Answer product questions from an up-to-date catalogue

Start with the operational bottleneck

Retail product information changes faster than a static FAQ. Stock, variants, prices, bundles, and shipping rules can shift while older help articles remain searchable. An assistant needs both approved explanatory material and current system data.

Start with routine product discovery questions. Separate descriptive answers from actions such as cancelling orders or authorising refunds. The first release can improve lookup and handoff without receiving permission to change an order.

Design the target workflow

Classify the question into a small set of intents. Retrieve approved product descriptions and policies. Query the catalogue for the exact variant, price, or availability needed, using a restricted read-only service account. Feed only those results to the response model.

Return a short answer with the product link and the relevant policy source. If the product cannot be identified or the requested advice exceeds the approved content, ask a clarification or hand off. Keep conversation context bounded, and record the source versions used. Cache stable descriptions separately from fast-changing inventory.

Connect to the existing systems

Connection	Information	Control
Commerce catalogue	Variants, price and availability	Use live reads for changing values
Approved help content	Product facts and policies	Assign content owners and revision dates
Support queue	Conversation and retrieved context	Transfer unresolved questions with useful evidence

Use stable identifiers and versioned inputs. A retry should resume the same piece of work, with the original permissions and evidence attached. Treat an upstream outage as a visible workflow state.

Fit the model to the task

Evaluate a small local model on product identification, policy grounding, and refusal to invent details. Use structured tool calls for catalogue access. Escalate only questions that need stronger reasoning and are permitted to leave the local environment. Medical, safety, or technical suitability claims require an approved content boundary.

Keep a clear human boundary

Keep identifiers and tool arguments validated. Text from a product description must not gain authority to invoke tools. The assistant cannot turn customer persuasion into a policy override. A support agent owns exceptions and consequential actions.

Measure the complete operating cost

Include retrieval, catalogue requests, maintenance, and handoff effort. A low token bill can conceal a costly custom integration. Track successful grounded answers and the quality of transferred conversations, not simply the fraction of chats that end.

Pilot checklist

• Limit the first release to read-only product questions.
• Test stale descriptions against current catalogue values.
• Validate product IDs and tool arguments.
• Measure grounded-answer quality and handoff usefulness.
• Treat refunds and cancellations as a separate reviewed capability.

Agree on acceptance thresholds before testing, then compare the proposed workflow with the same work handled through the current process. Expand only after the operation can explain both successful and failed cases.

Further implementation detail

• Choose the lowest-cost model that meets the task requirements
• Serve a local model to a team
• Ollama’s local API reference