API ReferenceDocumentation
Documentation

TetraScience AI Services v1.4.x Release Notes

Release notes for TetraScience AI Services v1.4.x

The following are the release notes for TetraScience AI Services versions 1.4.x. TetraScience AI Services empower developers and scientists to transform data into decisions through seamless, secure APIs, and enable intelligent automation across scientific, lab, and enterprise workflows.

📘

NOTE

TetraScience AI Services are currently part of a limited availability release and must be activated in coordination with TetraScience along with related consuming applications. For more information, contact your customer account leader.

v1.4.1

Release date: 4 November 2026

TetraScience AI Services v1.4.1 adds queued batch inference for Bring Your Own Model (BYOM) on the AI Runtime, ties each synchronous inference request to its audit record, improves telemetry sink management, and fixes bugs in vector stores, AI Workflow installation, multipart uploads, and BYOM model metadata.

Here are the details for what's new in TetraScience AI Services v1.4.1.

Prerequisites

TetraScience AI Services v1.4.1 requires the following:

New Functionality

New functionality includes features not previously available in TetraScience AI Services.

Bring Your Own Model on the AI Runtime

You can now run BYOM inference as queued batch jobs on a pool of GPU workers that stays running for the model, instead of starting a Databricks job for each request. This suits models that process many inputs at once and models with weights too large to load on every request. POST /ai-platform/v1/ai-runtime/inference queues a job and returns a run identifier. GET /ai-platform/v1/ai-runtime/inference/{runId}/status returns the job's status, output location, and any error message.

Bring Your Own Model on the AI Runtime is a beta release. To get access, contact your customer account leader.

Enhancements

Enhancements are modifications to existing functionality that improve performance or usability, but don't alter the function or intended use of the system.

  • Each synchronous inference request is tied to its audit record: a synchronous inference response includes a sync_<uuid> request identifier. The same identifier is the target.id of the ai-workflow.inference-sync.v1 event in the Tetra Platform System Log, which can take several minutes to appear after the call, and is stored as client_request_id on the matching row of the Databricks inference table.
  • Telemetry sink changes are attributed: a telemetry sink record now stores who made each change and when. Changes made by TetraScience are attributed to the provisioning service.
  • A telemetry sink can be moved to a different SQL warehouse: If your telemetry sink needs to use a different SQL warehouse, TetraScience Support can now move the existing sink instead of provisioning a new one. Existing clients continue to work.

Bug Fixes

The following bugs are now fixed:

  • Vector store job failures can now be diagnosed: vector store jobs now write their output, including Python tracebacks, to the Log File Location shown on the job run's details page in Tetra Platform.
  • Deleting a vector store removes its Databricks resources: DELETE /ai-platform/v1/vector-store/{id} now removes the vector search index and endpoint, the Delta table, and the knowledge base files. The request returns 202 and completes in the background, 404 for an unknown store, and 409 while a create or re-index job is running. Send force=true to cancel the job and delete anyway.
  • AI Workflow installations with long bundle deployments complete: an installation whose Databricks bundle takes more than 2 minutes to deploy no longer stays at Provisioning Infrastructure indefinitely. Installations now wait up to 10 minutes for the deployment.
  • Invalid install and uninstall requests return a client error: uninstalling a version that isn't installed, installing a version that doesn't exist, or repeating an install that is still in progress now returns a 4xx error instead of success.
  • An incorrect ETag when completing a multipart upload returns 400: completing a multipart upload with an incorrect ETag now returns 400 instead of 500, so you can re-upload the part.
  • BYOM model metadata no longer repeats training information: a BYOM AI workflow's stored model metadata no longer carries the same training information under both training and training_parameters.

Deprecated Features

This release deprecates no features. For more information about Tetra Platform deprecations, see Tetra Product Deprecation Notices.

Known and Possible Issues

The following are known limitations of TetraScience AI Services v1.4.1:

  • BYOM Registration Does Not Enforce Organization Access Settings: An organization whose Bring Your Own Model access is disabled can still register models through the API. Do not rely on the BYOM access setting to prevent model registration. Only users with a role that includes the Tenant Admin, Organization Admin, Developer, or Machine Learning Engineer policy can register models. To restrict registration, limit which users have these roles.
  • Telemetry Does Not Enforce Organization Access Settings: An organization whose AI observability and telemetry access is disabled can still use telemetry through the API. Do not rely on the telemetry access setting to disable telemetry. Telemetry requests still need an active telemetry sink that TetraScience provisions for your organization and a Tetra token authorized for your organization. Upgrading AI Services doesn't start telemetry collection.
  • Trace Search Isn't Bounded by Time: A trace search request isn't bounded by a time range, so a search can return traces across the full history of a sink. Result counts are capped, and you can narrow results by filtering on time in your query.
  • No Telemetry Deprovisioning Path: A provisioned telemetry sink can't be removed through any supported path. Contact TetraScience Support before removing a telemetry sink.
  • Vector Store Update Status Remains syncing After a Re-index Job Completes: After a vector store update is triggered, the status returned by the vector store API might remain syncing even after the re-index job completes successfully. The updated content is indexed and queryable, but because the status doesn't transition back to ready, any automated process that polls the status to confirm completion can wait indefinitely. To verify that an update completed, query the vector store and confirm that the results reflect the updated content. TetraScience is working with the vendor on a fix.
  • AI Workflows Generated by ts-cli Need Manual Edits Before They Install: An AI workflow created with ts-cli init ai-workflow fails to install until the generated files are edited. The corrected template ships with the next ts-cli release. Until then, contact TetraScience for the required changes.
  • Model Training UI Upload Failure: The Upload Supporting Files button in the AI Services UI does not work for model training workflows. File uploads fail because of an incorrect S3 file path constructed by the UI. As a workaround, upload model training files and initiate model training by using the AI Services API directly. The related cross-origin (CORS) error was fixed in Tetra Platform v4.6.0.
  • Analyst Policy Install Workflow UI Access: On Tetra Platform v4.4.1, users assigned a role with an Analyst policy can view, but not use, the Install Workflow button even though the policy doesn't allow them to install an AI workflow or run an inference. This issue is fixed in Tetra Platform v4.4.2 and later.
  • GPU Worker Pool Can Shrink After a Quiet Period: If you use Bring Your Own Model on the AI Runtime, idle GPU workers can exit after a period with only occasional single jobs and aren't restarted. A later burst of jobs can run on one worker instead of in parallel. The jobs complete, but more slowly.
  • A Queued AI Runtime Job Can Run Twice If a GPU Worker Stops Responding: If you use Bring Your Own Model on the AI Runtime and a GPU worker stops reporting for too long while it is still running a job, the job is handed to another worker. If the first worker later finishes, both write the job's status and output, and the last write wins. If a job's output doesn't match its reported status, run it again.
  • Agent-Deployed Serving Endpoints Serve Only the Most Recently Installed Version: If an AI workflow's installation deploys its own model serving endpoint, such as an agent deployed with databricks.agents.deploy, all installed versions share that endpoint, which serves only the version installed last. A POST /ai-platform/v1/inference/online request for an older version runs the latest installed model and returns success, while the request and its System Log record show the version you asked for. Endpoints enabled through POST /ai-platform/v1/inference/online/enable are not affected, because each version has its own endpoint. To run an older version, uninstall it and install it again.
  • Real-Time Inference Can't Be Enabled for Bundle-Served Versions: When a version's installation job created its own serving endpoint, POST /ai-platform/v1/inference/online/enable returns 422 saying the version must be installed, and GET /ai-platform/v1/inference/online doesn't list the endpoint. Inference requests to the version still succeed.
  • A Forced Re-Install Leaves the Workflow Registration Inconsistent: POST /ai-platform/v1/install?force=true for an installed version runs the installation job but doesn't re-register the workflow, and the response echoes an earlier request. Uninstall and install again instead.
  • An Install Request Can Intermittently Fail with a 502 Response: POST /ai-platform/v1/install can occasionally return 502 with Failed to validate artifact version: Request timeout. No installation starts. Retry the request.
  • Install Status Omits the Serving Endpoint Details: GET /ai-platform/v1/install/{slug} no longer returns the outputs field (serving endpoint, model, alias, and AI Runtime readiness details) for a successful install in a private-* or client-* namespace. The install succeeds and inference works. Use the inference links in the response, or run inference through POST /ai-platform/v1/inference/online, which resolves the endpoint for you.
  • Re-Installing an Installed Version Starts a Hidden Second Install: POST /ai-platform/v1/install for a version whose installation has already completed returns 201 and runs a new installation, but GET /ai-platform/v1/install/{slug} keeps returning the original request, so you can't see the re-install's progress or result. To re-install a version, uninstall it and install it again.
  • BYOM Registration Reports Success Before the Artifact Is Published: POST /ai-platform/v1/byom/register-start reports success while the artifact is still being built, so approving the version right away returns a 400 response that says the version isn't found. The delay can be seconds to minutes. Retry the approval.
  • Batch Inference Status Response Doesn't Match the API Reference: GET /ai-platform/v1/inference/{inferenceId}/status returns SUCCESS and FAILED with results under outputFiles. The API reference documents completed, failed, and files. Poll for SUCCESS or FAILED. For the full list of status values, see Get Inference Status in the TetraScience AI Services User Guide.
  • Vector Store Sentence Chunking Needs Internet Access: The sentence chunking strategy downloads a tokenizer at run time, so it fails on clusters without outbound internet access. Use the default character strategy there.
  • A Transient Token Error Fails LLM Requests for Up to 50 Seconds: After a single transient authentication failure, every /ai-platform/v1/llms/* request returns 500 for up to 50 seconds. Retry after that interval.
  • ts-cli publish --dry-run Exits with Code 1 and ts-cli unpublish Needs a Terminal: A successful dry run exits with code 1, and the unpublish prompt can't read piped input, so neither can be scripted. Remove a version with DELETE /v1/artifact/ai-workflows/{namespace}/{slug}/{version} instead.

Upgrade Considerations

To upgrade to the latest AI Services version, see Update the AI Services UI Version in the TetraScience AI Services User Guide.

If you have automation that deletes vector stores, update it for the new DELETE /ai-platform/v1/vector-store/{id} responses (202, 404, and 409) described under Bug Fixes.

After the upgrade, verify the following:

  • Existing AI Workflow installations continue to function correctly
  • Inference requests complete successfully
  • Knowledge base and vector store operations function as expected
  • Asset promotion completes successfully for models, volumes, and aliases
  • If telemetry has been enabled for your organization, telemetry writes and trace read-back complete successfully
  • The AI Services UI displays correctly

GxP Impact Assessment

All new TetraScience AI Services functionalities go through a GxP impact assessment to determine validation needs for GxP installations.

Enhancements and Bug Fixes do not generally affect Intended Use for validation purposes.

TetraScience AI Services is currently a limited availability release. Model Promotion, Bring Your Own Model as a container image, and Bring Your Own Model on the AI Runtime are beta releases. Items marked as either a beta release or a limited availability release are not validated for GxP by TetraScience. However, customers can use these prerelease features and components in production if they perform their own validation.

📘

Security

TetraScience continually monitors and tests the codebase to identify potential security issues. Security updates are applied on an ongoing basis.

📘

Quality Management

TetraScience is committed to creating quality software. Software is developed and tested following the ISO 9001-certified TetraScience Quality Management System.

For instructions on how to validate core AI Services functionality programmatically, see Test TetraScience AI Services Programmatically in the TetraConnect Hub. For access, see Access the TetraConnect Hub.

For more information, see the TetraScience AI Services documentation.

v1.4.0

Release date: 8 October 2026 (updated 3 November 2026)

TetraScience AI Services v1.4.0 adds support for registering Bring Your Own Model (BYOM) models as container images and fixes bugs in private-namespace publishing, AI Workflow installation, batch inference, model promotion, and BYOM registration.

Here are the details for what's new in TetraScience AI Services v1.4.0.

Prerequisites

TetraScience AI Services v1.4.0 requires the following:

New Functionality

New functionality includes features not previously available in TetraScience AI Services.

Bring Your Own Model as a Container Image

You can now register a model packaged as a container image and receive an inference endpoint for it, instead of supplying model artifacts alone. To register, save the container image as a TAR archive and publish it with ts-cli.

Bring Your Own Model as a container image is a beta release. To get access, contact your customer account leader.

Enhancements

Enhancements are modifications to existing functionality that improve performance or usability, but don't alter the function or intended use of the system.

  • Pay-per-token LLM serving endpoints no longer mask personal information: AI Services no longer applies a personally identifiable information (PII) masking guardrail to pay-per-token LLM serving endpoints, including endpoints configured before the upgrade. Request content, including any personal information it contains, now reaches the model unmasked. If your use of these endpoints depends on PII masking, contact TetraScience before you upgrade.

Bug Fixes

The following bugs are now fixed:

  • Model promotion returns a clear error for unsupported asset types: promoting a view now returns a client error that names the problem, instead of a server error.
  • Model promotion to a different region works: promoting a model to a target environment in a different AWS region from the source now succeeds.
  • Private-namespace publishing enforces permissions, file size, and namespace ownership: publishing to a private-* namespace now checks the caller's permissions, enforces the file-size limit, and confirms the caller's organization owns the namespace.
  • AI Workflow installations no longer appear stuck: an installation now completes when its Databricks job succeeds, even if the completion event is missing.
  • AI Workflow installation completes when an alias update fails: a failed alias update no longer aborts the whole installation or leaves the installation record unusable.
  • Installation records include the namespace: GET /v1/install returns the top-level namespace field again.
  • Batch inference works for existing AI workflows: batch inference no longer fails with a missing-weights error for AI workflows created before the upgrade.
  • Telemetry trace search works as documented: the telemetry trace search request now matches the published API reference.
  • BYOM retraining keeps the trained model: retraining a BYOM AI workflow saves the trained model again.
  • BYOM AI Workflow installation completes: installation no longer fails while registering the inference endpoint.
  • BYOM registration rejects a container image that doesn't exist: registering a model against a missing image now fails, instead of reporting success.
  • BYOM registration accepts organization and tenant names containing hyphens or uppercase letters.
  • Re-uploading an unchanged BYOM version is stable: re-uploading the same version no longer changes its manifest identifier each time.
  • BYOM scaffolding targets your environment: generated scaffolding no longer contains a hard-coded internal environment.

Deprecated Features

This release deprecates no features. For more information about TDP deprecations, see Tetra Product Deprecation Notices.

Known and Possible Issues

The following are known limitations of TetraScience AI Services v1.4.0:

  • BYOM Registration Does Not Enforce Organization Access Settings: An organization whose Bring Your Own Model access is disabled can still register models through the API. Do not rely on the BYOM access setting to prevent model registration. Only users with a role that includes the Tenant Admin, Organization Admin, Developer, or Machine Learning Engineer policy can register models. To restrict registration, limit which users have these roles.
  • Telemetry Does Not Enforce Organization Access Settings: An organization whose AI observability and telemetry access is disabled can still use telemetry through the API. Do not rely on the telemetry access setting to disable telemetry. Telemetry requests still need an active telemetry sink that TetraScience provisions for your organization and a Tetra token authorized for your organization. Upgrading AI Services doesn't start telemetry collection.
  • Telemetry Sink SQL Warehouse Is Fixed at Creation: A provisioned telemetry sink's SQL warehouse can't be changed after the sink is created. To use a different warehouse, contact TetraScience to provision a new sink.
  • Sink Change History Isn't Tracked: A sink record shows who created it, but not who has changed its configuration since.
  • Trace Search Isn't Bounded by Time: A trace search request isn't bounded by a time range, so a search can return traces across the full history of a sink. Result counts are capped, and you can narrow results by filtering on time in your query.
  • No Telemetry Deprovisioning Path: A provisioned telemetry sink can't be removed through any supported path. Contact TetraScience Support before removing a telemetry sink.
  • Vector Store Deletion Does Not Remove Databricks Resources: Deleting a vector store removes the internal database record, but the associated Databricks vector search endpoint and index, the Delta tables, and the knowledge base files in Amazon S3 are not removed and require manual cleanup. Contact TetraScience Support to remove the orphaned resources.
  • Vector Store Update Status Remains syncing After a Re-index Job Completes: After a vector store update is triggered, the status returned by the vector store API might remain syncing even after the re-index job completes successfully. The updated content is indexed and queryable, but because the status doesn't transition back to ready, any automated process that polls the status to confirm completion can wait indefinitely. To verify that an update completed, query the vector store and confirm that the results reflect the updated content. TetraScience is working with the vendor on a fix.
  • Some Install and Registry Requests Report Success When They Should Be Rejected: For example, uninstalling an AI workflow version that is not installed returns a success message. Confirm the result by reading the installation or registry state after the request.
  • AI Workflows Generated by ts-cli Need Manual Edits Before They Install: An AI workflow created with ts-cli init ai-workflow fails to install until the generated files are edited. Contact TetraScience for the required changes.
  • An Incorrect ETag When Completing a Multipart Upload Returns a Server Error: Completing a multipart file upload with an incorrect ETag returns a 500 response instead of a 400, and marks the file as failed. Restart the file upload instead of retrying the completion request.
  • BYOM Model Metadata Can Repeat Training Information: A BYOM AI workflow's stored model metadata can contain the same training information under both training and training_parameters. The AI Services UI is unaffected.
  • Model Training UI Upload Failure: The Upload Supporting Files button in the AI Services UI does not work for model training workflows. File uploads fail because of an incorrect S3 file path constructed by the UI. As a workaround, upload model training files and initiate model training by using the AI Services API directly. The related cross-origin (CORS) error was fixed in Tetra Platform v4.6.0.
  • Analyst Policy Install Workflow UI Access: On Tetra Platform v4.4.1, users assigned a role with an Analyst policy can view, but not use, the Install Workflow button even though the policy doesn't allow them to install an AI workflow or run an inference. This issue is fixed in Tetra Platform v4.4.2 and later.

Upgrade Considerations

To upgrade to the latest AI Services version, see Update the AI Services UI Version in the TetraScience AI Services User Guide.

If your use of pay-per-token LLM serving endpoints depends on PII masking, contact TetraScience before you upgrade. See the PII masking entry under Enhancements.

After the upgrade, verify the following:

  • Existing AI Workflow installations continue to function correctly
  • Inference requests complete successfully
  • Knowledge base and vector store operations function as expected
  • Asset promotion completes successfully for models, volumes, and aliases
  • If telemetry has been enabled for your organization, telemetry writes and trace read-back complete successfully
  • The AI Services UI displays correctly

GxP Impact Assessment

All new TetraScience AI Services functionalities go through a GxP impact assessment to determine validation needs for GxP installations.

Enhancements and Bug Fixes do not generally affect Intended Use for validation purposes.

TetraScience AI Services is currently a limited availability release. Model Promotion and Bring Your Own Model as a container image are beta releases. Items marked as either a beta release or a limited availability release are not validated for GxP by TetraScience. However, customers can use these prerelease features and components in production if they perform their own validation.

📘

Security

TetraScience continually monitors and tests the codebase to identify potential security issues. Security updates are applied on an ongoing basis.

📘

Quality Management

TetraScience is committed to creating quality software. Software is developed and tested following the ISO 9001-certified TetraScience Quality Management System.

For instructions on how to validate core AI Services functionality programmatically, see Test TetraScience AI Services Programmatically in the TetraConnect Hub. For access, see Access the TetraConnect Hub.

For more information, see the TetraScience AI Services documentation.

Other Release Notes

To view other TetraScience AI Services release notes, see TetraScience AI Services Release Notes.

Updated 8 days ago

Did this page help you?