According to technical reports published on Dev.to and Odiseo’s corporate blog, software developers are building a native, self-hosted AI shopping assistant specifically tailored for the Sylius e-commerce platform.
Unlike standard third-party tools that rely on external servers and catalog synchronization, this new plugin reads directly from the store database, operates on the actual shopping cart, and executes locally using the store's own API credentials.
Core Architecture and Symfony Integration
The development team explained that the system is built by porting Anthropic’s reference implementation for commerce agents from Python into PHP and Symfony.
While the specific shopping vertical is tailored for e-commerce, the non-commerce-specific core has been decoupled and released as an open-source, MIT-licensed Symfony bundle named odiseoteam/ai-conversational-agent-bundle.
This reusable core encapsulates the agent loop, skills, memory management, budgeting controls, evaluation harnesses, and browser streaming capabilities, allowing developers to utilize the same underlying architecture for other automated assistants.
Features and Administrative Controls
The shopping assistant integrates directly into the customer-facing chat interface to handle complex queries, such as comparing products side-by-side or building multi-item purchasing plans based on natural language descriptions.
The system modifies the real Sylius cart, calculates precise shipping rates using built-in store methods, answers return policy questions, and tracks signed-in customer orders.
On the management side, store operators receive an administrative interface to review conversation transcripts, track assisted revenue, set spending caps, choose specific AI models per channel, and define custom tones of voice or store-specific rules.
The developers noted that the assistant has been successfully tested on Sylius versions 2.0 through 2.3.
Guardrails and Performance Metrics
To address common reliability concerns like price hallucinations or prompt injection, the architecture incorporates strict guardrails enforced directly in code rather than relying entirely on prompt engineering.
Provenance constraints ensure the agent can only display products whose IDs are returned via tool calls during the active conversation.
Additional safeguards include grounding lookups for catalog data and fencing mechanisms that treat product descriptions and tool results as strictly delimited data.
In internal benchmarks cited across the reports, conversations processed using Claude Haiku 5.5 configured at low effort performed exceptionally well, maintaining an average operational cost of under US$0.006 per conversation.
Why Self-Hosted E-Commerce AI Architecture Matters
The introduction of a native, self-hosted AI shopping assistant for Sylius highlights a broader architectural shift for online merchants who want advanced automation without surrendering data ownership to third-party software-as-a-service providers. Well-known conversational commerce tools, such as Fin and Gorgias, are predominantly optimized for platforms like Shopify and typically require store owners to synchronize their product catalogs to external vendor servers. While functional, that model forces merchants to trust third-party infrastructure with sensitive inventory, customer cart states, and pricing rules, while also leaving pricing models and model selection entirely in the hands of the vendor. By contrast, building natively within PHP and Symfony allows Sylius-based stores to maintain absolute control over their database queries and infrastructure costs. Utilizing custom API keys means store operators can seamlessly swap underlying language models, implement strict spending caps, and ensure that customer conversation data remains strictly within their own corporate boundary. Furthermore, open-sourcing the core framework enables the developer community to adapt conversational agents for alternative use cases, ranging from post-sales support to entirely non-transactional websites. As AI integration matures across enterprise software, this self-hosted paradigm offers a blueprint for developers seeking deep framework integration without vendor lock-in, emphasizing code-level guardrails and strict data provenance over black-box cloud connectors.
Editorial graphic and headline text by Signal Titans.
Source: Why we're building a shopping agent for Sylius · Why we're building a shopping agent for Sylius | Odiseo
SIGNAL TITANS
Leave a Reply