Release 2026.10.14-1

Evaluation baselines, deployment models changed without a redeploy, and the first self-hosted image since 2026.10.02-3, with single sign-on and the Agent Review step.

Cloud, United States
Cloud, Australia
Self-hosted image
ghcr.io/docaifabric/docaifabric:2026.10.14-1

Security

  • Self-hosted image: the first image since 2026.10.02-3. It carries every change of the cloud releases 2026.10.09-1 to 2026.10.13-2, their security fixes included, and of the cloud releases of 3 to 6 October, which have no page. Those add, among others, access checks on transaction lists, bulk actions, the work queue and the Assistant’s editing tools, storage paths confined to their own organization, project and data set, business rule web requests that cannot reach internal addresses, and project ZIP import and folder import connectors kept inside their own folders. Read the upgrade notes below before updating.
  • Evaluation analytics read only evaluation data sets, and leave out the transactions you may not open.
  • A download link an AI agent gets for an export file opens only what that agent may read.
  • Checkpoint and model test responses no longer include internal error text.

New

  • Evaluation baselines: keep a run as a baseline and compare every configuration change against it, on the Analytics tab and in the viewer.
  • Deployment models without a redeploy: on a self-hosted deployment, add, change or retire the models everyone uses without a redeploy, and see what uses a model before removing it. See Model configuration.
  • Self-hosted image: new to the image since 2026.10.02-3 are Single sign-on, the Agent Review step (Preview), What Memory collects, Copy to ground truth and Deployment administration by AI agents, and, from the cloud releases of 3 to 6 October, Cost per transaction, License usage forecast and history, Document processing by AI agents and Workflow commands in the CLI. Voice input in the Assistant can use Azure AI Speech. See Enable voice input.

Improved

  • Agent Review (Preview): the built-in agent learns Lessons from your reviewers’ corrections, and people can read and pin them. The Activity Log shows the agent’s run turn by turn, with its reasoning.
  • Agent Review (Preview): Human Review pauses after an agent’s Reject or Send To and lets through what the agent approved. A new project’s workflow has the step on, with Approved going to Export.
  • Reporting shows review taken by agents, and who decided, by kind of reviewer and by step.
  • The Activity Log says what each step did: Validate counts, issues to review, steps that were off, the exit taken.
  • A step that names a model it cannot use fails, instead of running on another.
  • Workflow model calls try the next endpoint when one refuses the key or cannot find the model.

Fixed

  • Transaction integrity rules match documents by their class.
  • A Review step reached from both a failure exit and a decision exit waits for the decision.
  • Saving rule code refuses code the rule engine would refuse to run.
  • Data Transform scripts get each page’s text and barcodes.
  • Bounding boxes are found for amounts printed with a decimal comma, grouping or no decimals, and table cells the model reformats keep their original text.

API

  • da admin models: add, change or remove the deployment’s models, and da admin models who lists what uses a model. See Model configuration.
  • get_export_files: an AI agent reads the text export files of up to 100 transactions in one call. See AI agents.
  • Self-hosted image: a step’s settings (its model, thresholds, what it sends) are changed on the workflow step. PATCH .../settings refuses them with 400 (step_settings_on_workflow). See Where settings live.
  • Self-hosted image: an AI agent connected over MCP sees one short tool list on every client and finds every other tool with find_tools, running it through execute_read, execute_write or execute_write_destructive; load_tools and tool groups are gone. See AI agents.
  • Self-hosted image: da skill install is gone; an agent reads the deployment’s playbook with da skill show. See Let your AI agent do it.

Upgrade notes

For a self-hosted deployment updating from image 2026.10.02-3 or earlier. Take a backup first.

  • Encryption key: if .env has neither CREDENTIAL_ENCRYPTION_KEY nor TENANT_MODELS_ENCRYPTION_KEY, set CREDENTIAL_ENCRYPTION_KEY. Without either, saving or testing a model with an API key is refused, and a model with an API key is left out of processing. See Updating to a version that needs an encryption key.
  • Secrets re-encrypted per organization: at the first start, every saved secret is re-encrypted in the background. An older image cannot read them, so to roll back, restore the backup you took before updating. With several nodes, update every node or stop the older ones first. See Updating to the version with a key per organization.
  • Step settings: after it starts, the application converts every project so that each step holds its own settings, and keeps the originals in each project’s step_settings_backup/ folder. da admin migrations status shows where it stands. A Data Transform script that read a step setting from project_settings no longer finds it there. See Updating to a version where steps hold their own settings.
  • Folder import from a mounted share: a connector watching a folder you mounted yourself (custom mounted path) needs that folder allowed in INGEST_CUSTOM_ROOTS, set to the container-side paths of your shares. Set it before or with the update, or those connectors stop with a configuration error. See Allowing custom folders.
  • Models: at the first start, the deployment stores its models from the LLM_* variables in .env. From then on, change models with da admin models, not in .env. See Where the models come from.

All releases