Release notes

Version 3.4.1

About Release

• Release date: October 17, 2025
• Compatible platform versions: Work.AI 10.2.8+

The Tara 3.4.1 release enhances the data enrichment, entity recognition, and decision intelligence capabilities to strengthen data validation, reduce processing errors, and deliver more consistent decision outcomes.

• The GLEIF connector is added as a separate step to the PSS Enrichment Connectors Business Process for entity information lookup, supporting entity name validation and address enrichment. The new service allows a configurable search for company information using security identifiers such as LEI, ISIN, or BIC, with multi-ID batch calls and caching by source.

• The sanctions-api-client is upgraded to v1.2.10. The SecurityInfo object now includes additional fields from GLEIF enrichment, which enables richer entity profiles and improved compliance checks.

• The AI Agent configuration is upgraded with the Decision model v5.2.0, introducing the following improvements:

  • Added support for the SecurityData object from enrichment data to extract names for enhanced name-matching accuracy.

  • Improved the decision-making process, resolving the issue where the model returned inconsistent best-matched aliases for the same input. When multiple names on one side produce the same matching score, Tara now sorts names by length to ensure deterministic and reliable outputs.

• The NER model is upgraded to v2.6.6, delivering higher precision in entity extraction and improved model robustness through the following updates:

  • Enhanced recognition of ISO codes as LOCATION entities.

  • Improved ID handling by excluding specific false-positive patterns.

  • Refined quote and punctuation preprocessing for cleaner tokenization.

  • Enhanced preprocessing to split alpha prefixes from numeric suffixes when the suffix contains [./-].

  • Improved validation when, on invalid input, the model returns a single OUTSIDE entity (confidence 0.1) and logs a warning to ensure traceability.

  • Improved location detection by reducing confidence when no digits are present and every word has a long consonant sequence (more than five consonants).

  • Refined organization detection:

    • Organization name merges are only allowed when the second token is a pure legal or business indicator.

    • The confidence for organization-like names that resemble locations (for example, those containing strong street, avenue, or parkway markers) is reduced to minimize false positives in mixed-address text.

  • Cleaner tokenization around the word BANK, ensuring that real words like banks, banking, banker(s), and bankruptcy are no longer split, while concatenated names (for example, BANKOFAMERICA → BANK OF AMERICA) are still separated for better recognition and consistency.

  • Minor stability and configuration updates.

Bug fixes

• Fixed null pointer exceptions during date conversion in message parsing, which previously caused incomplete records and quality control report failures when some data fetched from the pss_processing Data Store lacked decisions.

• Resolved the issue where Orbis query results occasionally returned null or missing values.

• Resolved the issue with NER input truncation for multi-tag content beyond the 35th character. Tags with the same name (without numerical extension, such as "[ABC]") and values shorter than 36 characters are now joined into full text, improving data integrity and recognition accuracy.