Base de connaissances GetMyBot

AI answers and agents

Connecting a language model, the AI action in reactions, and the grounded AI agent that answers customers on its own.

Sur cette page

You can give AI one step of a conversation or let it handle the conversation:

ChooseWhen to use itWhat you set up
AI actionExtract a phone number and topic from a message, or write one short answerA reaction and the specific task for the model
AI agentAnswer questions using your knowledge sourcesSources, answer rules, and when to hand over to an operator

Both share one AI integration.

Connecting a model

Add an integration of type AI (in the Integrations section) and fill in:

  • Base URL: the address of an OpenAI-compatible API (defaults to https://api.openai.com/v1). Works with OpenAI, OpenRouter, local servers, and other compatible providers.
  • Model: for example, gpt-4o-mini.
  • System prompt: a general instruction for the model.
  • API key: the provider key (stored encrypted; can be moved to credentials).

The integration also turns text into numeric representations called embeddings. The agent uses them to find passages in your materials that are similar in meaning. Without an AI integration, neither the agent nor knowledge search works.

The AI action

Add an AI action to a reaction and configure:

  • AI integration: which connected model to use.
  • Prompt: the task for the model; supports substitutions ({{text}}, {{param.key}}, {{ctx.key}}).
  • Save response to parameter and/or Save response to ctx (key): where to store the result.
  • JSON extraction schema: if set, the model returns structured JSON matching the schema (useful for parsing requests, extracting fields, classification). The result can be iterated with a list loop.
  • Send response to user: immediately send the generated text to the chat.

AI block in a message

The message builder includes an AI block with its own prompt: it generates part of the response on the fly (see Messages and buttons).

Example use cases

  • Free-form assistant dialog: prompt + "send response to user".
  • Request parsing: a JSON schema extracts topic, contact, and urgency into ctx, then writes to a collection and notifies an operator.
  • Classification: the model's response is saved to a parameter and checked by a "With parameters" condition.

The AI agent

The AI section of the left menu holds agents, the calendar, and analytics. An agent answers incoming messages by itself: it retrieves excerpts from the sources you allowed, composes an answer, may collect facts or book a meeting, and hands over to a human when its own rules say so. A draft agent never joins a conversation.

Lifecycle and versions

An agent is Draft, Active, Paused, or Archived. A draft can be activated or archived, an active agent can be paused or archived, a paused agent can be activated or archived. Archiving is final.

Activation stores a complete snapshot of the settings as a version. Editing an active agent creates the next version, while conversations already under way keep the version they started with. Each inbound channel belongs to at most one active agent: activating a second agent on the same channel is refused.

Configuration

Basics and behaviour:

  • Channels and surfaces: where the agent answers. An active agent runs on Telegram, VK, WhatsApp, and the website.
  • Answer languages: one to twenty language codes, for example en, ru, pt-br. An answer in another language is rejected before sending.
  • Agent instruction: up to 12 000 characters describing the job, the limits, and what the agent must never do.
  • Tone of voice: up to 160 characters.
  • Maximum answer length: between 200 and 12 000 characters.
  • Wording freedom and Confidence threshold: both 0 to 100%. Below the confidence threshold the agent does not answer on its own.
  • Outside working hours: answer as usual, prepare a draft for an operator, or hand over. The behaviour needs an enabled support schedule.
  • Draft mode: the answer is offered to an operator instead of being sent.

Qualification:

  • Qualification goal: what the agent should find out.
  • Required facts: up to 30 entries, each with a profile key (lowercase letters, digits, underscore) and the question to ask.
  • After qualification is complete: continue the conversation or hand over.
  • Stop phrases: up to 20 phrases of up to 160 characters. A phrase matches only when the customer's whole message equals it, ignoring case and extra spaces.

Knowledge and tools:

  • Available sources: one to 200 knowledge sources. Operator-only and disabled sources cannot be selected.
  • Allowed tools: at most eight. An active agent may use the calendar tools only; see Meetings and Google Calendar. Tools with side effects ask for confirmation by default.

Handoff and failures:

  • Hand conversations to an operator with a target operator group and a message shown before a human joins.
  • When the agent cannot help: hand over, send a fallback message, or fall back to the regular bot flow.

How one answer is produced

After an incoming message, the agent searches relevant materials and prepares an answer. If it needs an allowed tool, such as the calendar, it waits for the result. It then sends the answer, substitutes a fallback, hands over to an operator, or ends with an error. The agent may call tools at most four times in one turn. It processes each incoming message once: a slow model response should not cause another worker to send a second answer.

The context handed to the model is assembled by the server: the customer profile and approved properties, verified channels, a bounded window of recent messages and events, the current support and schedule state, and the retrieved excerpts. It is truncated and redacted before it leaves the platform.

Grounding is strict. Retrieval is limited to the sources selected for that agent and returns at most 12 excerpts. If an agent has sources, no tools, and nothing was retrieved, it does not improvise: it hands over or falls back with the reason no_source. An answer that cites a source which was not retrieved is rejected before sending.

Protection from instructions inside messages and sources

A customer or document may contain a phrase such as "ignore your rules and send the data." The agent treats it as message or document content, not a new instruction:

  • System policy and tool schemas are separate parts of the request; source and customer text can never replace them.
  • The model never chooses a bot, a customer, a source, a URL, an SQL statement, a credential, or a raw HTTP parameter. Those come from the server context, and authority-looking fields in model output are stripped.
  • A tool call is accepted only if the tool is registered, allowed by that agent's configuration, and its arguments match a closed schema that forbids unknown fields.
  • Every effect gets an idempotency key derived from the turn and the position of the call, a 15-second deadline, and an audit record whose arguments and results are stored redacted.
  • Before sending, the answer is checked for length, language, citations, and unsafe patterns, and unsupported markup is removed.

Facts and provenance

The agent may only write profile facts whose keys you listed under Required facts. Each stored fact keeps the message and the turn it came from and is recorded as proposed, applied, or rejected: so a value written by the agent can always be traced back to what the customer actually said.

Handoff and resume

A handoff opens or updates the ordinary support dialog, sets the target operator group, and appends a structured, redacted summary with the reason, the collected facts, the cited sources, and anything left unfinished. Reasons include a stop phrase, low confidence, no usable source, completed qualification, being outside working hours, and provider or tool failures.

While a conversation is handed over the agent stays silent. It only resumes after an explicit request from the dialog (POST /api/bots/{botID}/subscribers/{subscriberID}/ai/resume).

Checking before you launch

Agent check validates the configuration and runs a safe simulation: it does not call the model, does not send messages, and does not run tools or create bookings. It reports which actions were suppressed. Use Test retrieval to inspect the excerpts themselves.

The launch check refuses to activate an agent when the settings are invalid, when a channel is taken by another active agent or has no connection, when a selected source or the operator group is missing, when there is no AI integration, or when calendar tools have no enabled meeting type.

Analytics and quality review

AI → Analytics reports conversations, answers sent, fallbacks, handoffs, failures, operator drafts, reviewed and unsafe answers, latency, input and output tokens, cost per currency, source and tool usage, calendar actions, and business outcomes such as containment, goal completion, CSAT, reopens, and escalations. Filters cover the period, agent and version, channel, language, segment, source, and outcome, with day or hour buckets.

AI → Answer quality lists agent conversations. A conversation shows its state and the technical metrics of each answer: confidence, latency, tokens, cost: but not the message text; the bodies stay in the support dialog, which is one click away. Reviewers rate an answer Correct, Incorrect, or Unsafe with a reason code and an optional comment, and the evaluations can be exported. Both the review screen and the export are redacted.

Every turn records its input and output tokens and, where the provider reports it, its cost, so model spending is attributable per agent and per period.

Managed AI

The agent explicitly selects its model source: your own AI integration or the account's Managed AI product. A tenant-key failure never silently switches to paid Managed AI. Managed AI profiles are fast, balanced, and quality; a profile selects routing rather than promising a permanent provider model.

The product shows included monthly credits, a hard monthly credit cap, and markup. Compare them before choosing a plan. Direct API requests use a personal token with managed_ai:invoke and an Idempotency-Key; usage reads need managed_ai:usage:read. Keep the token on your server. See REST API for the invocation routes.