Main features

1. ILLMSummary behavior with allow_llm_summary and llm_summary fields.

2. generate_summary_for(obj, request, properties=None) — single-call API that resolves the right agent for the object’s portal_type and runs it via the core executor.

3. Content-type-specific agents via naming convention summarizer:<PortalType> (falls back to summarizer).

4. generic_metadata enricher (GenericMetadataProvider) — contributes title, description, language, geographic coverage, and temporal coverage to the user prompt. Reads from in-progress properties if passed, so unsaved edits are reflected.

5. Event subscriber for automatic summary generation on save.

6. @llm-summary REST endpoint for single-object generation.

7. @llm-summary-batch REST endpoint for batch generation.

8. Default agent definitions shipped in agents.json — override via the control panel agents_json.

9. POST <container>/@image-metadata for validated, stateless metadata generation before ordinary Image creation.

10. POST <image>/@image-alt for a validated, non-persisting block-local alt suggestion.

11. POST <object>/@image-metadata-regenerate for validated, non-persisting metadata suggestions when editing existing Dexterity content.

Image APIs

@image-metadata accepts a base64 image payload plus a safe language code on a folderish container. The optional portal_type identifies the content being added (default Image for existing callers), not its container. The caller must have the ordinary Add portal content permission, the requested FTI’s construction permission, and satisfy local container type constraints. It validates and bounds the image entirely in memory, calls a structured multimodal agent, and returns plain title and description values. It creates no content, blob, catalog record, or temporary file.

@image-alt is registered for the standard Plone IImage interface and requires Modify permission on the Image. It analyses verified Image bytes and metadata and returns plain alt text without changing the Image.

@image-metadata-regenerate requires Modify permission on the existing object (guarded by cmf.ModifyPortalContent). It accepts field (the schema image field’s name) and an optional language. With no image payload it reads the stored field; a supplied base64 image takes precedence and is validated as an unsaved replacement, never persisted. Download descriptors and URLs are not fetched. The field must be a named-image schema field on the object or one of its behaviors. The endpoint returns plain title and description suggestions through the same generation pipeline as Add without mutating the object.

The form widget supports both Add and Edit for the explicitly registered image and preview_image fields. On Edit it uses the loaded object’s content type, confirms replacement of existing metadata, and sends any unsaved replacement bytes instead of asking for the stored source. Stored regeneration changes only unsaved title/description; the file remains unchanged. Normal form Save persists edits; form cancellation discards them. Late results are ignored after source/context changes or cancellation, and metadata edited during analysis is preserved with an explicit retry message.

None of the Image endpoints uses AI markers or depends on the ILLMSummary behavior. Generic summary behavior remains opt-in and behavior-driven.

Image AI is disabled by default. In the shared GenAI controlpanel, enable both the global GenAI switch and Enable Image AI features to allow Image metadata and alt generation. Disabled requests return HTTP 403 with the image_ai_disabled error code; ordinary uploads and manual editing remain available.

Allowed content types defaults to Image only. An explicitly empty list allows all types; it does not bypass permissions or add support for otherwise unsupported image sources. All Image endpoints enforce this policy and return HTTP 403 with image_ai_not_allowed for an excluded target type. Add metadata generation checks the requested new object’s type; regeneration checks the existing object’s actual type. Alt generation checks the linked Image’s actual type, never a submitted type or the containing page’s type. The public Image settings include allowed_content_types; frontend controls fail closed when this policy is missing or failed to load.

The Image AI section defines positive character limits: a shared title/alt maximum (default 40) and a description maximum (default 500). Limits apply to trimmed generated text, not manually edited metadata. They are supplied in prompts and enforced on returned values. Oversized output is retried once; if it still exceeds a limit, generation fails without truncation. Metadata fields are retried together; filename-derived titles share the same single retry budget.

The section is contributed through a named IGenAISettingsSection utility in configure.zcml. Its records are installed by this package’s registry.xml using the core settings prefix. On existing sites, the owner must run the eea.genai.summary profile upgrades to 2.1 after deploying/restarting the backend, before using the extended controlpanel. The 1.0 to 2.0 step installs Image AI settings; the 2.0 to 2.1 step adds the content-type policy on sites that already have those settings. The upgrade imports only the registry step, without purging records or running import dependencies. It preserves existing settings and is safe to repeat. No eea.genai.core upgrade or legacy Image behavior cleanup is included. Missing Image records use safe runtime defaults and do not enable generation.

The theme loads allowlisted flags from @genai-settings during SSR into state.reduxAsyncConnect.genAISettings, just like header settings. Changes take effect in frontend controls on the next full page load; generation requests always check current backend settings.

Install

• Add eea.genai.summary to your requirements.txt.

• Install the GenericSetup profile.

• Activate the eea.genai.summary behavior on the desired content types (Site Setup → Dexterity Content Types → Type → Behaviors).

Customization

Override the default summarizer agent for a content type by registering a more specific agent in either ZCML or the control panel agents_json:

<genai:agent
    name="summarizer:EEAFigure"
    class=".agents.FigureSummarizerAgent"
/>

To inject extra prompt fragments (e.g. dataset-specific instructions), write an IEnricher in your own package and reference it from the agent’s enrichers list — no need to subclass anything in eea.genai.summary.

Copyright and license

The Initial Owner of the Original Code is European Environment Agency (EEA). All Rights Reserved.

All contributions to this package are property of their respective authors, and are covered by the same license.

The eea.genai.summary is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 2 of the License, or (at your option) any later version.

Secret Scanning

This repository uses the Betterleaks GitHub Action to scan the current repository content on every push and pull request. The scan uses the rules in .gitleaks.toml and uploads a betterleaks-report artifact when a finding is detected.

If the optional SMTP secrets are configured, failed scans also send an email to the last commit committer. The workflow expects these repository or organization secrets:

• SMTP_URL

• SMTP_PORT (optional, defaults to 25)

• SMTP_EMAIL

• SMTP_PASSWORD (optional if the SMTP server does not require authentication)

Port 465 is sent with direct TLS; other ports use the default SMTP handshake. The email includes a short finding summary from the redacted Betterleaks report, including the redacted matched line from each finding.

There are three common outcomes:

1. Everything is OK. The Betterleaks / Scan for secrets check is green and no action is needed. Regular references to runtime values are OK, for example:

   token_from_cookie = request.cookies.get("auth_token")

2. A real secret was found. The check is red and the workflow log asks you to download the betterleaks-report artifact. Open the artifact from the GitHub Actions run and check the reported file, line and rule. Remove the committed value, move it to the proper secret store, and rotate it if it was exposed. A report entry looks like this:

   {
     "RuleID": "secret-literal-assignment",
     "File": "src/config.py",
     "StartLine": 12,
     "Secret": "[REDACTED]"
   }

3. The finding is a false positive. Keep the value only if it is clearly not sensitive, such as a test fixture, placeholder, or public example. Add betterleaks:allow on the same line and include a short explanation in the pull request:

   test_password = "admin"  #betterleaks:allow

Do not add betterleaks:allow to real credentials.

Changelog