Assistant combines the live WooCommerce catalogue with a local semantic index. The catalogue remains authoritative for product identity, price, stock, variations and cart actions; embeddings help match how shoppers describe a need.

Requirements and scope

WooCommerce must be active for product tools. In Settings → Products & Cart, enable WooCommerce integration. Assistant can still operate as a site-content assistant when WooCommerce is absent or the integration is disabled.

Eligible products must be real, published WooCommerce products. Hidden, unavailable or conflicting catalogue data should be corrected in WooCommerce rather than compensated for with an instruction prompt.

Build the index

1. Configure and test Gemini or OpenAI as the embedding provider.
2. Open NAI Assistant → Dashboard.
3. Compare Products indexed with the live total.
4. Choose Reindex all after the first setup, an embedding provider/model change or a product-knowledge schema update.
5. Keep background processing available through Action Scheduler or WordPress cron until the state finishes.

Assistant automatically reacts to product saves, variation saves, stock changes and product deletion. It only re-embeds changed products during normal maintenance.

Decide what product knowledge contains

Under Settings → Products & Cart → Product knowledge:

• Enable full product content when long descriptions contain facts the assistant should use.
• Include variation data when shoppers need size, colour, configuration, variation price or variation stock.
• Add only safe, relevant custom-field names. A custom field containing internal margin, supplier notes or personal data must not enter customer-facing knowledge.

Index text can include the product title, short description, categories, tags, SKU, relevant attributes and configured knowledge. Current price and availability are still resolved from live WooCommerce data when the answer is produced.

Configure result volume

• Products shown at once controls the first broad browsing batch.
• Maximum products shown in chat caps the conversation total and cannot be lower than one batch.
• Products retrieved (N) under Advanced controls semantic candidates, up to 20.

Use a small visible batch for mobile usability and a larger retrieval pool for ranking. Do not display every retrieved candidate.

Use grounded filters

Assistant understands natural-language price, colour, size, brand, stock, category and sort requests and converts them to catalogue constraints. Accuracy depends on consistent WooCommerce data.

Before launch, standardise:

• category and brand names;
• global attributes and variation values;
• numeric prices and sale prices;
• stock status and backorders;
• SKU and model identifiers;
• measurement units and custom fields used for comparisons.

An exact model number or SKU should be tested separately from a broad need such as “warm waterproof jacket for commuting.”

Control weak and ambiguous matches

Use Product result accuracy and Advanced settings deliberately:

• Keep lexical evidence enabled for weaker semantic matches so technical terms and identifiers cannot drift to unrelated products.
• Keep unsupported-answer protection on Strict for a new site.
• Enable clarifying questions so a vague request can produce one short follow-up with real catalogue choices.
• Set the relevance threshold near the default 0.6 and change it only after inspecting diagnostics.
• Use synonyms for genuinely equivalent customer vocabulary, one group per concept.
• Add excluded terms when a recurring word attracts the wrong catalogue area.

Lowering the threshold cannot repair missing products, inconsistent attributes or an index built with the wrong embedding model.

Accessories, alternatives and explicit rules

When Exclude accessories when the shopper asks for a main product is enabled, Assistant uses configured multilingual accessory terms and selected accessory categories. A customer who explicitly asks for an accessory can still receive accessories.

Keep Allow clearly labelled alternatives off until zero-result behaviour has been tested. When enabled, alternatives remain capped and visibly labelled; they do not become exact matches.

Use Query-to-product rules for important ambiguous phrases. A rule can define trigger terms, required terms, excluded terms and a required WooCommerce category. Rules should correct known catalogue language, not create store-wide hard-coded behaviour for every query.

Zero-price products

Choose one policy for a numeric price of zero:

• show the formatted zero price;
• show custom text such as “Request a quote”;
• hide the price.

Then choose whether the action opens the product page or allows add to cart. Configure per-language replacement text under Languages & Localization. Test both the product card and the text answer so a hidden or replaced zero price does not reappear as “0” in the explanation.

Search test plan

Run Settings → Advanced → Search test and a real private-session chat for:

1. exact product name;
2. SKU/model identifier;
3. category browse;
4. use-case description;
5. price ceiling and cheapest sort;
6. colour/size/brand filter;
7. unavailable variation;
8. accessory exclusion and explicit accessory request;
9. zero-price product;
10. deliberate no-match query.

For each result, verify title, URL, image, current price, stock, variation and action against WooCommerce.

Troubleshooting

• Index count is lower than product count: wait for background work, inspect the first failed item and confirm product publication state.
• Exact product is missing: inspect SKU/title/category/attribute evidence and the active integration setting.
• Results come from the wrong category: standardise taxonomy and add a narrow query rule only after diagnostics confirm the ambiguity.
• Old price or stock appears: clear page/object caches and confirm the card resolves live WooCommerce data.
• Changing embeddings broke search: complete a full product and Knowledge Base rebuild.
• Too many weak alternatives: disable alternatives, raise evidence quality and correct catalogue vocabulary.

Verification checklist

• All intended products are indexed with the current embedding signature.
• Live price, stock and variations agree with WooCommerce.
• Exact identifiers and natural-language needs both work.
• Vague queries clarify instead of dumping unrelated products.
• Alternatives, accessories and zero-price behaviour match store policy.
• Deliberate no-match requests do not invent products.