Kavesta

Back to the exhibition

Build assistants that read your content, then answer from it.

Node types

Retrieval stages

To embed

2 lines

The problem

The question arrives close to midnight.

Someone is ready to buy. Nobody is awake to answer. The message sits for nine hours, and by morning they have booked with whoever answered first.

What happens today

Both of the usual answers cost you.

Either it waits until morning and they book elsewhere, or somebody drops their own work to answer it. One loses the customer. The other loses the day — your best people spending it on the same forty questions instead of the ones that need them.

The bots list in a workspace: six assistants, each with its own path, flow mode and channel. — The workspaceEach one a separate assistant, with its own flows and conversations.

The conversations view: three live chats, one open showing a customer asking how to reset their password and the bot answering, with a Take Over Conversation control in the header. — Live chatsThe bot is handling this one. Take over, and it hands you the whole conversation.

The answer

Kavesta takes charge from here.

It understands the question, asks what it needs, and answers from your own documents — your prices, your policies, your hours. Never from a guess. Your team can continue to take rest.

One turn, end to end

What happens between the question and the answer

EmbedA two-line script mounts the widget. The origin is checked against the sites you allow.
RouteA master router reads the message and picks the right node anywhere in the bot.
RetrieveA retrieval sub-graph runs — query rewriting, vector search, filtering, guardrails.
AnswerThe reply carries prose and structure: tables, dates, slots, choices, ratings.
Hand overWhen a person is needed, the whole conversation goes with the handover.

Retrieval

Ten stages, each one replaceable

Query rewriting — HyDE, step-back, multi-query
Router
Cache
Loader
Vector search
Filter
Web search
Reranker
Summariser
Guardrail

The studio

Twelve node types on one graph

llm
rag
websearch
api
condition
function
delay
step
calendar-availability
calendar-book
note
end

Show your work

You can see what it saw.

Every retrieval leaves a per-stage trace, readable in the playground and the logs. Skipped, errored and not-yet-implemented are three different things, and the trace says which. Most demos are a chat window; this one opens the machinery.

Tenancy

Isolation is structural, not a filter

ClientA person
OrganizationThe isolation boundary
WorkspaceA region — EU, US
BotOne assistant
FlowOne graph

Install

Paste this before </body>

<script
  src="https://widget.quana.ai/embed/v1/embed.js"
  data-client-id="your-client-id"
  defer
></script>

Built with

Application: Next.js 16 · React 19 · TypeScript
Data: MongoDB Atlas — documents and vector search
Default model: gpt-5.4-mini · text-embedding-3-small
Bring your own store: Qdrant · Chroma
Loader: Vanilla JS, no build step, lazy iframe

One surface

Built once. Answering everywhere.

The same journeys serve your website, WhatsApp and the phone. Every conversation is logged, so for the first time you can read what customers are actually asking for — and change the answer the same afternoon.

The knowledge area of a workspace, where the documents an assistant answers from are kept and indexed. — KnowledgeThe documents it answers from.

Client to Organization to Workspace to Bot to Flow, each one-to-many, with members, audit, quota and knowledge bases hanging off the organization and channels and conversations off the bot. — TenancyOne tenant, many workspaces, many bots. Access is a grant, and the most specific object wins.

View project