AI Module¶

The AI module answers questions inside the platform. It reads the knowledge base and live platform data, and replies in the language of the question, with the sources it used.

Two audiences, one service:

Employees — backoffice and CRM. The assistant explains how the platform works and looks up customers, accounts, trades and margin state.
Clients — the trading terminal. The assistant explains the same mechanics in client language and looks up only the account of the current session.

The module also turns platform events into notifications: a margin call or a verification change becomes a text that a manager approves before it reaches the client.

What an interface gets¶

Capability	What it means for the interface
Asynchronous answers	`AIChatSendMessage` returns a `runId` immediately; the answer arrives as events or is polled with AIChatGetRun
Streaming	the answer arrives in fragments through module events as `module:event`
Sources	every answer carries the knowledge documents it was built from
Actions	a document may offer a navigation button; the model cannot invent one
Tool trace	which platform data the assistant read for this answer
Idempotency	a repeated `requestId` returns the existing run instead of paying for a second answer

What the module never does¶

It is read-only. There is no method that opens, closes or modifies a trade, moves money, or changes an account. The assistant reads and explains; every action stays with the person or with the existing platform methods.

It also does not give investment advice. That is a product decision built into the system prompt, not a limitation of the model.

Where to start¶

You are building	Read
Chat in the terminal, backoffice or CRM	Chat integration
The approval queue for generated notifications	Notifications
Administration screens (agents, prompts, knowledge, spend)	Administration
Anything over HTTP	REST API
Anything over TCP, plus streaming events	TCP API

Methods¶

Names in the table are the TCP command names; the REST column is the HTTP entry point of the same method.

Chat¶

Available to managers and to clients. A client session is limited to its own account.

Method	REST	Description
AIChatCreateConversation	`POST ai/chat/conversation`	Start a conversation
AIChatGetConversations	`GET ai/chat/conversations`	Conversation list of the current session
AIChatGetConversation	`GET ai/chat/conversation`	One conversation with a page of messages and the active run
AIChatUpdateConversation	`PUT ai/chat/conversation`	Rename or archive a conversation
AIChatDeleteConversation	`DELETE ai/chat/conversation`	Delete a conversation
AIChatSendMessage	`POST ai/chat/message`	Ask a question; returns a `runId` without waiting for the model
AIChatGetRun	`GET ai/chat/run`	State and result of a run
AIChatCancelRun	`POST ai/chat/run/cancel`	Stop a running answer
AIFeedbackAdd	`POST ai/chat/feedback`	Rate an answer
AIGetAgents	`GET ai/chat/agents`	Agents this session may talk to

Notifications¶

The queue of texts the assistant generated from platform events.

Method	REST	Description
MngGetAINotificationsByFilter	`GET ai/notifications`	Queue and journal, with the state of delivery channels
MngApproveAINotification	`POST ai/notification/approve`	Approve and deliver, optionally with an edited text
MngRejectAINotification	`POST ai/notification/reject`	Reject with a reason
MngRetryAINotification	`POST ai/notification/retry`	Deliver a notification that stayed pending

Agents and tools¶

Method	REST	Description
MngGetAIAgentsByFilter	`GET ai/agents`	Agent configurations
MngAddAIAgent	`POST ai/agent`	Create an agent
MngUpdateAIAgent	`PUT ai/agent`	Update an agent, its tools and its budgets
MngDeleteAIAgent	`DELETE ai/agent`	Soft-delete an agent
MngGetAIModels	`GET ai/models`	Models available to the provider, with prices
MngGetAITools	`GET ai/tools`	Tool catalogue and what is available on the bus right now

Prompt templates¶

Method	REST	Description
MngGetAIPromptTemplatesByFilter	`GET ai/prompts`	Templates of the brand
MngAddAIPromptTemplate	`POST ai/prompt`	Create a template
MngUpdateAIPromptTemplate	`PUT ai/prompt`	Update a template
MngDeleteAIPromptTemplate	`DELETE ai/prompt`	Soft-delete a template
MngTestAIPromptTemplate	`POST ai/prompt/test`	Preview substitution; with `run` also calls the model

Knowledge base¶

Method	REST	Description
MngGetAIKnowledgeStatus	`GET ai/knowledge/status`	Index state, per-brand breakdown, file errors
MngGetAIKnowledgeDocuments	`GET ai/knowledge/documents`	Loaded documents and their metadata
MngReloadAIKnowledge	`POST ai/knowledge/reload`	Rebuild the index without a restart
AIKnowledgeSearch	`POST ai/knowledge/search`	Search as the assistant sees it, with scores

Events and bindings¶

Method	REST	Description
MngGetAIWorkflowEvents	`GET ai/workflow/events`	Platform events an agent can subscribe to
MngGetAIWorkflowBindingsByFilter	`GET ai/workflow/bindings`	Event-to-agent bindings
MngAddAIWorkflowBinding	`POST ai/workflow/binding`	Bind an agent to an event
MngUpdateAIWorkflowBinding	`PUT ai/workflow/binding`	Update a binding
MngDeleteAIWorkflowBinding	`DELETE ai/workflow/binding`	Soft-delete a binding
MngTestAIWorkflowBinding	`POST ai/workflow/test`	Dry run: what would happen, and why not

Spend, audit and service¶

Method	REST	Description
MngGetAIUsageByFilter	`GET ai/usage`	Requests, tokens, cost, errors and latency
MngGetAIRunsByFilter	`GET ai/runs`	Run journal with filters
MngGetAIRunDetails	`GET ai/run`	One run with its prompt, sources and tool calls
MngGetAIFeedbackByFilter	`GET ai/feedback`	Ratings left by users
AIPing	`GET ai/ping`	Cheap liveness probe with a problem list
AIHealth	`GET ai/health`	Full state: database, knowledge, model, tools, platform commands

Entities¶

Entity	Scope	Purpose
`agents`	brand	Model, system prompt, allowed tools, daily budgets, and three independent modes: `modeChat`, `modeClient`, `modeWorkflow`
`conversations`	session	Chat history of one manager or one account
`messages`	conversation	Questions and answers with tokens, cost and sources
`runs`	conversation	One generation: state, tool rounds, spend, error
`promptTemplates`	brand	Text with placeholders for notifications and service prompts
`workflowBindings`	brand	Platform event to agent, with filters, cooldown and delivery channels
`notifications`	brand	Generated texts, their approval state and delivery result
`toolCalls`	run	Audit of every platform call the assistant made
`usageDaily`	brand	Daily aggregate of requests, tokens, cost and latency

The knowledge base itself is not an entity: documents are files in the module repository, reviewed like code and reloaded into memory by MngReloadAIKnowledge.

How answers are built¶

flowchart LR
    q["Question"] --> k["Knowledge search<br/>(language + audience)"]
    k --> p["Prompt:<br/>rules, agent, knowledge, history"]
    p --> m["Model"]
    m -->|needs data| t["Read-only platform tools"]
    t --> m
    m --> a["Answer + sources + actions"]

Three rules hold that picture together, and they explain most of what an interface sees:

The model never decides access. Brand, account and session type come from __access; a tool call for another account is refused by the platform, not by the prompt.
Figures come from tools, not from memory. If a tool fails, the assistant says so instead of inventing a number.
Knowledge is data, not instructions. Text inside a document cannot change the assistant's behaviour.