What makes it trustworthy

Four guarantees, enforced by the architecture, not by a clever prompt.

F.01

Grounded answers only

Every reply cites a document from your knowledge base. If no source supports the answer, the message escalates. It does not guess.

F.02

Immutable escalation register

Every conversation that reaches a human is numbered, dated, and stored permanently per tenant. Nothing is overwritten.

F.03

Zero hallucination by design

The model can only answer from the passages you ingest. No ingested source, no answer: by construction, not by prompt.

F.04

Multi-tenant from day one

Each client gets a separate knowledge base, prompt set, WhatsApp number, and escalation queue. Isolated end to end.

The escalation register

Every hand-off becomes a permanent, numbered record. Auditable per tenant, never overwritten.

append-only
The escalation register
RefDateTenantStatusNote
R.00012026-02-04ferreteria-crescalatedBulk discount, no source
R.00022026-02-04clinica-esteansweredHours · hours.pdf p.2
R.00032026-02-05ferreteria-cransweredReturns · policy.pdf p.1
R.00042026-02-05estudio-legalescalatedCase status, needs human

Solid work behind the chat

Retrieval, orchestration, and hand-off are wired from proven parts, no black boxes.

  • WhatsApp Cloud API
  • Anthropic Claude
  • Voyage AI
  • pgvector
  • Chatwoot
  • n8n
  • Next.js
  • FastAPI

Put Antesala in front of your WhatsApp inbox.

Ingest your documents, connect a number, and let it answer what it can prove, escalate the rest.