OpenAI AI Assistants & Agents Manager
Description
Highlights
- Central agent manager for the 5 Elements AI plugin suite (configure once, reuse across plugins)
- Backend Playground to test agents and validate tool calls before going live
- Self-growing knowledge base (Agent Logs): unanswered questions become reusable answers
- Chat QA & regression testing: Scorecards from chat threads (tools, tokens, latency) + prompt replay
- Includes the pre-configured 5 Elements Backend Assistant (BaBu) for server testing and validations
Features
- Create and manage agents (model, instructions, tools)
- Assign a knowledge base (vector store) per agent
- Playground chat for iterative testing and prompt/tool validation
- Central log overview for all AI plugin interactions (monitoring + troubleshooting)
- Agent Logs as knowledge entries with dedicated **Answer** field
- Unresolved logs still appear in search results (flagged as unresolved)
- `tagAliases` to connect related tags without duplicating entries
- `searchTerms` for exact phrases (fallback) to reduce duplicates caused by wording variations
- “Fetch URLs” per log entry with automatic retrieval + sanitizing when the log is used
- Optional `fetch_urls` tool for on-demand URL fetching
- Knowledge-base tools: `tags_for_logs`, `list_logs`, `search_logs`, `log`
- Shopware API tool integration (orders, payments, products, etc.)
- Backend Agent tools for knowledge maintenance (`get_unresolved_logs`, `edit_log`, `delete_log`, `staged_content`)
- Supports Context-List-Pattern
- Chat thread logging (optional) + Scorecard generator: review full conversations, token usage, caching ratio and latency; export prompts for auto-chat regression runs
About the Extension
OpenAI AI Assistants & Agents Manager is the backend control center for creating, configuring and testing OpenAI-powered agents in Shopware 6. It is the foundation for our advanced AI plugins ("NextGen OpenAI Chatbot", "AI Product Advisor", "Backend AI Analyst"…).
Under the hood, the manager uses the OpenAI Responses API (the “Assistants” naming is legacy).
Why this exists
This manager started as an internal configuration layer for our first customer support chatbot. When we planned additional plugins (for example a Product Advisor), it became clear that duplicating agent configuration in every plugin would create unnecessary overhead for shops. So we decoupled the manager, extended it into a central backend control center, and made it free.
The goal is simple: shop owners should be able to configure and test agents as long as needed, directly in the backend, and decide if the results match their expectations. If the results do not fit, it should be safe to remove and replace without having paid just for an “AI” label.
Shopware-native, self-growing knowledge base (human-in-the-loop)
Agent Logs turn unanswered questions into reusable answers—directly in the Shopware backend. They form a long-term knowledge system with a dedicated Answer field.
- A user asks a question and the agent searches the knowledge base (if enabled).
- If nothing matches, a new log is created automatically and marked as unresolved.
- You add an Answer (and optionally aliases/search terms) and mark it as resolved.
- From then on, agents can reuse that Answer in future chats.
Unresolved logs can still appear in results (flagged as unresolved). This helps agents recognize that the topic already exists, even if the final Answer is not available yet.
URL-backed knowledge
Each knowledge entry can store multiple source URLs (“Fetch URLs”). When an agent finds the entry, the system automatically fetches and sanitizes the URL content and returns it as part of the entry content. No extra agent tool call is required.
Playground chat and observability
Use the Playground chat to test agents and validate tool calls in the backend. A unified log area helps you trace AI plugin activities and troubleshoot issues.
Important: The Playground chat runs in the context of the selected Sales Channel (language, currency, rules, etc.), so it reflects the same conditions as the storefront. With Backend mode enabled (default), responses are intentionally more verbose, and additional debug details are available in the logs/trace area. Disable Backend mode to see responses that match what the same agent will return in the storefront for that Sales Channel.
Note on product search results: To keep tool payloads lean, agents work with a condensed product representation, and the backend Playground shows exactly what the agent returns. Storefront-facing plugins (e.g. the Chatbot) additionally wrap the optimized products.response in a dedicated storefront template to present richer listings. This template layer is intentionally maintained only in storefront plugins to avoid duplicating UI maintenance (the product listing template alone is 300+ lines of Twig). We may add a bridge later to close this last UI gap between backend and storefront.
What makes this manager different
Built for transparency (fully inside Shopware)
This extension implements its tooling directly in the Shopware backend. It does not rely on embedded external dashboards or third-party UIs. That means you can inspect and change what your agents do in the same place you configure them.
- Agent instructions are visible and editable (including examples and schemas), prominently placed in the agent create/edit route.
- A Content Checker provides a human-friendly interface to test all available tools directly in the backend.
- The Content Checker UI is generated from the same tool definitions agents use, so the available parameters match exactly.
- Tool responses shown in the UI are the actual payload agents receive (not a simplified preview).
Quality assurance (Scorecards)
If chat logging is enabled, final agent conversations are stored and reviewable in the backend. You can open a full thread view, inspect per-message token usage, and see aggregated totals.
For quality assurance, the manager can generate a Scorecard from any chat thread. A Scorecard bundles the complete chat context plus tool metadata and performance data (tokens, caching ratio, latency), so you can analyze behavior consistently and repeatably.
You can also export the user prompts as a plain text list (one prompt per line) to replay the same sequence with the auto-chat utility and verify the system still behaves like yesterday.
See what your AI sees (no black box)
Test tools with an interactive UI and inspect what data your agents receive before you deploy anything.
- Live tool testing for available tools
- Input/parameter preview and validation
- Tool response visualization
- Usage tracking per agent
Knowledge base that improves over time
When customers ask new questions, missing information becomes a new Agent Log entry. You resolve it once (Answer + optional aliases/search terms) and agents can reuse it in future chats.
- Unresolved entries can still appear in searches (flagged as unresolved)
- Exact searchTerms (fallback) and tagAliases help prevent duplicates caused by wording variations
Built to avoid duplicated setup across plugins
This manager exists as a shared foundation so you do not end up configuring the same agent logic multiple times when you use multiple AI extensions.
Requirements / important notes
- An OpenAI API key is required (with API credits). The key must be configured for all sales channels you use.
- This is a base plugin. Without one of our AI plugins, it does not provide standalone storefront functionality.
- If you use the older plugin “Custom GPT Sales and Support Assistant – Chatbot, Shopping Helper”, uninstall it first to avoid conflicts.
- Shopware 6.7 compatible; updated to the OpenAI Responses API.
Documentation:
https://docs.5-elements.cloud/oai-assistants-manager/
OpenAI AI Assistants & Agents Manager is the backend control center for creating, configuring and testing OpenAI-powered agents in Shopware 6. It is the foundation for our advanced AI plugins ("NextGen OpenAI Chatbot", "AI Product Advisor", "Backend AI Analyst"…).
Under the hood, the manager uses the OpenAI Responses API (the “Assistants” naming is legacy).
Why this exists
This manager started as an internal configuration layer for our first customer support chatbot. When we planned additional plugins (for example a Product Advisor), it became clear that duplicating agent configuration in every plugin would create unnecessary overhead for shops. So we decoupled the manager, extended it into a central backend control center, and made it free.
The goal is simple: shop owners should be able to configure and test agents as long as needed, directly in the backend, and decide if the results match their expectations. If the results do not fit, it should be safe to remove and replace without having paid just for an “AI” label.
Shopware-native, self-growing knowledge base (human-in-the-loop)
Agent Logs turn unanswered questions into reusable answers—directly in the Shopware backend. They form a long-term knowledge system with a dedicated Answer field.
- A user asks a question and the agent searches the knowledge base (if enabled).
- If nothing matches, a new log is created automatically and marked as unresolved.
- You add an Answer (and optionally aliases/search terms) and mark it as resolved.
- From then on, agents can reuse that Answer in future chats.
Unresolved logs can still appear in results (flagged as unresolved). This helps agents recognize that the topic already exists, even if the final Answer is not available yet.
URL-backed knowledge
Each knowledge entry can store multiple source URLs (“Fetch URLs”). When an agent finds the entry, the system automatically fetches and sanitizes the URL content and returns it as part of the entry content. No extra agent tool call is required.
Playground chat and observability
Use the Playground chat to test agents and validate tool calls in the backend. A unified log area helps you trace AI plugin activities and troubleshoot issues.
Important: The Playground chat runs in the context of the selected Sales Channel (language, currency, rules, etc.), so it reflects the same conditions as the storefront. With Backend mode enabled (default), responses are intentionally more verbose, and additional debug details are available in the logs/trace area. Disable Backend mode to see responses that match what the same agent will return in the storefront for that Sales Channel.
Note on product search results: To keep tool payloads lean, agents work with a condensed product representation, and the backend Playground shows exactly what the agent returns. Storefront-facing plugins (e.g. the Chatbot) additionally wrap the optimized products.response in a dedicated storefront template to present richer listings. This template layer is intentionally maintained only in storefront plugins to avoid duplicating UI maintenance (the product listing template alone is 300+ lines of Twig). We may add a bridge later to close this last UI gap between backend and storefront.
What makes this manager different
Built for transparency (fully inside Shopware)
This extension implements its tooling directly in the Shopware backend. It does not rely on embedded external dashboards or third-party UIs. That means you can inspect and change what your agents do in the same place you configure them.
- Agent instructions are visible and editable (including examples and schemas), prominently placed in the agent create/edit route.
- A Content Checker provides a human-friendly interface to test all available tools directly in the backend.
- The Content Checker UI is generated from the same tool definitions agents use, so the available parameters match exactly.
- Tool responses shown in the UI are the actual payload agents receive (not a simplified preview).
Quality assurance (Scorecards)
If chat logging is enabled, final agent conversations are stored and reviewable in the backend. You can open a full thread view, inspect per-message token usage, and see aggregated totals.
For quality assurance, the manager can generate a Scorecard from any chat thread. A Scorecard bundles the complete chat context plus tool metadata and performance data (tokens, caching ratio, latency), so you can analyze behavior consistently and repeatably.
You can also export the user prompts as a plain text list (one prompt per line) to replay the same sequence with the auto-chat utility and verify the system still behaves like yesterday.
See what your AI sees (no black box)
Test tools with an interactive UI and inspect what data your agents receive before you deploy anything.
- Live tool testing for available tools
- Input/parameter preview and validation
- Tool response visualization
- Usage tracking per agent
Knowledge base that improves over time
When customers ask new questions, missing information becomes a new Agent Log entry. You resolve it once (Answer + optional aliases/search terms) and agents can reuse it in future chats.
- Unresolved entries can still appear in searches (flagged as unresolved)
- Exact searchTerms (fallback) and tagAliases help prevent duplicates caused by wording variations
Built to avoid duplicated setup across plugins
This manager exists as a shared foundation so you do not end up configuring the same agent logic multiple times when you use multiple AI extensions.
Requirements / important notes
- An OpenAI API key is required (with API credits). The key must be configured for all sales channels you use.
- This is a base plugin. Without one of our AI plugins, it does not provide standalone storefront functionality.
- If you use the older plugin “Custom GPT Sales and Support Assistant – Chatbot, Shopping Helper”, uninstall it first to avoid conflicts.
- Shopware 6.7 compatible; updated to the OpenAI Responses API.
Documentation:
https://docs.5-elements.cloud/oai-assistants-manager/
Details
- Available: English, German
- Latest update: 9 February 2026
- Publication date: 25 June 2025
- Version: 2.0.7
- Category: Administration
Resources
Reviews (0)
About the Extension Partner
5 Elements Websolutions
Partner Status
-
Shopware
Extension Partner
Details
-
Ø-Rating:
1
Average rating of 1 out of 5 stars
- Partner since: 2019
- Extensions: 9
Support
- Based in: Germany
- Speaks: German, English
- Response time: Very quickly
Follow these steps to install and configure the OpenAI Assistants Manager plugin, including setting up the 5 Elements Backend Assistant (BaBu):
Install and Activate
- Install the plugin via the Shopware 6 Plugin Manager.
- Activate the plugin after installation.
Set Up API Connection
- Navigate to the plugin’s settings in the Shopware backend.
- Enter your OpenAI API key under the "API Settings" section.
- Attention: The API-Key need to be set for all Saleschannels
- Click the "Test API Key" button to verify the connection.
Clear Cache
- Clear the Shopware cache to ensure all changes are applied.
Create and Manage Assistants
- Open the plugin main interface (Content → 5E OAI Assistants Manager) to create and configure assistants.
- Assign assistants to specific AI plugins, such as the Customer Support Chatbot or Product Advisor.
One-Click BaBu Setup
- Go to Content → 5E OAI Assistants Manager in the Shopware backend.
- Click the button to automatically create the 5 Elements Backend Assistant (BaBu).
- Use BaBu to test server functions, run stress tests, and validate function call efficiency.
Utilize the Playground
- Use the playground area to test and refine assistant behaviors and responses.
- Adjust settings to optimize performance based on testing results.
Review Logs
- Access logs for all AI plugins in the consolidated log section.
- Monitor assistant interactions and performance for troubleshooting or insights.
Save and Sync
- Save all changes and synchronize settings to apply them across integrated AI plugins.
2.0.7
- Fixed display of OpenAI internal error messages such as "insufficient_quota"
- Refined Chat Tool instructions for the "autoResolveAgentLogs" workflow
- Added an extensive help section about the self-growing knowledge database
- Renamed navigation labels for better clarity (e.g. "Agent Logs" → "Knowledge Base")
2.0.6
- Standard instructions improved: Context-first tool calls are prioritized so that agents have a good starting point.
2.0.5
- Added a per-agent Background mode toggle that sets `background: true` to enable asynchronous chats.
- Added "Custom Policy / Behavior Fix" field—injected as developer message (high priority) on every turn for quick per-agent behavior overrides.
- Added Developer Message textarea in Playground for testing high-priority agent instructions.
- Introduced `staged_content` tool—paste large content (FAQ, docs, data) into Playground and let agents retrieve it on demand.
- Added `BackendToolsExecutor` pattern for backend-only tools with `isBackendChat` protection.
- Added file purpose selector to uploads (assistants, batch, fine-tune, vision, user_data, evals) with dynamic file type filtering.
- Added Backend/Frontend toggle in Content Checker to test tools with different access levels.
- Updated default tag set for one-click installation.
- Extended message logs to include `_meta_data` from agent responses alongside arguments.
- Refined minor layout issues in the UI.
- Added a ScoreCard service that analyzes each chat run, generates scoring data, and dispatches downloadable reports for further human or LLM review.
- Introduced a beta “Self-Learning Agent” chat tool to experiment with autonomous training loops inside the Agents Manager.
2.0.4
- Tool selection now highlights [CLP] (Context-List-Pattern) tools—agents learn your catalog dynamically, no training needed.
- Improved utility instructions for Backend Assistant tools.
2.0.3
- Fixed UTF-8 encoding corruption in tool arguments where unicode escapes were malformed (e.g. `\u0000e4` → `\u00e4`).
- Added [Context List Pattern] to default instructions—guides agents to fetch available data first, then query precisely.
2.0.2
- Introduced the `list_product_names` tool so agents can grab a lightweight, real-time index of product names before crafting searches.
- Polished the default instructions/snippets to better match the Responses workflow and reduce redundant tool usage.
- Fixed UTF-8 handling for German umlauts (and other multibyte characters) inside tool arguments, preventing failed searches and noisy log entries.
2.0.1
- Added hook-based extension system for plugins to register custom tools.
- Improved search results with fallback cases for categories and max_price.
- Optimized default instructions for gpt-5-mini.
- Fixed category fetching to load all categories instead of only 100.
- Added abort request controller to playground chat.
- Added 4 new tools for backend agents.
- Added configurable scheduled database cleanup with dual protection (age and count limits).
- Added "Initial Instructions" field for agents to set per-agent initial chat guidelines.
- Various UI enhancements and refined translations.
2.0.0
Updated the plugin for Shopware 6.7
1.0.6
- Removed code_interpreter from UI (adds complexity without benefit for e-commerce; use vector_store instead)
- Agent Export/Import: Backup and restore agents as JSON files with "skip existing" or "update existing" modes
- Fixed: File upload to OpenAI was missing content-type
- Transition to Responses API complete
1.0.5
- Migration to Responses API: Due to the deprecation of OpenAI's Assistants API ~2026, the plugin now fully uses the Responses API. Thread management has been moved from OpenAI servers to the shop database, providing better control and reliability. This brings faster, more stable multi-tool answers and the backend playground chat was updated accordingly.
- Chat log view improved: every message is stored as its own entry, so you can scroll and filter conversations much faster. A new log route powers the view; the former endpoint stays available for older integrations until the next release.
- Message Markers: Added the ability to mark individual user messages with markers (interesting, training, review, example, flagged). Multiple markers can be applied to a single message. Filter conversations by marker type to quickly find messages marked for training data, review, or as examples. Markers use an efficient MySQL SET column type with indexing for fast filtering.
- Cached Token Counter: Token usage display now shows cached tokens alongside total tokens (e.g., "Tokens: 116,754 (77,952 cached)"). This provides visibility into prompt cache effectiveness and cost savings from OpenAI's cache feature, which can reduce costs by 50-90% on cached tokens.
1.0.4
- update support version up to <6.7
1.0.3
- fetchUrl Extension: Added new functionality to support fetching predefined URLs set in the shared knowledge database without requiring assistants to have the fetchUrl tool enabled. You can now define URLs to be loaded internally during content searches with the search_logs tool. The fetched content is sanitized and returned directly as output, reducing unnecessary tool calls by combining search and fetch into a single process, thereby improving performance.
1.0.2
- Updated composer dependency constraints: Now supports Shopware versions up to 6.6.10.5
1.0.1
- Fixed price display issue: Sorting by parentId was confusing when variants have different prices. Added price-range display when necessary.
- Added a separate page for query logs that assistants have used in search_logs and product_search. Accessible via the "Manage Tags" page.
1.0.0
- Initial plugin release