We’re excited to announce the release of version 1.15.0 of the Certara.AI platform! This release brings a host of new features, performance improvements, and critical fixes across multiple components of the ecosystem. Here's a breakdown of what's new:
🧠 Cortex
Features:
Introduced caching for RAG document sets to improve performance.
Switched to distributed search for embedding creation and document retrieval.
Expanded data fabric integration across services, including chemistry jobs and document processing.
Added endpoints to list and download models for easier model management.
Enabled streaming responses for the OpenAI-compatible
/completionsendpoint.Added support for key/value metadata tags on sources used in GPT generate requests.
Fixes:
Various bug fixes and performance optimizations to improve stability and efficiency.
🧪 Curate
Features:
Added support for multi-label classification models, enabling more flexible data labeling.
Introduced the ability to include metadata when creating questions for enhanced context.
Fixes:
Resolved text rendering issues for bioRxiv and medRxiv sources.
Improved error handling for document metadata and classification edge cases.
🌐 Layar
Features:
N/A
Fixes:
Fixed an issue where the welcome tour displayed too frequently, improving the onboarding experience.
🧬 Reference Data Fabric
Features:
Updated the ClinicalTrials.gov connector to use the new v2 schema API for improved compatibility.
Fixes:
Improved error handling and schema validation for external data sources.
🧬 Chemistry & Featurizers
Features:
Enabled users to select and configure featurizers for chemistry models.
Added endpoints for inserting SMILES strings and source IDs.
Fixes:
Addressed issues with featurizer configuration persistence and validation.
🛠️ DevOps & Infrastructure
Features:
Upgraded vector database to v1.30.2 for better performance and stability.
Enabled backups for vectorstore deployments and improved monitoring.
Updated Helm charts from v0.5.521 to v0.5.552, adding support for new clients and environments.
Fixes:
Improved deployment reliability and monitoring coverage across services.
🌐 Web Applications
Features:
Composer now supports rendering Markdown-formatted LLM responses, improving readability of tables, links, and structured content.
Fixes:
Enhanced retry logic and error handling in Composer for a more robust user experience.
🔍 Certara-GPT
Features:
Overhauled tokenizer caching to reduce startup time and memory usage.
Added support for DeepSeek, Blossom, and LLaMA 4 models, including Korean variants.
Enhanced logging for tokenizer and document source issues.
Improved error messages and fallback mechanisms for tokenizer creation.
Updated vector database schema to support dynamic tokenization and better memory management.
Fixes:
Resolved issues related to tokenizer initialization and document source resolution.
🧠 LLM & RAG Pipeline
Features:
Enabled RAG skipping by default to improve performance in applicable scenarios.
Enabled streaming for OpenAI-formatted endpoints to support real-time interactions.
Added retry logic for specific error types without skipping RAG.
Streamlined validation steps during the RAG process to reduce latency.
Fixes:
Improved error handling and fallback logic during RAG execution.