AI Descriptions - Descriptions for Products and Categories

Added

• Bulk generation: New option "Variants take over the new content". It removes the variants' own content of the generated fields for the processed main products, so the variants inherit the new content and show it in the storefront. Previously, variants with their own description (e.g. after a migration from Shopware 5) kept showing the old text. The removed content is backed up and brought back with "Undo changes".

Changed

• Prompt: If the global instructions contradict the plugin's general rules (e.g. on length, completeness or structure), the global instructions take precedence. Content they exclude no longer counts as missing information.
• Global instructions: up to 8,000 instead of 4,000 characters are passed to the AI.
• Bulk generation: meta description and keywords are derived from the description generated in the same run instead of the old description that was still stored. The meta title is still based on the product name and product data only.

Fixed

• Variants: Generating a description or SEO texts directly on a variant now uses the data inherited from the main product (name, manufacturer, categories, properties, description). Previously the AI received an almost empty product. The variant's own options (e.g. "Size: M") are passed to the AI as well and shown in the data preview.
• Variants: The switch "Use existing description as data source" is also available when the variant inherits its description from the main product.

Added

• Prompt: The structure follows the specifics of each product instead of a fixed template, so descriptions no longer all read alike.

Changed

• Existing description as data source: The AI rewrites the content completely instead of summarizing it. The text is at least as detailed as the source. Besides measurements and materials, unique selling points, special features, symbols and their meaning, gemstones and their traditional properties, origin, craftsmanship and application are now carried over explicitly. The example texts are left out of the prompt in this mode because they suggested short texts with a fixed layout.
• Existing description: up to 40,000 instead of 10,000 characters are passed to the AI.
• Higher output limit for descriptions (16,000 instead of 3,000 tokens), so long texts are generated completely. Only the tokens actually generated are billed.
• Longer timeouts for generations in the administration and on the server, so long texts and reasoning models do not abort early.
• OpenAI: Pro and Codex models are no longer listed in the model selection because they do not support the chat API.

Fixed

• OpenAI: GPT-5 and newer models (e.g. gpt-5.4-mini, gpt-6-luna) and the o-series can be used. Previously they failed with "Unsupported parameter: 'max_tokens' is not supported with this model". The plugin now sends max_completion_tokens, omits temperature for reasoning models and reserves additional tokens for reasoning. If a model rejects an optional parameter or allows a lower output limit, the request is adjusted and sent again automatically.
• The configured minimum length had not been passed to the AI since 2.1.0. It is taken into account again.
• Cut-off AI responses (output limit reached) now lead to an understandable error message instead of "not valid JSON".
• Anthropic: Models with a lower output limit are automatically requested with their maximum.

Added

• Data preview: In the generation dialogs for products and categories and in the bulk generation, "Show data preview" shows which data reaches the AI: product data, categories, properties, the existing description (cleaned, as it is passed), global and custom instructions after sanitizing, and the selected options. The bulk generation lets you page through the selected products. Nothing is generated and no costs are incurred.
• New setting "Use existing description as data source by default" ("Content & Instructions" card): preselects the switch in all generation dialogs and the bulk generation.
• Hint in the generation dialogs when custom or global instructions contain placeholders such as $product.translated.description or Twig expressions. They are not evaluated.
• Hint in the generation dialogs when the global instructions exceed 4000 characters and are therefore truncated.

Changed

• Prompt: If the existing description and the Shopware fields (e.g. weight in kg) disagree, the value from the description wins. Values from the description are no longer converted into other units.
• Prompt: Custom instructions are now followed for content, structure and wording as well, not only for style and tone. The protection against role changes and changes to the response format remains.
• Custom instructions are passed to the prompt without HTML escaping (quotes and & reach the AI unchanged).
• The category dialog uses the same validation as the product dialog.
• SEO generation: Meta description and keywords receive a longer excerpt of the saved description (up to 1000 instead of 300 or 400 characters). It is prepared like in the description generation: lines and bullet points are kept, and it is cut at a word boundary. Details such as dimensions and weights are therefore available for the meta texts as well.

Fixed

• Typo in the German help text of the custom instructions.
• SEO generation for categories: The meta title prompt contained the description with escaped HTML instead of plain text.

Added

• Bulk Generation: select up to 500 products in the product list and use "Generate AI content" to create the description, meta title, meta description and keywords in one run. Results are saved directly. Option "Only fill empty fields" keeps existing content.
• Two execution modes: in the background via the message queue (default, with a note about the required worker) or directly in the browser with a progress bar, for shops without a reliable queue. Browser runs can be cancelled at any time; the default mode and the concurrency are configurable.
• Cost notice with an estimate of requests and tokens. The generation only starts after explicit confirmation.
• Backups: the previous text is saved before it is overwritten. Restore a single product on the product page or a whole run in the dialog. Backups are deleted by a scheduled task after a configurable period (default 14 days); expired or all backups can also be deleted immediately in the plugin configuration.

Changed

• Provider error handling: reached spend limits, missing credit and invalid API keys are detected and no longer retried. A bulk run pauses in that case and can be continued once the problem is fixed. Rate limits are retried after the wait time sent by the provider (including retry-after-ms).

Fixed

• OpenAI: internal options such as _timeout were sent to the API in SEO generation requests.
• Length limit error messages in the generation modals did not show the limit values.
• Uninstall without "keep data": replaced the call to executeUpdate(), which was removed in DBAL 4.

Added

• Existing Description as Data Source: new switch in the generation modal for products and categories. When enabled, the saved description (cleaned, line structure preserved) is sent to the AI. All facts it contains, such as dimensions, weights, diameters, materials or gemstones, are carried over completely with exact values and additionally listed. Disabled by default.
• Global Instructions: new field in the plugin configuration ("Content & Instructions" card) for binding rules that apply to every description generation. Up to 4000 characters, independent of the custom prompt limit in the modal.
• Fixed Text Block: new editor field in the plugin configuration, e.g. for disclaimers or legal notices. The text is appended word for word after generation and is not sent to the AI, so it can no longer be shortened or replaced by placeholders. Can be switched off per generation in the modal.
• Physical Product Data: width, height, length (mm) and weight (kg) from the Shopware product fields are sent to the AI when maintained.

Changed

• Prompt: the rule "don't repeat specifications verbatim" was replaced. Numeric values and units must now be kept exactly; their benefit may be described in addition. New example for products with measurements.
• Generation Parameters: removed frequency_penalty, presence_penalty and top_p, which made it harder to keep exact values. Temperature 0.7, or 0.4 with the data source enabled or the "Technical" style. max_tokens raised to 3000.
• The custom instructions field in the modal is now a multi-line textarea.
• Product data is passed to the prompt without HTML escaping (quotes and & reach the AI unchanged).

Fixed

• Generation and SEO preview modals: after closing, the modal sometimes could not be reopened (Vue insertBefore error) because sw-modal moves its element to . The modals are now mounted inside a stable wrapper.
• The "Max Custom Prompt Length" and "Max Keywords Length" settings are now respected in backend and administration (capped at 2000 and 500 characters). Previously 500 and 200 characters were hardcoded.
• The "Enable Prompt Sanitization" and "Block on Injection Detection" settings are now applied during description generation.
• Replaced calls to EntitySearchResult::first(), which is removed in Shopware 6.8, with getEntities()->first().

Released 2026-05-04

Added

• Anthropic Claude Support: generate descriptions and SEO content with Claude alongside OpenAI. New provider selector in plugin configuration (OpenAI / Anthropic). The shared API key field applies to the active provider.
• Dynamic Model List: the model dropdown is now populated by calling the active provider's /v1/models endpoint, with a refresh button to reload after switching providers or keys. No more hardcoded model options that go stale between releases.
• Structured Output Enforcement: generated content is now constrained by a response schema. OpenAI uses native Structured Outputs (response_format: json_schema, strict mode), Anthropic uses a forced tool call with input_schema. Eliminates stray code-fence wrapping and "Here is your description:" preambles.

Changed

• Plugin Label: renamed from "AI Descriptions - OpenAI Integration" to "AI Descriptions - OpenAI & Anthropic Integration" so the admin plugin list and store listing reflect both supported providers.
• Provider-Agnostic Naming: internal classes renamed for the multi-provider architecture. OpenAIController → AiController, OpenAiService → AiService, OpenAiConfigStruct → AiConfigStruct, admin JS service openAiService → aiProviderService. Only relevant if you have overrides referencing these classes.
• Retry Strategy: retryable HTTP failures now honour the provider's Retry-After response header (capped at 30 seconds) instead of a fixed linear delay.
• Database Lookups: product, category, and language fetches use Shopware's indexed primary-key path (new Criteria([$id])) instead of an EqualsFilter on id.

Fixed

• Translation lookup in description generation now returns the requested field instead of always returning the translated name.
• The "Test API" button no longer sends the saved organizationId header when verifying an unsaved key, preventing false-negative results when switching organizations.
• Connect/transport failures from the AI provider now produce clean status-coded errors in the admin notifications instead of generic 500s.

Removed

• Hardcoded model options in config.xml (replaced by the dynamic model select).
• output-format.html.twig and output-management.html.twig prompt templates (their output rules are now enforced by the response schema).

Added

  - SEO Generation: AI-powered generation of meta titles, meta

  descriptions, and keywords for products and categories

    - Preview modal with character count validation for optimal SEO lengths

    - Review and edit generated content before applying

    - Visual feedback for character limits (green/yellow/red indicators)

  - Writing Style Variants: Five different writing styles for content

  generation

    - Standard, Technical, Emotional, SEO-focused, and Luxury variants

    - Configurable default variant in plugin settings

    - Style selector in product and category forms

  - Modular Prompt System: Component-based prompt architecture

    - Reusable prompt components (shop-context, entity-data, constraints,

  output-format, examples)

    - Language-specific context templates (DE/EN)

    - Better structured prompts with clear sections

  - Custom Prompts: Additional instruction field for fine-tuning AI

  generation

  - API Retry Mechanism: Automatic retries with exponential backoff for

  transient API failures

  - New OpenAI Models: Support for GPT-4.1 variants, o3-mini and o4-mini

  reasoning models

  Changed

  - Shopware 6.7 Compatibility: Full compatibility update for Shopware 6.7

    - Updated dependencies and service configurations

    - Migrated build system from webpack to Vite

  - Default Model: Changed from gpt-3.5-turbo to gpt-4o-mini for improved

  performance

  - UI Improvements: Streamlined generation workflow

    - Moved "Generate All SEO" button to bottom of form

    - Modal opens immediately with loading overlay

    - Added sparkle icons to generation buttons

  - Administration Notifications: Updated to use Shopware's built-in

  notification mixin

  - Configuration Page: Improved spacing and layout

  Fixed

  - Category form variant selector initialization issues

  - Template variable references in category detail menu

  - Prompt variant passing to API endpoints

  - Various UI state management bugs in admin components

  Security

  - Prompt Injection Protection: New security layer with PromptSanitizer

  service

    - Detection and blocking of prompt injection attempts

    - Secure boundaries in templates for user-provided content

    - Comprehensive input validation with configurable limits

  - Security Settings: New configuration card for customizable validation

  limits

Compatibility with Shopware 6.6 established

Release of the plugin