Run the right assistant for each customer path.
Kelvia agents are tenant-owned assistant profiles, entry points, and routing policies. Use them to separate sales, support, bookings, languages, departments, and channels without creating a public marketplace or custom code runtime.
Profiles
Role, tone, behavior, tools, language policy, and channel scope
Entry points
Webchat choices that start with the right profile and intent
Routing policies
Channel, intent, department, business-hours mode, and priority
Release controls
Versions, simulations, eval suites, readiness checks, and audit events
Routing preview
How the control plane chooses a profile
Channel
webchat
Detected intent
booking
Department
clinic reception
Hours mode
after_hours
Selected assistant profile
Bookings - evening desk
selected_by: routing_policy
Different assistants, one operating system
Profiles are practical operating profiles. They decide what the assistant says, which capabilities it can use, which channels it can serve, and where it hands off.
Behavior and tone
Choose receptionist, sales, support, booking, billing, after-hours, triage, reservations, or intake behavior with a tone preset.
Language and channel scope
Scope a profile to phone, webchat, WhatsApp, Telegram, Instagram, or all channels, then apply language policy where needed.
Capabilities and transfer policy
Limit allowed capabilities and transfer behavior so each profile only uses the tools that fit its job.
Route by path, not by guesswork
Routing policies select active profiles by channel, detected intent, department, business-hours mode, priority, and conditions. Webchat entry points can also start a conversation with a chosen profile and initial intent.
Give each profile the business sources it should retrieve
Add websites, manual notes, supported text files, services, and FAQs. Group sources into datasets, keep retrieval metadata visible, and leave a source shared or bind it to an assistant profile and channel.
Websites and structured sources
Ingest one page, a sitemap scan, or a plan-gated deep scan. Services and FAQs synchronize from their owning product screens.
Notes and supported text files
Add manual notes or editable text from .txt, .md, .csv, and .json files. Knowledge does not perform OCR; document OCR belongs to Intake.
Datasets and retrieval metadata
Group related sources and label visibility, purpose, and trust so retrieval can filter and rerank within the tenant. These labels do not replace product authorization.
Profile and channel binding
Unbound sources stay shared. A binding can scope retrieval to one profile, one channel, or both, then the team can test search and reprocess stale sources.
Kelvia retrieves source text as business data, not instructions or model training. Embeddings may use the configured provider; sources, chunks, bindings, and search remain tenant-scoped.
Preview before changing production behavior
Versions and rollback
Create drafts, publish versions, archive old versions, and roll back when a profile should return to a known state.
Simulations and evals
Run profile simulations and evaluation suites before changing how the assistant behaves for customers.
Readiness and audit
Knowledge readiness checks and reviewable audit events help teams see what changed and why.
Where it fits
Separate the work without splitting the platform.
Use assistant profiles and routing policies when one generic receptionist is no longer enough for the operation.