 10.1.0/10.1.0/1.11.1
New Features
Optional Query Parameters: Queries will now ignore parameters if they’re not provided.
Notebook Compatibility for VDB: The new InteractiveExecutor allows seamless vector database connections from notebooks.
Simplified Query Syntax: Users can now define queries more intuitively, e.g.,
.similar(space, 3)instead of.similar(space.number, 3).Enhanced JSON Logs: Stack traces are now included in logs for better debugging.
Fixes
Notebook Plot Rendering: Fixed compatibility issues in various environments.
Redis Query Errors: Addressed a bug where certain queries returned incorrect results due to naming conflicts.
NLQ Filtering Accuracy: Prevented NLQ from generating categories that don’t exist.
25-09-2024
Framework/Server/Batch: 9.33.0/9.33.0/1.4.0 → 9.43.0/9.42.1/1.8.0
New Features
Expanded NLQ Operators: IN and NOT_IN operators are now supported in natural language queries.
Unified Logging for Qyver Components: Centralized logging improves debugging visibility.
Text Embedding Caching: Added caching for 10,000 items to speed up repeated queries.
Fixes
Chunking & Hard Filters: Ensured chunking doesn’t interfere with filtering constraints.
Category Filtering in NLQ: NLQ now correctly adheres to available categories, preventing false results.
Changes
Updated Image Embedding Model: Default embedding model in notebooks has been replaced.
11-09-2024
Framework/Server/Batch: 9.22.1/9.33.0/1.1.2 → 9.33.0/9.33.0/1.4.0
New Features
Modular Code Structure: The app is now split into
index.py,query.py, andapi.pyfor improved maintainability.Additional Operators for Notebooks: OR and CONTAINS operators added based on user feedback.
Fixes
Recency Weighting Fix: Addressed an issue where full zero vectors affected weight calculations.
NLQ Hard Filter Compatibility: Ensured new hard filters work as expected.
Chunking & Hard Filters: Fully tested and confirmed working correctly.
Changes
Improved Startup Messages: Clarified server startup logs to reduce user confusion.
28-08-2024
Framework/Server/Batch: 9.12.1/9.12.1/1.1.0 → 9.22.1/9.22.0/1.1.2
New Features
Event Handling in Batch Processing: Batch jobs can now include event-based calculations.
Expanded Hard Filter Support: Added operators such as LT, LTE, GT, GTE, AND, OR, CONTAINS, NOT_CONTAINS, IN, and NOT_IN.
Optimized GPU Utilization: The system now automatically switches between CPU and GPU based on dataset size (~10k embeddings threshold).
GPU Detection in Tests: The framework now verifies GPU availability and utilizes it when possible.
Score Display in Notebooks: Vector similarity scores are now visible by default in examples.
NLQ Filtering & Temperature Adjustments: Improvements based on user feedback.
Fixes
Redis Query Ordering: Results are now properly sorted.
Embedding Normalization: Long categorical embeddings no longer distort results.
Consistency in
source.put: Standardized behavior across various input types.Index Temperature Handling: Integer values are now correctly processed.
Changes
Formatted Logs in Tests: Enhanced readability for logs during testing.
(Breaking Change) Removed Status Endpoints: This change optimizes for a stateless executor with high availability.
14-08-2024
Framework/Server/Batch: 9.7.0/9.6.0/1.0.2 → 9.12.1/9.12.1/1.1.0
New Features
Default Similarity Scores: Included in results by default for deeper insight into ranking distributions.
Expanded Feature Notebooks: Added examples for similarity scores, recency queries, and event-based parameters.
Automated Code Quality Checks: Ensured consistent testing and formatting in server code.
Fixes
Recency Context Alignment: Now uses a unified reference point for consistency.
07-08-2024
Framework/Server/Batch: 9.7.0/9.6.0/1.0.2
New Features
Updated Sentence-Transformers to 3.0.1: Now supports the top-performing models from the MTEB leaderboard.
Support for Empty List Embeddings: Allows datasets where not all rows contain embeddings.
Default Query Limits: Redis and Mongo connectors now return a default of 10 items.
NLQ Parameterization: Users can now define query parameters dynamically.
Logarithmic Number Embeddings: Supports non-linear scaling for large numerical values.
Fixes
Negative Weighting in Recommendations: Fixed cases where negative event weights were incorrectly handled.
Last updated