Released versions
Release notes for generally available versions are listed below.
Versions 43.0.x, 43.1.x, 43.2.x are available for SaaS customers only.
43.2.2 (9 Sep 2026)
Highlights
As a leader in intelligent document processing, Hyperscience continues to advance its platform with greater accuracy, flexibility, and ease of use. Hyperscience v43.2 introduces ORCA 2, simpler model management, and improvements to your workflows. Read on for a closer look at what’s new.
ORCA
ORCA 2 — v43.2 introduces ORCA 2, our latest Vision Language Model (VLM). ORCA 2 improves extraction accuracy across a wide range of document types. Learn more in ORCA Vision Language Models.
Out-of-the-box table extraction — ORCA 2 can extract tables as rows and columns without requiring you to train a specialized model. To learn more, see General Prompting Block.
You can continue using ORCA 1.0 and specialized models trained on ORCA 1.0 while transitioning to ORCA 2.
ORCA 1.0 remains supported in v43.2, but is planned for deprecation in a future release.
We recommend transitioning your specialized models to ORCA 2 for better accuracy and automation across document groups.
The ORCA base models available under Administration > Assets depend on your organization’s license.
You can install one base model at a time. Learn more in Installing ORCA VLMs.
Configure model selection in the Document Processing with ORCA subflow — In v43.2, the ORCA Extraction Block includes a Use live model setting. This setting determines whether the live model version in TDM or the ORCA base model selected in the flow controls processing.
Choose between managing specialized models centrally through VLM Field Extraction model definitions or keeping a flow tied to a particular ORCA base model.
Hyperscience also verifies base-model compatibility, preventing a model version trained with ORCA 1.0 from being used when the flow is configured for ORCA 2, or vice versa.
You can configure this in Flow Studio (Flows > Edit flows > Settings Type — ORCA).
.png?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
When Use live model is selected:
Hyperscience finds the VLM Field Extraction associated with the layout.
If the Model Definition has a live specialized model, Hyperscience uses that version.
If no live model version is available for the layout, processing stops, and the flow returns an error.
The base model selected in the flow does not control processing.
When Use live model is not selected:
The ORCA base model selected in the flow controls processing.
Hyperscience uses a live model version only when it belongs to the matching VLM Field Extraction model definition and was trained with the selected base model.
A live model version trained with a different base model is ignored. For example, if the flow specifies ORCA 2, Hyperscience does not use a live model version trained with ORCA 1.0.
If no compatible live model version is available, the system uses a compatible model version when available.
Learn more in Document Processing Flow.
Manage multiple ORCA base models in Training Data Management (TDM) for VLM Field Extraction — In v43.2 and later, you can install and manage multiple ORCA base models, such as ORCA 1.0 and ORCA 2, in TDM for VLM Field Extraction.
Multiple base-model support lets you adopt ORCA 2 gradually:
You can train and evaluate specialized models built with ORCA 2, while continuing to use models trained with ORCA 1.0.
On the Overview tab in TDM for VLM Field Extraction, the Required Documents indicator shows how many eligible documents you have and the lowest number of documents you need before training a specialized model.
Hyperscience also prevents training and deployment when the required base model is unavailable, reducing configuration errors.
The History tab shows the base model used by each specialized model.
To learn how to work with TDM for ORCA VLM Extraction, see TDM for ORCA VLMs. Learn how to specialize a model in Training a Specialized Model.
If a base model is not installed, but specialized models trained with it already exist:
On the VLM Field Extraction page, affected model definitions are grayed out. To learn more, see Model Definitions.
On the Model Details page, specialized model versions trained with the missing base model are also grayed out, and Deploy is unavailable.
You can still download or archive the affected specialized model versions.
The base model’s row on the History tab in TDM for VLM Field Extraction displays Not installed and includes an Install action.
To install the base model, click Install. This opens the Assets page (Administration > Assets). Learn more in Installing ORCA VLMs.
User Experience
Updated
Updated keyboard shortcuts for Supervision and QA tasks — Keyboard shortcuts have been standardized across supported Supervision and QA tasks on Mac and Windows.
The redesigned keyboard shortcuts panel displays the shortcuts in a list
You can search by action to see the key combination for your operating system.
The updated shortcuts are off by default in v43.2.
To try them, turn on the new shortcuts from the keyboard shortcuts panel.
You can turn them off at any time to continue using the previous shortcuts.
.gif?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
For a full list of the updated shortcuts, see Keyboard Shortcuts.
Flows
New
Parse and standardize addresses — In v43.2 and later, the out-of-the-box Document processing flow can parse address transcriptions into components such as house number, street, city, state, and postal code, and standardize them for use in downstream systems. This provides a consistent alternative to creating custom address-processing logic in each flow.
Address parsing is off by default and configured by flow. When it is disabled, address fields are returned as plain transcription strings without parsing or standardization.
To enable address parsing:
Go to Flow Studio (Flows > Edit flows).
Select Address Parsing from the Settings Type drop-down list.
Select the checkbox for structured address parsing.
You can use the built-in US or international address parser or connect to a supported external provider, including Google Geocoding, Smarty, or Geoapify.
All supported parsers return address components in a consistent format. Parsed addresses appear in the Machine Transcription, Manual Transcription, and Flexible Extraction Block outputs.
To learn more, see Document Processing Flow.
Updated
Faster PDF Transcription for additional file types — This feature reduces transcription time, enabling documents to move through your workflow faster. In v43.2 and later, it supports .docx, .xlsx, .eml, .msg, .html, and .xps files.
Additionally, in v43.2 and later, you can use Faster PDF Transcription with Image Correction. Previously, enabling Image Correction prevented you from using Faster PDF Transcription.
For documents containing both digital and scanned pages, Hyperscience uses embedded text where available and applies Image Correction to the remaining pages.
Faster PDF Transcription is off by default and is configured separately for each flow. To enable it:
Go to Flow Studio (Flows > Edit flows).
Select Classification from the Settings Type drop-down list.
Select Faster PDF Transcription. To use both features, also select Image Correction.
To learn more, see Document Processing Flow.
Flow Blocks
New
Searchable PDF output — In v43.2 and later, the Document Renderer Block can add a searchable text layer to PDF output. You can search the PDF’s contents, including text that spans multiple lines. Downstream systems can also index the output more easily.
If text segments are unavailable, the Document Renderer Block runs Segmentation and OCR to generate them.
Search highlights may be approximate because their placement is based on OCR-derived word bounding boxes and estimated text positioning.
Updated
Email attachment processing — In v43.2 and later, the Email Listener and Submission Bootstrap Block process supported files found inside an attached .eml or .msg file. This supports one level of nested email attachments. Files inside emails nested more deeply are not processed automatically. Note that this is not part of the standard IDP processing flow. To process additional levels, create a custom flow using the Flows SDK:
Add a Custom Code Block before the Submission Bootstrap Block to extract the nested files. This configuration is not included in the standard Document Processing flow. See example in Flows SDK documentation.
Salesforce Listener Block no longer supported — Beginning with 43.2, the legacy Salesforce Listener Block is no longer supported. If you use this block, replace it with a customer-managed Salesforce integration that creates submissions through the Hyperscience Submission API using POST /api/v5/submissions. Learn more in [43.2 and later] Salesforce Listener Block Deprecation.
Hyperscience does not require a specific Salesforce integration method. Salesforce can call the Submission API directly, or your organization can send requests through middleware, an integration platform, or a scheduled process.
Machine-to-machine authentication with OAuth 2.0 is required for SaaS deployments and strongly recommended for on-prem deployments where available.
If your submission references a protected Salesforce file URL, configure OAuth authentication separately in the Submission Bootstrap Block or HTTP Downloader Block that retrieves the file. Authentication to the Submission API does not give Hyperscience access to protected files referenced in the request. For the complete submission contract, see the API documentation.
To learn more, see Document Processing Flow.
Flow Runs
Updated
Improved JSON viewer — In v43.2 and later, large JSON payloads are faster and easier to navigate when you inspect flow runs, jobs, and submission output. You can use the updated viewer in the following locations:
Flows > Flow Runs — Click a Flow Run ID, and then select a block to inspect its inputs and outputs.
Flows > Flow Runs — Click a Flow Run ID, and then open the flow-level input, output, or errors tab.
The updated viewer allows you to:
Use your browser’s search function to find values in the default text view.
Follow references between related block inputs and outputs.
Open or download referenced Hyperscience objects from their UUIDs.
Link directly to a selected tab or property and return to the same location after refreshing or sharing the URL.
Copy or download the JSON from either view.
.gif?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
Submissions — Open a submission (Submissions > Submission ID), and then click Actions > View JSON Output.
Administration > Jobs — Open a Job and click on a flow run to inspect its JSON details.
JSON opens in the original line-numbered text view. From there, you can switch to the tree view to expand or collapse individual objects and arrays instead of scrolling through the entire payload.
Only the visible parts of the tree are loaded, helping the page remain responsive when you inspect large payloads. To learn more, see Flow Runs Page and Submission Output Page.
Training Data Management (TDM)
New
Archive models — In v43.2 and later, you can archive Classification, Identification, and VLM Extraction models that are no longer actively used. Archived models are removed from the Model Details page, but are retained in a dedicated Archive view for future reference. Before archiving a model:
A live model must be undeployed before you can archive it.
A candidate model must be rejected before you can archive it.
A base model cannot be archived.
To archive a model:
In the Model History table, find the model you want to archive
Click the options drop-down (
.png?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
)Click Archive.
To view archived models, open the options menu in the Model History table and select View Archive. To return to the active models, select View Active.
You can also permanently delete an archived model, but note that this action cannot be undone.
Access model UUIDs from Model History — In v43.2 and later, you can view and copy a model’s UUID from the Model History table. The UUID is shortened in the table; hover over it to view the complete value or click it to copy the value to your clipboard.
The UUID column is hidden by default. To display it:
In the Model History table, open the menu next to Filter.
Click Manage columns.
Select the UUID, and then click Save.
This feature is available for Classification, Identification, and VLM Field Extraction models. Learn more in our Models category.
Updated
Long tag names on the Training Data table — v43.2 fixes an issue where long tag names could expand the Tags column and obscure other content in the table. Long names are shortened with an ellipsis. Hover over a shortened tag to view its full name.
VLM Field Extraction model names — v43.2 fixes an issue where renaming a VLM Field Extraction model with more than 128 characters returned a database error. Model names are validated before being saved, and a clear message indicates the 128-character limit. Long model names no longer expand the Name column or disrupt the Model History table. They are shortened with an ellipsis; hover over a shortened name to view it in full.
Training data export notifications — When you export training data as a CSV file, Hyperscience v43.2 confirms that the export is being prepared and directs you to the Notifications menu for the download. This confirmation is shown for Classification, Identification, and VLM Field Extraction models.
Exporting training data while documents load — In v43.2, CSV export options are unavailable when the export would include documents that are still loading. This prevents exported files from containing links to annotation pages that are not yet ready. You can export selected rows as soon as those documents have loaded.
To export all training data, wait for every document to finish loading.
Hover over an unavailable option to see why the export is not yet available.
This fix applies to Classification, Identification, and VLM Field Extraction models. To learn more, see the Models category.
Fixed
Importing duplicate models — v43.2 fixes an issue that allowed duplicate Classification and Identification models to be created when the same model was imported more than once.
Hyperscience now rejects an imported model when its UUID already exists. The error message identifies the UUID and indicates that the existing model may be in the archive.
Field Identification
Updated
Remove occurrences without changing pages in Field Identification Supervision and QA — You can now remove a field occurrence while keeping your position on the current page. This makes it easier to review and correct occurrences in long documents without being moved to the page containing the previous occurrence.
To remove an occurrence and keep the current page in focus:
Use OPTION + SHIFT + X on Mac
Use ALT + SHIFT + X on Windows.
You can also hold SHIFT and click the Delete icon.
To learn more, see Field Identification.
Flexible Extraction
New
Search by text segment — In v43.2, you can search for specific text in Flexible Extraction tasks. Matching text is highlighted in the document, making content easier to locate during annotation and review.
This feature is off by default. To enable it:
Go to the Flow Studio (Flows > Edit flows).
In Flow Studio, select Flexible Extraction from the Settings Type drop-down list.
Select Enable text segment search.
Populate fields from text segments — You can now populate fields in Flexible Extraction by selecting text segments directly in the document.
Click a segment to populate the selected field, or draw a box around multiple segments to combine their values.
The selected content and its location are associated with the field, reducing the need to enter values manually.
Only characters supported by the field’s data type are added to its transcription value. For example, if you select a segment containing “Dunlop2026” for a numeric field, the field is populated with “2026”. Selecting a checkbox or signature segment automatically updates the corresponding field. Learn more in Flexible Extraction.
Transcription
Fixed
Reliable Transcription Quality Assurance (QA) processing — Fixed an issue that could prevent Transcription QA responses from processing successfully. Transcription QA and downstream reporting are now complete as expected.
Cases
New
View original file names for Case documents — The Documents table on the Case Details page now shows the names of the files originally uploaded with the submission. This makes it easier to identify the source file associated with each Case document. To learn more, see Cases.
Export Knowledge Store collections — In v43.2, you can export an entire Knowledge Store collection as a CSV file. The exported file is compatible with the existing CSV importer, allowing you to edit collection items offline, re-import them, or move a collection between environments. To export a collection:
Open the collection.
Click Export.
When the export is ready, open the notification and click Download.
Learn more in Knowledge Store.
List Knowledge Store collections through the API — In v43.2, you can use the public v5 API to retrieve an alphabetical list of Knowledge Store collections and the number of items in each collection. This makes it easier to discover available collections before retrieving or managing their items through an integration. For endpoint and response details, see the API documentation.
Updated
Knowledge Store audit records — In v43.2, the audit records for collection imports, exports, and deletions are improved. Audit records distinguish queued exports from completed exports, record rejected export attempts, identify imported source files, and provide clearer information about deleted collections and items.
Layouts
Fixed
Preserving existing field notes after upgrading — After upgrading to v43.1, field notes, including VLM prompts, could appear blank or be removed when a layout was viewed, edited, copied, or added to a release. Table notes were not affected.
This update adds a safeguard that preserves existing field notes during these actions. It prevents data loss and keeps field instructions and VLM prompts consistent across layout versions and releases.
This mitigation does not automatically restore notes that were already removed. Notes that are still available in a working layout version can be recovered separately. Some cases may require additional recovery.
For help identifying affected layouts or recovering available notes, file a ticket on our Support portal.
Releases
Fixed
Fixed an issue that could prevent users from locking a release — Previously, layouts that shared only some pages could be incorrectly marked as duplicates, preventing the affected release from being locked. In v43.2, layouts are marked as duplicates only when all the pages in one layout are also present in another layout. This restores the behavior from v42.2 and earlier.
Reporting
Updated
Automatic transmission of usage data — v43.2 improves the reliability of scheduled usage-data transmissions. Hyperscience detects missing reporting days and automatically sends the missing data during a scheduled transmission. The Usage page shows the bundle-transmission status based on the most recent transmission, whether it was triggered automatically or manually:
Healthy — The latest scheduled transmission completed successfully.
Latest transmission needs attention — The latest transmission encountered an issue and is still being retried.
Latest transmission failed — The latest transmission could not be completed after all retries.
Disabled — Automated usage-data transmission is turned off.
Select Show… to view the transmission history. The table includes an Updated column showing when each transmission was last updated.
You can also view the individual transmissions and filter those with a Warning status.
Application bundle
Updated
Reduced full application bundle size — In v43.2, the full on-prem application Bundle is approximately 29% smaller. In testing, its compressed size decreased from 46.95 GB to 33.44 GB, reducing the bandwidth and storage required to download the bundle.
During installation, Hyperscience reconstructs the original block packages before importing them. This process preserves the packages byte-for-byte but adds approximately 2 minutes to the block-import step and requires temporary disk space.
By default, temporary files are written to
$HS_PATH/media.To use another location, configure
HS_SDM_REHYDRATE_DIR.
This update applies only to the Docker/ Podman on-prem application bundle. SaaS and K8S deployments are not affected.
Installations
Updated
New artifact distribution service — Hyperscience v43.2 distributes application bundles, container images, and Helm charts through a new artifact service called Harbor, replacing Cloudsmith. You will continue to receive authenticated download links or curl commands, and no changes to your download process are required.
File storage
Updated
Faster image loading from object storage — If your deployment uses supported cloud object storage, in v43.2, you can deliver original document images directly to the browser through short-lived, pre-signed URLs. This can improve loading times in image-heavy views, including Supervision, while reducing load on the application server.
This behavior is off by default and only available to Docker/Podman and Kubernetes deployments:
For a Docker or Podman-based deployment, add
REDIRECT_CLIENT_FACING_FILE_SERVING_TO_OBJECT_STORE=trueto the application’s .env File.For a Kubernetes deployment, add the variable under
app.dotenvin the Helm chart’svalues.yaml.
Before enabling the setting, ensure that browsers in your instance can access the object store and that its access policy allows requests from your Hyperscience application. Images that require resizing, cropping, or rotation continue to be served through the application.
Kubernetes
Fixed
Refreshing the System & Health page — v43.2 fixed an issue that caused an authorization error to occur when users clicked the Refresh button on the System & Health page in on-premise Kubernetes deployments.
Databases
Updated
Support for PostgreSQL 18.4 — Hyperscience v43.2 adds support for PostgreSQL 18.4. From v44.0+, PostgreSQL 15.x will be deprecated.
API
Fixed
Filtering halted submissions — Fixed an issue where GET /api/v5/submissions?halted=true could include legacy halted submissions whose response incorrectly reported "halted":false. Submissions returned by this filter now consistently report "halted": true. To learn more, see our API Documentation.
43.2.1 (29 Aug 2026)
Reporting
New
VLM GPU image counts usage reports — Usage reports now include a # of images metric alongside the existing request count. This metric shows the number of pages processed during Field Extraction or the number of images submitted through general prompting, providing greater visibility into ORCA usage for each task.
Training Data Management (TDM)
Fixed
Bulk actions could affect unselected training data — When filters were active, a bulk action on selected training data could also affect other results that matched those filters. Bulk delete, status change, and update actions now affect only the selected results. This prevents unintended changes. Select-all actions continue to affect all results that match the active filters.
43.1.7 (10 Sept 2026)
Layouts
Fixed
Preserving existing field notes after upgrading — After upgrading to v43.1 or later, field notes, including VLM prompts, could appear blank or be removed when a layout was viewed, edited, copied, or added to a release. Table notes were not affected.
This update adds a safeguard that preserves existing field notes during these actions. It prevents data loss and keeps field instructions and VLM prompts consistent across layout versions and releases.
This mitigation does not automatically restore notes that were already removed. Notes that are still available in a working layout version can be recovered separately. Some cases may require additional recovery.
For help identifying affected layouts or recovering available notes, file a ticket on our Support portal.
43.1.6 (29 Aug 2026)
Reporting
New
VLM GPU image counts usage reports — Usage reports now include a # of images metric alongside the existing request count. This metric shows the number of pages processed during Field Extraction or the number of images submitted through general prompting, providing greater visibility into ORCA usage for each task.
Layouts
Fixed
Preserving existing field notes after upgrading — After upgrading to v43.1 or later, field notes, including VLM prompts, could appear blank or be removed when a layout was viewed, edited, copied, or added to a release. Table notes were not affected.
This update adds a safeguard that preserves existing field notes during these actions. It prevents data loss and keeps field instructions and VLM prompts consistent across layout versions and releases.
This mitigation does not automatically restore notes that were already removed. Notes that are still available in a working layout version can be recovered separately. Some cases may require additional recovery.
For help identifying affected layouts or recovering available notes, file a ticket on our Support portal.
43.1.5 (18 Aug 2026)
Training Data Management
Fixed
Bulk actions in TDM for Classification — We fixed an issue where bulk status changes and deletions affected all filtered documents, even when you selected only specific documents. Bulk actions now apply only to the documents you select.
43.1.4 (27 Jul 2026)
Cases
New
Filter case documents by status — Users can now filter documents within a case by their status. This enhancement allows for easier identification and triage of specific files within the Documents table.
Authentication
Fixed
SAML-enabled deployments no longer require a manual administrator-role configuration during upgrades — Previously, database migrations could fail when SAML_ADMIN_PERMISSION_ROLE used its default value, preventing the upgrade from completing. The setting now has a valid default value hs_amdin, so upgrades can proceed without this workaround.
43.1.3 (22 Jul 2026)
Versions 43.1.0 — 43.1.2 were not released and are not supported.
Submissions
New
Flow Run Status for submissions — We've added a Flow Run Status column to the Submissions table on the Submissions page (Submissions > Submissions).
This status aligns with that of the main flow run used to process the submission, categorizing it into one of four mutually exclusive and collectively exhaustive statuses:
Running
Failed
Completed
Canceled.
It can also be viewed on the Submission Output page and on the Document Output pages for the submission's documents. While it is being processed, each submission passes through stages, which appear in the Stages column (previously Status) of the Submissions table (e.g., Manual Classification, Manual Identification).
In earlier versions, the Stages column appears as Status
For more information, see [v43.1] Navigating the Submissions Page and [v43.1] Submission Output Page.
Filtering submissions by Flow Run status — A new Flow Run Status filter allows you to identify submissions whose Top-Level Flow run has failed, making it easier to isolate and troubleshoot failed processing. As part of this change, the “Halted” submission stage (previously status) has been superseded by the “Failed” Flow Run Status. If you previously filtered for halted submissions by using the submission state, use Flow Run Status = Failed instead.
Clicking the ID of a failed submission opens the Submission Output page, where you can select View Failed Flow Runs to investigate the failure.
Existing submission states continue to represent the current processing stage of the submission. To identify submissions that are actively being processed, combine the submission stage with Flow Run Status = Running.
To learn more, see [v43.1] Document Output Page.
Cancelling submissions — If you need to cancel a submission, you can now do so from its Submission Output page or the Submissions page. When you cancel a submission, any flow runs used in its processing are also canceled, along with any Supervision or Quality Assurance (QA) tasks based on the submission's content. You can also provide a cancellation reason of up to 512 characters. More information on canceling submissions can be found in [v43.1] Submission Output Page.
Classification
New
Group uncategorized pages by layout or filename — We've added new grouping options to the Uncategorized Pages panel in the Document Classification Task. You can now view uncategorized pages grouped by:
Layout to view pages grouped by their machine-predicted layout.
Filename to view pages grouped by their original submission filename.
These grouping options make it easier to review, compare, and process related pages during manual classification, reducing Average Handling Time (AHT). To learn more, see [43.1] Document Classification Task.
.png?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
Table Identification
New
Added keyboard shortcuts for switching between tables — We've added new keyboard shortcuts to the Table Identification annotation experience.
Use Shift + W to move to the next table and Shift + E to move back to the previous table.
These shortcuts are also available in the Keyboard Shortcuts panel.
.png?sv=2026-02-06&spr=https&st=2026-09-19T21%3A55%3A31Z&se=2026-09-19T22%3A50%3A31Z&sr=c&sp=r&sig=idJK%2FdHIyHxX6Ykgc7EcqlC0DhrE4gOTJ12JLzafang%3D)
Layout Editor
Updated
Shared notes across variations in Structured layouts — When a field is shared across multiple variations in a Structured Layout, changes made to its Notes value are automatically propagated to all matching fields. This update ensures that notes remain consistent across shared fields in layout variations. Learn more in Creating Structured Layouts, Layout Editor Tools & Functionality, and Editing a Structured Layout Variation.
Models
Updated
Updated infrastructure requirements for ORCA VLMs — ORCA VLMs in v43.1 introduce new infrastructure requirements. Machines running ORCA must have the NVIDIA driver 580 or later, and at least 10 GB of available space in the operating system’s /dev/shm directory. Verify that your ORCA infrastructure meets these requirements before upgrading. For more information on ORCA infrastructure requirements, see Infrastructure Requirements.
Training Data Management (TDM)
Updated
Model-action notification messages for VLM Field Extraction — We've updated the notification messages for common model actions in VLM Field Extraction, such as deploying and undeploying models, rejecting candidate models, scheduling and canceling model training, and exporting models.
The following notification messages are displayed when these actions are completed successfully:
Model successfully undeployed.
Candidate model successfully deployed.
Model successfully exported.
Candidate model successfully rejected.
Training successfully scheduled.
Model training successfully canceled.
To learn more, see VLM Field Extraction Models.
Preventing duplicate tags caused by capitalization differences — Tags with the same name but different capitalization are now treated as the same entry. For example, “Vendor”, “vendor”, and “VENDOR” are all mapped to the same value (e.g., “VENDOR”). This update helps prevent duplicate tags and maintain consistency across training data tagging. Learn more in TDM for Classification Models, TDM for Identification Models, and TDM for ORCA VLMs.
Improved training data table CSV export — Training data table CSV exports now include status names instead of numeric status codes and provide layout names alongside layout version identifiers. These changes make exported data easier to review, filter, and analyze. We've also updated exports to include only the most relevant columns for each model type, resulting in cleaner and more focused CSV files. See TDM for Classification Models, TDM for Identification Models, and TDM for ORCA VLMs for more information.
Reporting
Updated
Transmissions of the Usage Bundle to Hyperscience — We've made the following updates to the transmissions of Usage Bundles:
Usage Bundles are sent to Hyperscience through flows, allowing for automatic retries and improving the overall reliability of the transmission mechanism.
You can now manually initiate the transmission of the Usage Bundle by clicking Send on the Usage page (Reporting > Usage).
We've removed the eight-day redundancy of data in the Usage Bundle, helping to reduce file sizes and processing overhead.
The new Usage Bundle Transmissions table on the Usage page shows information on automatic and manually initiated transmissions of the Usage Bundle. This includes status, creation, and completion dates of each transmission. It also contains links to view the transmission's flow run and to download the transmitted file. More information about the Usage Bundle can be found in [v43.1] Usage Bundle.
Updated tracking of created pages and submissions — Metrics for created pages and submissions are now updated after successful progression through the Submission Bootstrap Block. In previous versions, these metrics were updated when submissions progressed through the Complete Block, preventing visibility into the volume of pages and submissions still being processed. As a result of this change, data in the Usage Bundle's application_usage CSV more accurately reflects processing volumes, regardless of whether submissions are in the Completed state.
Layout-level Classification accuracy metrics in the Usage Bundle — We’ve added layout-level metrics to the classification_accuracy CSV file in the Usage Bundle’s product_analytics folder. These metrics indicate the level of agreement between keyers and the machine in the classification of documents on a per-layout basis. Layout-level metrics are available only for documents processed after upgrading to v43.1.
Submission Processing
Fixed
Rendering of Checkboxes in PDF documents — Fixed an issue where certain checkboxes in processed PDF documents could be displayed as radio buttons after PDF pagination. Checkboxes are now rendered accurately, preserving the original appearance of PDF forms.
API
New
Cancel Submissions endpoint — You can now cancel one or more submissions through the API by using the Cancel Submissions endpoint. Canceling a submission stops any pending processing and updates its flow_run_status to cancelled. The submission’s current stage is preserved so you can see where processing was canceled. Any associated manual tasks are also canceled.
Because cancellation is asynchronous, the endpoint returns a 202 Accepted response. You can track the operation by using the returned status URL.
43.0.9 (14 Aug 2026)
Flow Blocks
Updated
Pagination can now be faster for native PDF files — Set PAGINATION_LOW_MEMORY_MODE to false to improve pagination speed.
Low-memory mode remains enabled by default.
Disabling it may increase memory usage.
Releases
Updated
Importing large layout releases — The application now uses significantly less memory when importing releases. This allows larger releases to be imported without running out of memory.
Fixed
Processing could fail after locking a release — The application could reuse layout information that was created earlier. Sometimes, that information contained different fields from the latest layout version. This could create an invalid release and cause processing to fail. The application now reuses layout information only when the fields match.
Training Data Management
Fixed
Bulk actions in TDM for Classification — We fixed an issue where bulk status changes and deletions affected all filtered documents, even when you selected only specific documents. Bulk actions now apply only to the documents you select.
43.0.8 (24 July 2026)
Cases
New
Filter case documents by status — Users can now filter documents within a case by their status. This enhancement allows for easier identification and triage of specific files within the Documents table.
Releases
Fixed
Fixed an issue that could prevent users from locking a release — Previously, layouts that shared only some pages could be incorrectly marked as duplicates, preventing the affected release from being locked. Layouts are now marked as duplicates only when all the pages in one layout are also present in another layout. This restores the behavior from v42.2 and earlier.
Authentication
Fixed
SAML-enabled deployments no longer require a manual administrator-role configuration during upgrades — Previously, database migrations could fail when SAML_ADMIN_PERMISSION_ROLE used its default value, preventing the upgrade from completing. The setting now has a valid default value hs_amdin, so upgrades can proceed without this workaround.
43.0.7 (14 Jul 2026)
Authentication
Fixed
Supporting secondary SAML authentication — Users were unable to log in via secondary SAML identity providers because the system was clearing required security information too early during the login process. This resulted in a system error that prevented access for users with specific security configurations. The fix ensures that all necessary security data remains available until the login is completed.
Submissions
Updated
Improving submission processing performance — We optimized the data retrieval process for submission pages to prevent database memory exhaustion and system-wide crashes during large-scale processing. Previously, certain requests triggered exhaustive database scans that could halt processing for all organizations on a shared server. These improvements ensure the system uses targeted lookups, maintaining stability and consistent performance regardless of the total volume of data in the system.
43.0.6 (2 Jul 2026)
Documents
Updated
Viewing source-file names for documents — You can now view the source-file names for documents in the File Names column of the Documents table on the Documents page (Submissions > Documents).
Cases
Updated
Viewing source-file names for documents — You can now view the source-file names for a case's documents in the File Names column in the Documents table on the Case Details page.
43.0.5 (19 Jun 2026)
Tasks
Updated
Queuing tasks for specific submissions, documents, and cases — When a keyer clicks the Perform Tasks button for a specific submission, document, or case, the system queues up to 10 tasks for the keyer rather than retrieving a single task at a time. As a result, keyers can complete tasks more quickly.
Kubernetes
Fixed
Refreshing the System & Health page — We've fixed an issue that caused an authorization error to occur when users clicked the Refresh button on the System & Health page in on-premise Kubernetes deployments.
43.0.4 (12 Jun 2026)
Flow Blocks
Fixed
Testing connections to S3 and GCS file stores — We’ve fixed an issue that caused testing of S3 and GCS (Google Cloud Storage) file-store connections to fail when a user tested them by clicking Test File Store Retrieval in the Submission Initialization Block’s settings.
“Clear value” button for secret-related settings in the Submission Bootstrap Block — Users can click Clear value for secret-related settings in the Submission Bootstrap Block to delete their contents in a single click. This update allows users to resolve authentication errors more quickly.
Training Data Management
Fixed
Scope of bulk deletion of training data for Classification models — We've resolved an issue that caused all training data for a Classification model to be deleted when a user requested the deletion of training data for specific layouts only.
43.0.3 (5 Jun 2026)
Flow Blocks
Updated
OAuth configuration for HTTP/HTTPS connections in the Submission Bootstrap Block — To support OAuth authentication for HTTP/HTTPS connections, we've added the following settings to the Submission Bootstrap Block:
HTTP/HTTPS OAuth Token URL
HTTP/HTTPS OAuth Client ID
HTTP/HTTPS OAuth Client Secret
HTTP/HTTPS OAuth Scope
The Generic Web Storage (HTTP/HTTPS) Configuration setting is still available.
For more information on these settings, see Flow Blocks, Document Processing Subflow Settings, and Document Processing with ORCA Subflow Settings.
Releases
Fixed
Locking releases with "duplicate layouts" — We've fixed an issue that prevented releases from being locked when they contained layouts where the Blanking tool was used. In these situations, an error message falsely indicated that the releases contained duplicate layouts.
Training Data Management
Updated
Improved training data table CSV exports — Training data table CSV exports now include status names instead of numeric status codes and provide layout names alongside layout version identifiers. These changes make exported data easier to review, filter, and analyze. We've also updated exports to include only the most relevant columns for each model type, resulting in cleaner and more focused CSV files.
43.0.2 (21 May 2026)
Custom Supervision
Fixed
Required document metadata fields — Document metadata fields marked as required in Custom Supervision tasks can no longer be left empty when completing a task. This update improves validation for tasks that rely on required metadata input and helps prevent incomplete submissions during review.
Reporting
Updated
New CSV files in the Usage Bundle — We've added the following CSV files to the Usage Bundle:
entry_accuracy— Includes data on the daily accuracy of Identification and Transcription tasks based on QA results. This data is grouped by flow and layout and differentiates between tasks performed by the machine and those performed by keyers.classification_qa_layout_revisions— Provides the names of the Semi-structured layouts predicted by the system and the layouts chosen by keyers during Classification QA. Its data is grouped by flow and layout.
Infrastructure
Fixed
Podman 5.6.0 compatibility during container startup — We’ve fixed an issue where deployments using Podman 5.6.0 could fail during container initialization due to changes in the Podman API behavior. This update improves compatibility with Podman 5.6.0 environments and helps prevent startup failures during deployment.
43.0.1 (9 May 2026)
Notifications
Updated
Hiding previously viewed notifications — To increase the visibility of unread notifications, we've divided the notifications in the Notifications panel into Unread and Read sections. The contents of the Read section are hidden by default but can be revealed by clicking on the Read section heading.
Layouts
New
"Notes" setting for fields in a Structured layout — We've added a Notes setting in the Layout Editor for fields in Structured layouts. These field-specific settings are optional and can be used to record information on the field's contents for your team's reference and for use in custom flows.
Training Data Management
Updated
Contents of "Actions" drop-down lists for Classification, Field Identification, and VLM Field Extraction training data — We've updated the training data Actions drop-down list for Classification, Field Identification, and VLM Field Extraction models for consistency across the three model types.
Fixed
Scope of bulk deletion of training data — We've fixed an issue that caused all training documents for a model type (e.g., VLM Field Extraction) to be deleted when a user requested the deletion of all training documents for single model.
Cases
Updated
Data in "Stats" card on the Case Details page — We've updated the Stats card on the Case Details page to include the following data:
Total documents
Completed documents
Total submissions
Completed submissions
Halted submissions
These statistics provide more information on the status of a case's submissions, making it easier to understand the status of the case as a whole.
Flow-based permissions and access to cases — Access to cases is now determined by the flow-based permissions for each case's submissions.
As part of this update, we've made following changes to the Cases user experience:
Cases page
If a user does not have access to any of a case's submissions, the name of the case will be grayed out on the Cases page, and they will not be able to click on the name of that case.
Case Details page
Users are not able to interact with documents, unmatched pages, and data types that flow-based permissions restrict their access to.
Deleting a case or editing its deletion date requires access to all pages in the case.
Updating Notes for a case requires access to at least one page in the case.
Cases API endpoints
Documents and unmatched pages are redacted in responses.
The contents of
submission_filesandsubmissionselements are filtered based on the permissions for the submission's flows.A
403response is returned if the user has no access to any of the case's pages.Responses from the Listing Cases endpoint omit information cases that are inaccessible to the user.
Fixed
"Show layout information" icon for users without the View Layouts permission — We've fixed an issue where the Show layout information icon on the Case Details page was shown to users who did not have the View Layouts permission. When they clicked on the icon, an authorization error occurred.
Supervision
Updated
Keyboard shortcuts for adding and removing table rows in Custom Supervision and Flexible Extraction — Keyers can now remove a selected row during Custom Supervision and Flexible Extraction tasks by pressing Alt + R (Windows) or Option + R (Mac) on their keyboards. They can add a row below the selected row by pressing Alt + N (Windows) or Option + N (Mac).
Custom Supervision
Updated
Viewing a bounding box for a single field — We've added a Toggle focused bounding boxes button to Custom Supervision tasks. When enabled, only the bounding box for the field selected in the right-hand sidebar is shown on the document. This option improves the readability of documents where the bounding boxes for multiple fields overlap.
Reporting
Updated
Flow-level Classification accuracy data in the Usage Bundle — We've added a classification_accuracy CSV file to the Usage Bundle's product_analytics folder. This file indicates the number of pages whose classifications were checked during Classification QA during the reporting period and how many of those classifications were correct. The data is provided at the flow level.
Hourly page metrics CSV file in Usage Bundle — The Usage Bundle includes a new CSV file, hourly_pages_metrics, in the product_analytics folder. This file enables review of hourly page volumes.
Hourly reporting data is available for the past 30 days only. If a Usage Bundle is created for a period that includes dates that are more than 30 days in the past, the CSV file will contain data only for the days within the available reporting window.
Field Identification
Fixed
Duplicate responses to Field Identification tasks — We've fixed an issue that created duplicate responses to Field Identification tasks in some situations, causing those tasks' flows to fail.
43.0.0 (6 May 2026)
Highlights
A leader in intelligent document processing, Hyperscience strives to consistently add value, drive innovation, and improve the user experience with every new version of Hypercell. As such, we’re introducing the following key features in Hyperscience v43.
Flow Studio
Flow Builder — With the Flow Builder in Flow Studio, you can add and delete blocks from a flow directly in the application, eliminating the need to edit the flow's code to make these updates. If your flow has Code Blocks, you can update the code for those blocks in the application, as well. As a result, the Flow Builder allows you to experiment with new flow blocks and iterate quickly on your changes.
You can access the Flow Builder by toggling the Build Mode switch in the upper-left corner of the flow's canvas. When this switch is in the ✔ position, Add block (+) buttons appear in the flow where you can add blocks.
Clicking on an Add block button reveals options to add flow controls, like Routing or Output Blocks, or blocks that are included in the Document Processing subflow. You can also click Get More Blocks to download additional sets blocks from the Flows SDK documentation.
When you click on a block, you can edit the block's settings. For some settings, you can specify the output of any preceding block as an input.
If you need to delete a block, you can click on the Delete button that appears above it.
Note that this feature does not replace the Flows SDK; you can still use the Flows SDK to create and edit flows.
For more information about the Flow Builder, see Editing Flows in Flow Studio.
Upgrade notes
These updates may impact your upgrade process or affect initial processing times after upgrading. For more information or assistance, contact your Hyperscience representative.
Python versions
Supported Python versions — Beginning in v43, the use of Python 3.11 in flows, including Code Blocks and external Python packages, is no longer supported. For more information on supported Python versions, see Developing Flows.
Submission retrieval storage
Deprecation of OCS (Open Computer and Software Inventory) submission retrieval storage — We are deprecating support for OCS submission retrieval storage in v43. If you are using an OCS submission retrieval store, you need to migrate to another storage option before support for OCS ends in v46.
For information on supported options for submission retrieval storage, see Submission Retrieval Storage Overview.
Additional features and enhancements
User experience
Login banner in FedRAMP deployments — We've added a banner to the Hypercell login page in FedRAMP deployments, which contains messaging required for government information systems. This banner appears above the Username and Password fields on the login page.
Models
Open list view filtered to a single task — You can now open the model list view pre-filtered to a single task (e.g., Field extraction, Field location) using URL parameters. This feature allows you to share links that take users directly to specific task content. The page and tab behavior have been updated to ensure the view renders correctly when scoped to a single task.
Training Data Management
View group information in the annotations page — You can now view group information directly in the annotations page for Identification models. Groups are assigned during Training Data Analysis, which organizes similar documents based on patterns in text and layout. This update makes it easier to review grouping behavior and validate your training data without interrupting the annotation process.
To learn more, see Training Data Analysis.
Managing document tags from the annotations page — You can now add, edit, and remove document tags directly in the annotations page for Classification, Identification, and VLM Extraction models. This update allows you to manage tags while reviewing documents, without switching back to the Training Data table.
For more information, see TDM for Classification Models, TDM for Identification Models, and TDM for ORCA VLMs.
Exporting training data table to CSV — You can now export training data table for Classification, Identification, and VLM Extraction models as a CSV. Doing so allows you to bulk download your training data table and work with the dataset outside the system. You can export the entire training data table or a selected subset of rows. The exported data includes all table columns, along with a link to each document’s ground truth item.
More information can be found in TDM for Classification Models, TDM for ORCA VLMs, and TDM for Identification Models.
VLM field extraction annotation experience — We’ve improved how bounding-box adjustments update transcriptions in the VLM field extraction annotation experience. Each transcription now reflects all text included in its bounding box.
To learn more, see TDM for ORCA VLMs.
VLM field extraction renaming — Model names can now be updated directly in the VLM field extraction Model History table. This feature allows you to rename VLMs without creating new entries, improving clarity and consistency when managing model versions. Changes apply only to the model name and do not affect the underlying model configuration or history.
For more information, see TDM for ORCA VLMs.
Flows
Configurations for low-latency processing — We've introduced a set of ".env" file variables that can be used to reduce latency in the processing of submissions. Note that the values of these variables should be changed only on the recommendation of your Hyperscience representative.
To learn more about the available ".env" file variables for low-latency processing, see Reducing Submission Latency.
Opening Flow Studio with the “View Flows” permission — We’ve fixed an issue that prevented users with the View Flows permission from opening Flow Studio unless they also had permission to access System & Health settings or the API. Users can now open Flow Studio with the View Flows permission alone.
Flow Blocks
Document Renderer Block includes download URLs in submission output — The Document Renderer Block now includes rendered document metadata directly in each document within the submission output. The metadata includes a download URL for each rendered document. This update eliminates the need to customize flows to generate the download URLs.
file_uuid removed from Submission Bootstrap Block output — We’ve removed the file_uuid field from page-level output in the Submission Bootstrap Block. This field was previously interpreted as a file identifier, which led to incorrect usage in some scenarios. Access to the original file remains available through the origin object.
Full Page Transcription raw_segments output removal — We’ve removed support for returning raw_segments in Full Page Transcription Blocks. This update reduces output size and improves performance, especially for text-dense documents. Raw segments are still available and can be accessed separately using the provided segment identifier.
Translation block — We’ve added a Translation block that can translate text from one language to another in any custom flow. The block supports translation across 400 languages.
Improved custom entity detection across lines and pages — The Entity Recognition Block now supports detecting entities that span multiple lines or pages. This update improves detection accuracy for real-world documents where values are often split across lines or continued on the next page.
For more details, see Entity Recognition Block.
Improvements to the user experience for resolving secret-related errors — To make it easier for users to address secret-related errors in the Submission Bootstrap Block, we've made the following updates:
Error messages provide more clarity on how to resolve their respective errors and which settings they apply to.
If no secret value has been entered for a setting, no placeholder value appears in the text box for that setting.
Flows SDK
Block-specific settings for IDP Blocks — We’ve separated the settings for IDP Blocks in the Flows SDK by block. When building a custom flow, you can now include settings only for the blocks used in that flow instead of including all IDP Core settings. Existing custom flows remain compatible with this update. Learn more in our Flows SDK documentation.
Submissions
Support for BMP and GIF files — We've added support for the processing of BMP and GIF files in Hypercell.
More information about supported file types can be found in the What is a Submission? article.
Classification
Visual indication of expected document pages — When creating new documents manually, the page editor now shows page boxes that indicate how many pages are expected. These boxes guide you when working with layout variations that contain multiple pages. They are visual aids and do not restrict adding additional pages.
To learn more, see Structured Document Classification.
Flexible Extraction
Field location overlay on documents — During Flexible Extraction for Structured documents, field locations are now displayed directly on the document to guide data entry.
For registered pages, field locations are shown precisely.
For unregistered pages, approximate locations are indicated.
Selecting a field in the right-hand sidebar highlights its corresponding location on the document.
For more information, see Flexible Extraction.
Custom Supervision
Validating Submission objects — We've added a submission-validation step to the Custom Supervision Block, which checks the contents of the Submission object and halts the submission if the fields required for Custom Supervision tasks are missing or invalid. If the submission is halted as a result of this validation, a detailed error message is provided that describes why the validation failed.
To learn more about Custom Supervision, see the Flows SDK documentation.
Adding and removing table rows — Keyers can now add and remove table rows during Custom Supervision tasks. By right-clicking on a table or using keyboard shortcuts, keyers can delete a specific row or all rows in the table. They can also insert a row above or below a selected row. This update makes the Supervision user experience more consistent across types of Supervision tasks, helping to increase keyer efficiency.
For more information about Custom Supervision, see our Flows SDK documentation.
Reporting
Submission and task metrics tracking in the Usage Bundle — The system now tracks submission and task metrics at the start of their lifecycle rather than waiting for completion, ensuring reports are accurate in real-time. This update prevents delays in metric visibility for uploaded pages and tasks without halting document processing.