Assistant provides two different operational views. Logs / Analytics helps improve demand and usage; Technical diagnostics captures controlled evidence for investigating a specific conversation.

Dashboard signals

NAI Assistant → Dashboard shows:

• indexed versus total products and pending work;
• queries this week and trend versus the previous period;
• chat-to-sale percentage and attributed orders;
• product-index progress;
• top searches with final outcome, answer preview, products, source count, language and time.

An “Answered” state means the response pipeline completed; it does not prove the answer was commercially or editorially good. Open repeated queries and verify the source and cards.

Logs / Analytics

Open Settings → Logs / Analytics to control:

• analytics and demand logging;
• answer previews;
• automatic retention;
• API usage and estimated cost;
• response caching and cache lifetime.

The operational sections distinguish resolved queries, resolved product queries, dismissed queries, zero-product demand, low-confidence/unanswered questions and trending terms.

Use the views as a workflow:

1. group repeated queries by intent;
2. reproduce the exact phrase and language;
3. identify whether the gap is catalogue data, site content, indexing, routing or wording;
4. correct the authoritative source first;
5. rebuild/refresh the relevant index and clear cache;
6. add a narrow training link only when ambiguity remains;
7. retest in a fresh conversation;
8. mark the item resolved or dismiss clear noise.

Product training should point to real products and must not override live price, stock or variation data.

Conversation logging

Open NAI Assistant → Technical diagnostics. Logging is separately configurable with:

• enabled state;
• Standard or Diagnostic level;
• retention from 1 to 365 days;
• email/phone masking;
• page-URL storage.

Standard is the normal production level. Diagnostic captures deeper intent, candidate and pipeline evidence and should be enabled only for a bounded investigation because it is more detailed.

Page URLs are stored without query strings in the hardened pipeline. Keep PII masking enabled unless a documented, temporary test requires otherwise.

Review and export

The conversation list shows time, conversation ID, language, page, message count and risk. Open a conversation to inspect the user turn, final answer, products/sources and warnings.

Use Download problematic logs only before sending a large all-log export. Review the file locally and remove customer content not needed for support. Never include API keys, licence keys, passwords, payment details or unrelated personal data.

Deleting logs is permanent. Confirm retention obligations and preserve only the minimal support case evidence before using Delete all logs.

Interpret common patterns

• Zero products, strong product intent: catalogue/index/filter gap.
• Unanswered policy question: missing or weak authoritative source.
• Correct text, wrong visible source: role/relevance or caching problem.
• Repeated clarification: inconsistent catalogue facets or overly broad request.
• High usage with identical questions: cache scope/lifetime or automation traffic.
• Clicks without cart actions: product fit, card action or WooCommerce-session issue.
• Chat sessions without attribution: event/session/cookie or order-completion path needs inspection.

Troubleshooting

• No analytics: confirm Analytics enabled and test as a logged-out visitor.
• No conversation logs: logging has its own switch under Technical diagnostics.
• Old answer preview: clear the Assistant response cache and begin a new conversation.
• Support export is too large: filter to problematic conversations and shorten the investigation window.
• PII appears despite masking: do not transmit the export; reduce scope, remove it and inspect custom fields/diagnostics before continuing.
• Resolved item returns: confirm the underlying source/index changed, not just the dashboard status.

Support evidence checklist

Include product/version, WordPress/PHP/WooCommerce versions, timestamp and timezone, exact safe reproduction steps, expected versus actual outcome, language, affected page without sensitive query parameters, and the smallest redacted diagnostic export that shows the first error.