Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.asapp.com/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

Unlike traditional bots with predefined flows, GenerativeAgent uses natural language processing to understand and respond to a wide range of customer queries and issues. GA works in Tasks. Each Task is a logical agent: it carries its own knowledge, tools, and instructions, and runs through an agentic loop to help the customer.
A GenerativeAgent Task running its agentic loop: drawing on the Knowledge Base, calling Functions to act in your systems, and invoking HILA when human input is needed
Within a given Task, your Knowledge Base is the source of truth. Tools are exposed as Functions, which GA uses to call your APIs, ask a human for help via Human-in-the-Loop (HILA), or escalate to a human agent via a Handoff.

​
What happens each turn

At the start of each turn, GenerativeAgent is inside a Task. GA assembles a fresh context window:
  • Limited set of the most relevant KB topics relevant to the conversation
  • The active Task’s instructions
  • The Functions available to call as tools
  • All current context variables
  • The conversation history
  • Any actions GA has already taken, including tool call results
From that context, GA picks one of four actions:
  • Answer from an topic: respond using the surfaced KB content.
  • Trigger a Handoff: route the conversation to a human agent queue.
  • Switch Task: move into a different Task when the conversation direction changes.
  • Continue the current Task’s procedure: follow the next step of the Task’s instructions. This can include calling a Function, asking the customer for more information, requesting Human-in-the-Loop (HILA) assistance, or anything else the procedure dictates.

​
Architecture

GenerativeAgent runs on different models depending on the channel. Voice (Talker-Reasoner): Two models split the work.
  • The Talker is the customer-facing model. Its default stance is to follow the Reasoner’s instructions, which in practice usually means progressing the active Task’s procedure. The Talker acts on its own in three cases:
    • Answering from a Topic
    • Initiating a Handoff
    • Switching Task
  • The Reasoner holds the Task’s instructions and runs the full agentic loop: calling Functions, evaluating results, planning the next step. It instructs the Talker on what to do next. In practice that’s often something like what to say to the customer, what information to ask for, or what to clarify before continuing.
Chat (single model): A single model handles the per-turn decision and the Task procedure together in one pass.

​
How deployments grow

Most customers start with GenerativeAgent answering from the knowledge base and routing callers to human queues via Handoffs. This covers the intake work a traditional IVR would handle. No Task configuration is required for this first stage. As teams identify which conversations could be automated, they layer in Tasks. A Task is where procedures, API calls, and multi-step logic live. Handoffs used during KB routing are equally available inside Tasks for human escalation. See Using GenerativeAgent to Route Conversations for the first-deployment walkthrough.

​
Core components

ComponentWhat it does
TopicsKB articles. Can range from a targeted snippet to a broader article covering multiple related intents. GenerativeAgent surfaces relevant chunks per turn; Handoff and Task references attached at the article level are always included regardless of chunk size.
HandoffsAvailability-aware transfers to human agent queues. Usable from Topics and inside Tasks. The Handoff resolves availability before GenerativeAgent is invoked. GenerativeAgent only sees the instructions for the current availability state.
TasksDeterministic workflows. Define procedures, connect Functions, and configure per-step logic. This is where API calls and complex automation live.
FunctionsConnect Tasks to your APIs (fetch data, perform actions), store conversation variables, or signal a transfer to an external system.
BricksReusable components shared across Tasks: common functions, prompt templates, conversation flows.
Knowledge BaseWhere Topics live. Import content from URLs, Zendesk, PDFs, or via API.
SettingsConfigure how the agent behaves across all Tasks.
  • Company Overview: name, welcome message, description, and live agent name
  • Input Safety Categories: categories GA should treat as safe so they aren’t filtered out (Scope and safety tuning)
  • In-Scope Categories: categories GA is allowed to discuss, with examples to anchor the boundary (Scope and safety tuning)
  • Sensitive API Keys: names of API key fields (e.g. account_number, ssn) that must never be shared with customers
  • Default Talker Prompt Override: use a company-wide Talker prompt to guide responses instead of per-Task prompts
  • Voice Configuration: preview and select the voice GA uses on phone channels
  • Language Support: language for chat and voice (Language support)
  • Channel-Specific Configuration: per-channel default model, with optional per-prompter LLM overrides

​
Environments

GenerativeAgent operates across three environments to support safe testing and progressive deployment:
  • Draft: Configure and test components before deployment.
  • Sandbox: Staging environment to validate behavior with real APIs.
  • Production: Live environment serving real customer conversations.
Configure in Draft, validate in Sandbox, and promote to Production when ready. For detailed deployment procedures, see Deploying to GenerativeAgent.