Z-Assist: a docs-grounded chat that refuses to hallucinate
The portal's assistant answers from the official docs.z.ai and zcode.z.ai text only — with a machine-audited anti-hallucination contract, a verification layer and staff-approved self-learning.
Support chats that invent pricing tiers are worse than no chat at all. Z-Assist — the chat in this portal's header and at /agent — was built around one rule: answer only from material you can point at.
How grounding works
- A curated knowledge base plus hourly-fetched official docs (docs.z.ai and all 29 zcode.z.ai wiki pages) form the corpus. The chat and the pages you read share one source of truth.
- An anti-hallucination contract is machine-audited: every factual claim — price, date, quota, model name, URL — must come from the corpus, a staff-approved learned answer, or the user's own message.
- A verification layer re-checks replies before they ship: URL allowlists, corpus cross-checks, verified-badge gating.
- A humanizer pass keeps answers readable — short paragraphs, honest limits, no docs-dump wall of text.
The parts that make it feel alive
Answers stream live with a multi-model fallback ladder, comparisons arrive as tables, and quota math runs as interactive tools you can tweak in the bubble. Accounts are a one-glance token (ZA-…): users come to get help, staff sign in to help users — reviewing conversations and approving learned answers back into the knowledge base.
The honest limit: Z-Assist knows the docs, not your account. When something is account-specific, it says so and links the right official page instead of guessing.