Answer quality & grounding

The widget answering "without making things up, with citations" isn't a single setting — it's the result of three complementary mechanisms: clean chunking, accurate search and strict grounding.

Smart chunking & dedup

Crawled pages are split into meaningful chunks. Boilerplate text that repeats on every page — menus, footers, cookie banners — is stripped out (dedup), so this noise doesn't leak into answers and search focuses on real content.

Hybrid search

When a question comes in, two searches run together:

• Keyword — exact term matching ("shipping").
• Semantic (vector) — meaning similarity; "delivery time" also finds "shipping" content (pgvector).

The combination catches both the customer who knows the term and the one who asks in different words.

Strict-KB grounding

The retrieved chunks are the only source for the answer. The model generates its reply only from those chunks and shows the source. If there aren't enough relevant chunks (below threshold), the widget doesn't make things up — it says "I don't have information on that" or, per your chosen behavior, redirects/escalates.