prod-20260830-weekly2026-08-30
Smarter product search, an image-aware knowledge base, and more reliable chat
This week focused on answer quality — multi-turn product search that keeps context, a knowledge base that understands images, and a fix for the recurring chat error message.
New Features
- Multi-turn product search keeps context across follow-ups — answers to questions like "anything else?" or "no preference" continue the established search instead of starting over, and catalog aliases (brand, product type, SKU nicknames) now apply across the whole conversation.
- Two optional search enhancements are available on request — an AI reranker that improves ordering of search results, and automatic hiding of known out-of-stock products from browse and recommend results (shoppers asking about a specific URL still get an honest "currently unavailable").
- The knowledge base is now image-aware — figures and charts inside PDF, DOCX, and PPTX sources are extracted with vision captions and rendered inline in chat answers. Re-index files from the dashboard and browse extracted images in the edit dialog.
- Knowledge base sources accept optional indexing hints (product name, SKU, brand, …) so deep document chunks stay searchable by the identifiers that matter to your catalog.
- New AI model options on storefront chat — GLM 5.3 Flash and Qwen 3.8 Flash are now selectable for all organizations.
- Human handover is more reliable — handover requests are saved before notifications are sent, shoppers switching back to the AI assistant close the handover cleanly, and repeated "agent please" asks no longer reset the pickup clock.
- New per-conversation KPI panel in the conversations Activity tab, with admin-only exclusions so internal or test conversations can be kept out of SLA reporting.
Fixes
- Fixed the recurring "Sorry, I am having trouble responding right now" chat error. The root cause was stalled AI provider streams; chat turns now time out quickly and retry — and fall back to an alternate provider if needed — instead of failing.
- Further guardrail false positives fixed — benign non-English shopper questions are no longer misclassified as data-exfiltration attempts.