Skip to main content
Version: 3.3.4

Release notes

Version 3.3.4​

About Release
  • Release date: July 3, 2025
  • Compatible platform versions: Work.AI 10.2.8+

The Tara v3.3.4 release aims to improve address matching flexibility, enhance entity recognition accuracy, enrich decision traceability, and support better precision in risk screening.

  • The logic is updated to parse known SWIFT tag values inside FUF content, primarily affecting F tags, such as SWF_4_50F and SWF_4_59F. The full address is now correctly set after parsing the SWIFT content.

  • To better handle * characters within the content, we updated the logic by removing 'New Lines From 36th Positions'.

  • A detailed decision narrative is now included in the HTML report and hit decision response, making Tara's decision-making fully transparent without requiring inspection of underlying data objects. A new optional additionalInfo field is available in both the response object and the HTML report. You can also enable this feature in the configuration UI.

  • Support is added for partial (fuzzy) address match scoring when external services return no results or are not configured. The enhancement helps escalate potentially suspicious transactions even when exact address matches are unavailable.

    • A probabilistic score is calculated using string similarity and structured data from the external service.

    • The addressMatch rule object now includes a new score parameter, and a sample custom rule is available for integration.

  • To prevent typing errors and ensure valid, standardized country entries, the free-text input for the Blacklisted countries field is replaced with a pre-populated country selection with validation and automatic suggestions available.

  • The narrative is updated for hits with empty tag content, improving clarity.

  • During historical data upload, analyst comments from messages.csv can now be applied to all hits, not just records with a single hit. A new setting in the configuration JSON controls this behavior, which is enabled by default.

  • The NER model is upgraded to v2.4.0 with the following improvements:

    • Enhanced model prediction accuracy for U.S. and Maldives addresses, Middle Eastern company names, and long personal names

    • Resolved BIC recognition issue where “ARQUITECTOS” was incorrectly classified as a BIC

    • Improved recognition quality when named tags are present in text, for example, "/ABIC/SBOSUS33 /NAME/Truist Financial MEXICO MEXICO"

    • Disabled merging of adjacent PER and ORG entities during post-processing

    • New post-processor to filter out overlapping entities

  • The AI Agent configuration is upgraded with the decision model v5.1.7, which includes:

    • Logic to stop parsing NER-based addresses when screened location or enrichment locations are already present, preventing incorrect country attribution. For example, "PHOENIX AZ" was misidentified as "Azerbaijan".

    • Improved Libpostal parsing of country names in SWIFT format where an ISO2 country code is followed by a slash and city name, for example, US/Boston.

  • The Name Matcher is upgraded to v1.1.4, incorporating the following updates:

    • Resolved issue where share listing information was not excluded from ORG name comparisons

    • Revised list of stop words

    • Nickname support for Russian, Italian, and French names

    • Improved handling of company acronyms and name tokens split by spaces

Bug fixes

  • Resolved the date conversion issue in the CLDS report by expanding the range of supported date formats.

  • Fixed the issue in computing full addresses.

  • Fixed the issue in entity resolution where the hit text displayed a country name, but the actual match was based on the ISO2/ISO3 code, leading to incorrect entity classification.

  • Resolved the failure in saving large (> 64 characters) tag name values to the pss_alerts_analytics_v1 Data Store. Tag names are now truncated from the beginning, cleanly breaking on slashes.

  • Fixed the compatibility issue where the Geocoder returned no Location object for unresolved addresses, causing decision models v5.1.2 and earlier to fail. A blank Location object is now inserted to ensure backward compatibility.

  • Resolved the entity resolution issue where the NER-detected entity did not directly match the hit text or sanctioned name, for example, "AL-RASHID" versus "ALRASHID" or "Al'RASHID".

Last updated on