A stated goal becomes a finished look, so a shopper who cannot phrase a query still gets a usable answer.
Solution Studio Shopping Greeter
Solution Studio
Solution Studio Product Insights Agent
Shoppers get an answer on the product page in under a second, grounded in documents the brand controls.
Conversational widget that answers shopper questions on the product detail page, grounded in product docs.
What the agent does
Answers a shopper question about a specific product from the brand's own documentation, refusing anything the source does not support.
- Contributed by
- MACH Alliance (Reference implementation)
- Industry
- Retail, Cross-industry
- Audience
- B2C
01
The problem
Why this exists
Shoppers ask product questions on detail pages and either give up, send the question into a contact form, or guess. Static FAQs cover the most common 20 questions and the rest go unanswered. Every unanswered question is a step closer to a return, a lost sale, or a support ticket the brand has to handle on the back end.
Read more
The questions themselves are not the hard part. Brands already hold the answers, scattered across product manuals, ingredient lists, specs sheets, and policy documents. The hard part is delivering an answer that is grounded in the actual documentation, fast enough that a shopper does not bounce, and safe enough that the brand will sign off on letting a model answer.
02
Outcomes
What this demonstrates
- An answer lands on the product page in under a second, retrieved from the product's own documentation rather than generated from what the model already knows.
- The brand controls what the model can say and can audit what it said, because every answer is grounded in documents the brand owns.
- A four-layer guardrail stack mapped to the OWASP LLM Top 10 keeps the agent inside its lane and rejects prompts that try to push it elsewhere.
Read more
- LangSmith traces every request end to end, so a team can catch a retrieval regression before it reaches a customer.
- The widget drops onto an existing product detail page with a single script tag. No rebuild of the storefront required.
03
How it works
Architecture
The reference implementation ships as two sibling Next.js apps. The product-insights-agent app holds the widget, the chat API, the RAG pipeline, the guardrails, and a small control panel. The storefront app is an example retailer site that demonstrates how the widget drops onto an existing product detail page through a single script tag.
Read more
The retrieval layer uses eval-driven chunking with contextual enrichment. Each chunk is graded by an LLM-as-judge step during ingestion, so the vector store ends up populated with chunks that retrieve cleanly rather than chunks that look correct to a developer but fail under real query patterns. The control panel lets the team rerun evals as documentation changes.
Guardrails live in four layers: input validation, retrieval scoping, output validation, and post-response moderation. Together they map to the OWASP LLM Top 10. A prompt that tries to extract internal data, jailbreak the assistant, or wander outside the product's documentation gets rejected at the closest layer that can catch it.
LangSmith captures the full trace for every request: input, retrieval candidates, scoring, response, and guardrail decisions. Engineers debug from the trace; product owners look at the same trace to validate that an answer was on-policy.
Open standards used
OWASP LLM Top 10REST APIs
Tech stack
Next.js 15React 19Tailwind CSSClaudeLangSmithOWASP LLM Top 10 guardrailsEvals-driven RAG pipeline
Next step
Take the code
This one is published in full. Read it, run it, fork it. Nothing here is behind a conversation.
Request an intro
This one is the MACH Alliance’s own work. Ask about answering product questions safely, and the people who built it will answer.
Not ready to talk to anyone? MACH Insider is the Alliance mailing list. New solutions, program openings, and the research behind them, once a month.
Related solutions
Demo catalogs that took days of sourcing arrive in minutes, with imagery and a mapping into the target system.
Solution Studio Catalog Generator
Solution StudioOrder questions and in-policy returns are settled in the chat, and every case that needs judgment reaches a person with the evidence attached.
Solution Studio Order Support Agent
Solution Studio
Verification
Public code
The implementation is published in full, so anything on this page can be checked against the repository. Anything the repository does not show is not established here.
Status: Reference prototypeAdded May 26, 2026A listing is an editorial decision, not a MACH Alliance certification.
Built something worth showing?
Put your agentic solution in front of the enterprise leaders who use the hub to decide what to build next. Submissions are open to MACH Alliance Members, Collaborators, or Ambassadors.