Use cases Chatbot widget
Customer-facing information chatbotA support chatbot that answers from your content only.
Turn repeat customer questions into instant, cited answers — a chatbot that reads only the files and pages you approve, running on your own infrastructure.
Built on a RAG knowledge base with input and output firewalls, so every answer is traceable to a source and every question is checked before it reaches the model.
- Answers only from approved content
- Dual input / output firewalls
- Web widget or OpenAI-compatible API
- On-premise or SaaS cloud
A chatbot built from content, not scripts.
There is no separate set of "scenarios" to write. The chatbot's answers come from a knowledge base you assemble and organise in aamp — configured once in a panel, no code required.
Any OpenAI-compatible provider, including local models.
Swap the model without touching the widget. The same configuration serves a frontier cloud model or a fully local one, depending on your policy and infrastructure.
Role, tone and topic scope, all set once.
The system prompt sets the agent's role, topic boundaries, tone and language, plus a rule to decline when the knowledge base has no answer. Conversation limits cap context window size, when long chats get summarised, and how many search rounds one answer can take.
Only what you explicitly attach.
The agent uses only the knowledge bases and helpers assigned to it — nothing else is reachable. A short prompt fragment on each base tells the model when to use it, which is enough to separate topics without writing rules.
Separate policies for what comes in and what goes out.
An inbound firewall screens each question against your policy before it reaches the model, with a decline message you write. An outbound firewall checks the answer before it is sent to the browser. Both are assigned per chat and per publication channel.
A web widget, or your own interface over the same API.
The web widget sets title, welcome message, button label, size and theme, with a ready address, test page and embed snippet. Its public server reaches aamp over an outbound tunnel, so your host is never exposed to the internet. Prefer your own interface — an app, a helpdesk, Teams? Use the OpenAI-compatible chat completions endpoint instead.
Six steps, every one logged.
Question in the widget
The user types into the chat window on your site.
Inbound firewall
The question is checked against your policy. A blocked question returns a message you set.
Agent searches the knowledge base
Search and document-read tools run only against the bases connected to this agent.
Answer with its source
The model streams a reply grounded in the retrieved passages and the system prompt.
Outbound firewall
The answer is checked before it reaches the browser.
Logged
The turn, model call, token count, tools used and firewall decisions are all recorded.
Only the widget is ever public.
Compliant with EU rules such as GDPR, the AI Act and the DSA. Control plane, secrets and knowledge stay on your server — only the widget component is publicly exposed.
The widget server never holds knowledge or secrets.
The aamp host — panel, agent, knowledge base, secrets, logs and the indexing helper VMs — stays in your internal zone. Only a small widget server sits in a DMZ or edge cloud, connected to aamp over an outbound tunnel with no access to knowledge or secrets.
Administrator, Operator, User — scoped per agent.
Access is granted per agent and per knowledge base. Keys are encrypted and rotate without ever revealing their values. The platform blocks a public widget outright when it is connected to non-public knowledge or risky tools.
Content without an assignment is invisible to the agent.
Search runs server-side and only against bases connected to the asking agent. Every query against the knowledge base is logged.
Conversations, costs and policy blocks, all visible in real time.
Mission Control reports conversation counts, errors, policy blocks, P95 response time and token usage. Immutable logs cover conversations, knowledge-base queries and firewall decisions.
Where repeat questions cost real support time.
Technical support
Answer product and troubleshooting questions from manuals, FAQs and known-issue notes, 24/7.
E-commerce returns
Handle returns, shipping and policy questions from your own terms and service pages.
Domain and hosting portals
Guide customers through transfers and account changes using your own documentation.
Internal helpdesk
Point employees at HR policies, IT procedures and internal wikis without exposing them externally.
Regulated public sector and finance
Publish a public-facing widget scoped strictly to approved public content, under GDPR, AI Act and DSA constraints.