Instabase Release Notes

Commercial

Enterprise

• You can export and import automation projects as ZIP files to move work across workspaces, organizations, and environments, or to create backups. Exports include the project details, schema, validation rules, and custom functions, including shared functions from the function library. Sample files or previously uploaded files aren’t included.
• In agent mode, the new Advanced checkbox detection file processing setting improves checkbox detection accuracy and can also improve results for table fields. Review the technical details before enabling in your project.
• Performance is improved when correcting tables in human review.
• The review dashboard has a new Date added column, letting reviewers sort and filter review tasks by date.
• You can create ground truth datasets from app runs that contained processing errors. The files that resulted in errors aren’t included in the dataset.
• Resolved issue | The number of fields per class was limited to 100, rather than the stated limit of 250. Adding 250 fields per class is now fully supported.
• Resolved issue | Hidden fields sometimes didn’t display correctly in human review.
• Resolved issue | Cleaning functions sometimes failed to rerun after reclassifying a document.
• Resolved issue | In human review, document sequence numbering was reversed. Reviewers now see the expected document 1 of 3 format, rather than document 3 of 1.

Enterprise

• The human review interface now fully supports packet-based review. Improvements to the review experience include:
  • Packet-processing app results can now exclusively be reviewed by packet. Previously, reviewing by file or by run was supported, which impacted visibility of cross-class fields. This change is also reflected in the deployment configuration tab: for packet-processing apps, configuring Review by file or Review by run isn’t supported.
  • The review dashboard makes it clearer that you’re reviewing by packet, and no longer requires a separate entrypoint for packet-based review.
  • When reviewing ranked cross-class fields, reviewers can select from a dropdown of extracted value options, or enter a custom value.
• The dropdown options show extracted values, not any manually corrected values. If reviewers have previously corrected the value of a class field, the dropdown options don’t reflect this change.

Original source Report a problem

Automate, deploy, & review

Commercial

• Resolved issue | When copying projects across workspaces, custom functions weren’t copied correctly.
• In agent mode, the new Force image OCR file processing setting lets you treat all pages as images and process them with OCR. When enabled, text extraction from machine-readable PDFs is skipped, which can improve processing speed and reduce memory usage for certain documents.
• This release includes memory optimization improvements to reduce peak memory usage during file processing. Expect improved stability and reliability when processing large files in agent mode.

Enterprise

• In cross-class custom function fields, you can now set the return type to Text choices, limiting valid returned values to a specified set. If the function returns a value not in that list, a validation error is shown. Additionally, in human review, reviewers can search and select from the allowed list rather than typing values themselves.

Original source Report a problem

This release includes security enhancements that disable local caching of potentially sensitive data.

Automate, deploy, & review

Commercial

• As previously announced, deployment integration functions now return results in the v2 format, which matches the structure returned by the Get run results API endpoint. To support this change, a migration helper is now available when editing existing integration functions. Use the helper to migrate your code and test changes before saving and permanently updating to the v2 format.

• Resolved issue | Users with tester permissions couldn’t view deployment configurations.

• The pop-out view for tables now displays wide, single-row tables more clearly.

• Single tenant | Resolved issue | When copying projects across workspaces, custom functions weren’t copied correctly.

Administration

Commercial

• When creating secrets, you can now use files as the secret’s value, such as uploading certificate files.

Enterprise

• When creating secrets, you can now use files as the secret’s value, such as uploading certificate files.

Original source Report a problem

Automate, deploy, & review

Community

• For extraction fields, handling of unclear or unavailable values is improved. Previously, N/A was returned, which could conflict with valid extracted values. Now, No value found is returned by default. This change is available immediately in automation projects and existing community apps, and in commercial and enterprise apps as AI runtime 2.0.1.
• Resolved issue | Renaming projects sometimes failed.

Commercial

• File splitting now supports files up to 400 MB and 2,000 pages when the Disable OCR setting is enabled. After splitting, individual documents should be 200 pages or fewer for optimal extraction results. For details about requirements and constraints, see file limitations.
• Single tenant | You can call LLMs from custom functions in validation functions, cross-class validation functions, and cross-class custom function fields. The LLM client uses the environment’s configured LLM provider and the configured AI runtime model. Supported in agent mode only.
• Resolved issue | When attempting to convert generated validation code to a custom function, the confirmation modal would close before users could confirm the action.
• Resolved issue | Users with tester permissions couldn’t view a project’s settings.
• Resolved issue | Apps with associated ground truth datasets failed to delete, despite showing a confirmation message. Now when you delete an app, the app and all associated ground truth datasets are deleted properly.
• Resolved issue | In projects, custom validation functions showed as having failed when the validation result exceeded the character limit. Improved error messaging now clearly indicates this root cause. This limitation doesn’t affect apps or deployments.

Enterprise

• Run results sent via downstream integrations or retrieved through manual export now include packet-level information, including cross-class fields. Previously, packet-level information was available only when getting results via API.

Administration

• AI Hub now enforces stricter access controls for cross-domain integrations. By default, requests from external websites are blocked unless explicitly verified as trusted, preventing unauthorized websites from performing actions on behalf of logged-in users.

Action required: If your organization integrates AI Hub with other applications (for example, embedding AI Hub in a portal or accessing it from a different domain), you must connect with Instabase Support to provide a list of URLs of your trusted applications. Support can configure the appropriate settings to maintain connectivity. Otherwise, AI Hub now allows same-origin requests only.

Original source Report a problem

Automate, deploy, & review

Community

Commercial

Enterprise

• The new Disable OCR setting lets you skip optical character recognition during file processing when using agent mode, significantly improving processing speed and reducing memory usage. When OCR is disabled, functionality that requires document text, such as text-dependent custom functions, doesn’t work.
• Deployment integration functions now return results in the v2 format, which matches the structure returned by the Get run results API endpoint. The v2 format organizes results by files and documents, providing a consistent structure and additional details compared to v1. Existing integration functions continue to work with the previous format for backward compatibility, unless edited. This is a breaking change. In 26.04, editing any integration function triggers conversion to the v2 results format, even if changes aren’t saved. In 26.08, a migration tool helps you update your code and test changes, and the conversion to v2 results format is completed only upon saving. To avoid unintended changes, wait until release 26.08 to edit any functions involved in business-critical workflows.
• Single tenant | For extraction fields, handling of unclear or unavailable values is improved. Previously, N/A was returned, which could conflict with valid extracted values. Now, No value found is returned by default. This change is available in commercial and enterprise apps as AI runtime 2.0.1.
• Single tenant | File splitting now supports files up to 400 MB and 2,000 pages when the Disable OCR setting is enabled. After splitting, individual documents should be 200 pages or fewer for optimal extraction results. For details about requirements and constraints, see file limitations.
• Single tenant | You can call LLMs from custom functions in custom function fields, classification functions, and splitting functions. The LLM client uses the environment’s configured LLM provider and the configured AI runtime model. Supported in agent mode only.
• Resolved issue | List-of-objects processing could error and return 0 due to a formatting bug. This issue affected derived fields and cleaning functions applied to list extraction results.

Enterprise

• Resolved issue | Users would encounter an error when editing ground truth datasets in human review. The issue occurred after scrolling to the 60th+ file in sets of 100+ files, then selecting a table or list field.

Administration

• Single-tenant | AI Hub now enforces stricter access controls for cross-domain integrations. By default, requests from external websites are blocked unless explicitly verified as trusted, preventing unauthorized websites from performing actions on behalf of logged-in users.
• Workspaces now have a default upload destination, the Data Drive, part of your organization’s 1 TB included storage. Each workspace has a separate Data Drive, ensuring data isolation between workspaces.

Conversations & chatbots

Community

Commercial

Enterprise

• Removal | The previously announced removal of conversations and chatbots is complete.

Administration

• Multi-tenant environments can now support IP allowlisting for outbound integrations, such as webhooks and deployment integration functions. Contact support for the egress IP ranges to allowlist.

API & SDK

• When using the generate_post_process_pdf parameter, PDF generation is now supported only for documents less than 100 pages in length.

Enterprise

• Resolved issue | input_fields values weren’t returned as expected with the Get run results endpoint.

26.02

26.02

Automate, deploy, & review

Community

Commercial

Enterprise

• Automation projects now support up to 250 fields per class. Optimizations in this release also improve processing performance for projects containing classes with large numbers of fields.

Commercial

Enterprise

• Shared functions let you reuse custom functions across multiple projects and fields within an organization, accelerating development and reducing maintenance. Save functions to the function library then import them into custom function fields.
• Validation functions for table and list fields can now return cell-level validation results.

Enterprise

• Resolved issue | Splitting functions now copy correctly when you copy a project. Previously, copied projects using splitting functions would error when running or editing the function.
• Resolved issue | Runs with only cross-class field validations were missing from the Review tab. These runs now appear in the Documents tab as grayed-out entries, and can be reviewed from the Runs tab.
• Resolved issue | Custom validation functions in cross-class fields auto-ran on every keystroke during editing. They now run only when you click Run code.
• Resolved issue | In human review, validation error messages weren’t displayed when classification confidence was 0.
• Resolved issue | Custom validation functions didn’t run when some arguments were null. Custom functions now run even when arguments are null, passing null as the argument value.

API & SDK

• The Get run results endpoint response has several updates.
• New metadata fields: run id, status, start_timestamp, and finish_timestamp.
• When the include_source_info query parameter is passed, a page_layouts object containing page dimension details (page_number, width, height) is returned.

25.50

25.50

Automate, deploy, & review

Commercial

Enterprise

• You can now create ground truth datasets directly from app runs in the App runs table. Select the overflow button for a run then click Create ground truth dataset.
• Webhook notification headers can be configured to pass custom keys defined in the deployment’s runtime configuration. Referencing secret keys isn’t supported.
• Resolved issue | In agent mode, fields sometimes got stuck in a loading state.
• Resolved issue | Workspace members with reviewer or review manager permissions were unable to filter runs by deployment on the Review tab.
• Resolved issue | File extension-based classification didn’t work as expected during app runs.
• Resolved issue | Extraction fields intermittently failed with authentication errors, returning No value found results. The issue was caused by OpenAI’s regional endpoint configuration and has been resolved by OpenAI.

API & SDK

• Resolved issue | The List runs endpoint didn’t filter by status correctly. The status query parameter now works as expected.

25.48

25.48

Automate, deploy, & review

Commercial

Enterprise

• Agent mode is generally available. Agent mode processes documents with next-generation models for enhanced extraction accuracy and simplified configuration. Update existing projects to agent mode to take advantage of these improvements.
• You can configure notifications to be alerted when run results are sent downstream.
• Upstream integrations include new processing options to align with how documents are grouped in your source systems and to better support packet-based processing. For connected drive integrations, you can choose to process subfolders as separate runs. For mailbox integrations, you can create a separate run for each email, instead of batching queued emails into a single run.
• Resolved issue | Webhook notifications weren’t sent when runs failed due to validation errors, such as when processing corrupted files.

Enterprise

• Classification custom functions let you write custom Python code to classify documents based on deterministic rules rather than relying on LLM classification.
• Splitting custom functions let you write custom Python code to split multipage files into documents using rule-based logic.
• Resolved issue | The cross-class validation rule Require matching fields incorrectly compared field values before cleaning lines were applied. It now compares final field values after cleaning lines are applied.

Administration

• Enterprise
• OAuth provider configurations now support a configurable scope claim (default scp) that lets you control per-token permissions using role-based access.

25.46

25.46

Automate, deploy, & review

Community

Commercial

Enterprise

• Resolved issue | App runs occasionally failed with the error Cannot open a non-existent file due to issues processing temporary files.

Commercial

Enterprise

• You can now create ground truth datasets from any app run in the selected workspace, not just app runs you executed yourself.

Enterprise

• Process packets of related documents with cross-class fields that consolidate data across multiple document types. In production, automation apps with cross-class fields process one packet per run. For details, see Extracting data from packets and Validating packets.

Conversations & chatbots

Community

Commercial

Enterprise

• Deprecation | Conversations and chatbots are deprecated and will be removed on February 1, 2026. You can continue to create and use conversations and chatbots until then, though feature updates and improvements are paused.
• Removal | URL upload for conversations and chatbots is no longer supported. Use alternative file upload methods.

Administration

• Enterprise

• Enterprise-tier organizations can now support IP allowlisting for outbound integrations, such as webhooks and deployment integration functions. Contact support for the egress IP ranges to allowlist.

API & SDK

• The Get run results API/SDK operation now includes the run ID and page layout information in its response.

Original source Report a problem

Community

Automation projects now support up to 250 fields per class. Optimizations in this release also improve processing performance for projects containing classes with large numbers of fields.

Commercial

Shared functions let you reuse custom functions across multiple projects and fields within an organization, accelerating development and reducing maintenance. Save functions to the function library then import them into custom function fields.

Enterprise

• Resolved issue | Splitting functions now copy correctly when you copy a project. Previously, copied projects using splitting functions would error when running or editing the function.
• Resolved issue | Runs with only cross-class field validations were missing from the Review tab. These runs now appear in the Documents tab as grayed-out entries, and can be reviewed from the Runs tab.
• Resolved issue | Custom validation functions in cross-class fields auto-ran on every keystroke during editing. They now run only when you click Run code.
• Resolved issue | In human review, validation error messages weren’t displayed when classification confidence was 0.
• Resolved issue | Custom validation functions didn’t run when some arguments were null. Custom functions now run even when arguments are null, passing null as the argument value.

API & SDK

The Get run results endpoint response has several updates.

• New metadata fields: run id, status, start_timestamp, and finish_timestamp.
• When the include_source_info query parameter is passed, a page_layouts object containing page dimension details (page_number, width, height) is returned.

Original source Report a problem

Commercial

• You can now create ground truth datasets directly from app runs in the App runs table. Select the overflow button for a run then click Create ground truth dataset.
• Webhook notification headers can be configured to pass custom keys defined in the deployment’s runtime configuration. Referencing secret keys isn’t supported.
• Resolved issue | In agent mode, fields sometimes got stuck in a loading state.
• Resolved issue | Workspace members with reviewer or review manager permissions were unable to filter runs by deployment on the Review tab.
• Resolved issue | File extension-based classification didn’t work as expected during app runs.
• Resolved issue | Extraction fields intermittently failed with authentication errors, returning No value found results. The issue was caused by OpenAI’s regional endpoint configuration and has been resolved by OpenAI.

Enterprise

API & SDK

• Resolved issue | The List runs endpoint didn’t filter by status correctly. The status query parameter now works as expected.

Original source Report a problem

Deprecation

New AI Hub Community sign ups are discontinued and organization members no longer receive a community account upon sign up. Existing community accounts aren’t immediately affected by this change. To get started with AI Hub, users can start a free AI Hub Commercial trial.

Automate, deploy, & review

• Community
• Commercial
• Enterprise

The description field for apps is increased from 500 characters to 3000 characters, allowing app developers to include more details about app functionality and usage.

• Resolved issue | Table viewer crashed when row or column data was missing.
• Resolved issue | In Safari, the Versions table in the app detail view was cut off and hidden behind a scroll bar.

Commercial

• The Python library fuzzywuzzy is now available for import in custom functions.
• Ground truth values are now preserved for renamed classes and fields using a generic ID instead of the class or field name.
• Workspace members with the developer role can create and edit deployments. This permission was previously limited to workspace managers and above.

Enterprise

• Conversations & chatbots

• Community

• Commercial

• Enterprise

• Resolved issue | Table viewer crashed when row or column data was missing.

Administration

• Community

• Commercial

• Enterprise

• You can now see more detailed usage information on the Usage settings page, using the Deep dive section.

• Commercial

• Enterprise

• You can connect Azure Blob Storage containers using managed identities for authentication.

API & SDK

• The create conversations and upload files and converse with a document API operations have new fast_mode parameters that can speed up conversation creation and response time.

• Resolved issue | API calls that used unusually long URLs failed for app runs, deployments, ground truth sets, and accuracy tests.

Original source Report a problem

Deprecation

New sign-ups for the AI Hub Community tier will be discontinued as of version 25.28. Existing Community accounts aren’t immediately affected by this change.

Enterprise

Single-tenant | Additional post-upgrade testing and validations have been added to ensure automation apps work as expected following scheduled maintenance.

Automate, deploy, & review

• Community
• Commercial
• Enterprise

Automation projects and apps can now factor in more input when processing documents, improving accuracy across all model-based processes. This change is available immediately in automation projects and existing community apps, and in commercial and enterprise apps as AI runtime 1.0.12.

• Table detection and extraction accuracy is improved for tables spanning multiple pages. This change is available immediately in automation projects and existing community apps, and in commercial and enterprise apps as AI runtime 1.0.12.

• Processing for Excel files is optimized for improved extraction accuracy. This change is available immediately in automation projects and existing community apps, and in commercial and enterprise apps as AI runtime 1.0.12.

• Resolved issue | CSV file upload failed if files used certain encoding.

Commercial

• Enterprise

• Resolved issue | In human review, the document view for non-admins displayed only the first 20 documents.

Conversations & chatbots

• Community
• Commercial
• Enterprise

This release upgrades the models used in conversations and chatbots, delivering improved accuracy and support for longer responses.

• Table detection and extraction accuracy is improved for tables spanning multiple pages.

• Processing for Excel files is optimized for improved extraction accuracy.

• Resolved issue | CSV file upload failed if files used certain encoding.

Administration

• Commercial
• Enterprise

Workspace role names have been updated. Member is renamed to Developer, and Manager is renamed to Workspace manager.

Enterprise

• Single-tenant | Organizations using Azure OpenAI endpoints can now authenticate using managed identities. For guidance, connect with Instabase Support.

API & SDK

• The create conversation operation supports documents written with non-Latin characters.
• The delete run operation returns status code 404 instead of 500 when the specified run doesn’t exist.
• The client.apps.runs.results() SDK method no longer throws incorrect errors about the post_processed_pdf_path option.
• The list batches operation takes an offset argument instead of file_offset.
• SDK version 0.11.0 is available, containing fixes described above.

Resolved issue

• Single tenant | API calls that used unusually long URLs failed for app runs, deployments, ground truth sets, and accuracy tests.

Original source Report a problem

Enterprise

Single-tenant | You can request to use Anthropic foundation models in your environment—through Amazon Bedrock—rather than the AI Hub default OpenAI models. Like other available models, Anthropic models are gated by AI runtime for automation apps. Connect with Instabase Support for more information.

Automate, deploy, & review

Community

Commercial
Enterprise
In the project field editor, changing the field type now preserves existing prompts or descriptions.

Commercial

Enterprise
Cleaning functions applied to table extraction fields now process large tables as a single unit, enabling global operations like sorting across all rows. This change is available immediately in automation projects and existing community apps, and in commercial and enterprise apps as AI runtime 1.0.11.
In human review, file search queries now match with either upper- or lowercase letters.
Resolved issue | When expanding the view for table extraction fields, an extra column—not included in the result—was displayed at far right.
Resolved issue | Accuracy tests failed to run and provided a vague error message if tables in the ground truth dataset were improperly formatted. A clearer error message now indicates which documents and fields need correction.
Resolved issue | In human review, changing document splitting or classification multiple times caused re-extraction to fail.
Resolved issue | Email notifications when items were queued for review included links to view, but not review, items.
Resolved issue | In the Hub, filtering for advanced apps returned results from only the first page.
Resolved issue | In deployment metrics, handling time averages were calculated inconsistently across charts. All handling time metrics are now standardized to include the full run time for each document.

API & SDK

The get run results API operation returns non-secret custom keys.
The get run results API operation includes error messages in its response body.

Conversations & chatbots

Community
Commercial
Enterprise
Table detection and extraction accuracy is improved for tables spanning multiple pages.
Resolved issue | When selecting page links in the Show steps view of the model response, the document viewer navigated one page ahead of the intended page.

Original source Report a problem