AI Responses

7 min

The AI Responses processor sends messages to an AI model and returns the model's response. It keeps a rolling conversation history across calls for multi-turn dialogue. Beyond a single response, you can give it tools, files, skills, and a persistent memory, which turns it into an agent that reads live and historical data, calls external services, and remembers what it learns. Use it when a flow needs an AI model that holds context, acts on the system, or serves as an HTTP endpoint.

How it works

The processor sends each incoming message to the configured AI model, along with the recent conversation history, and returns the model's response. Max Message History sets how many messages the processor keeps. When the history reaches that limit, each new message drops the oldest. This gives the model recent context without letting the history grow indefinitely.

The processor also copies an incoming request_id field to its output, so a Webhook in async mode, this processor, and a Webhook response form a loop that works out of the box. This turns the processor into an HTTP endpoint backed by an AI model.

Model configuration

Set AI Model to the unique name of the model to use, and AI Model Instructions to the system prompt that shapes its behavior. HTTP Timeout caps how long the processor waits for a response, Max Output Tokens caps how much the model generates, and File Download Timeout caps how long the processor waits for files referenced in a response.

Optional features

These capabilities run inside the model when it supports them:

Advanced Thinking / Reasoning enables extended thinking mode.

Enable Web Search lets the model search the web during generation.

Enable Code Execution lets the model run code during generation.

Custom Tools accepts a JSON array of tool descriptors that the processor appends to the request. The descriptor shape is vendor specific and passed through as is, so you can add MCP server entries or Anthropic skill, memory, and bash tool blocks.

Context from AI Suite files

You attach reusable content from the AI Suite page and select it by checkmark. The processor reads these files fresh on every event, so edits take effect without a restart, and the processor skips a deleted file.

Skills bundles skill files (.md, .txt) into the prompt, concatenated in selection order with per-file headers.

Instruction Files prepends always-on guidance to the prompt.

Context Files attaches images (jpeg, png, gif, or webp) and PDFs as multimodal content blocks, sent inline as base64. The request size limit is 50 MB on OpenAI and 32 MB on Anthropic.

Local tools

Caution: The litmus_cli tool can call any SDK function the credential allows, including writes. Use scoped credentials so the model cannot make changes you did not intend.

Local tools run inside the processor when the model calls them, in a loop bounded by Max Tool Iterations and the HTTP timeout. Raise HTTP Timeout, for example to 120, when you enable any of them, and pair them with a skill file that teaches the model the conventions. The processor reports executed calls in the tool_activity output field.

Enable Litmus CLI Tool exposes a litmus_cli tool that runs the Litmus Edge (LE) SDK CLI against this instance. Litmus CLI Path points at the binary, and Litmus CLI Environment provides the EDGE_URL, EDGE_API_CLIENT_ID, and EDGE_API_CLIENT_SECRET credentials as KEY=VALUE lines.

Enable DataHub Read Tool exposes a read-only datahub_read tool that subscribes to a DataHub topic and returns live values.

Enable TSDB Query Tool exposes a read-only tsdb_query tool that queries historical data from the local time-series database.

Tools selects AI Suite tool definition files (.json). A definition targets either a Webhook processor in this instance through its hook_path, which the processor dispatches in-process with no HTTP, or an external endpoint through its url.

MCP Servers selects AI Suite MCP server registrations. The processor connects over Streamable HTTP, discovers each server's tools per event, and exposes them to the model. The processor skips an unreachable server with a warning rather than failing the event.

Memory and knowledge graph

Note: Memory files are visible and editable on the AI Suite page and over FTP, so treat their content as untrusted input, since it is injected into the prompt.

Enable Memory gives the model a persistent folder it manages itself with the memory_list, memory_read, memory_write, memory_append, memory_delete, memory_move, and memory_search tools. The folder is this instance's own subfolder of the AI Suite Memory section, named by its identifier, so renaming the processor keeps its memory and no two instances share one. A memory index is injected into the prompt on every event.

Enable Knowledge Graph adds a knowledge_sync tool that maintains a Knowledge Graph of the entities the model observes, such as devices, tags, and Digital Twins, with typed relations and provenance. It requires memory. Leave it off unless the agent surveys the system, because the tool's schema is sent to the model on every request.

Shared Memory Write and Shared Memory Read share one namespace between agents. The single writer maintains the namespace, and every reader lists it as read-only.

Output fields

The processor adds the following fields to its output:

value: the assistant response text.

reasoning_summary: a summary of the reasoning steps, present when advanced reasoning is enabled.

optional_file_contents, file_name, file_id, and container_id: present when the model produces file output.

tool_activity: the list of tool calls executed during the event, present when a local tool ran.

Parameters

Parameter	Details
Name	The unique identifier for this instance of the AI Responses processor.
AI Model	The unique name of the AI model to use.
AI Model Instructions	The instructions, or system prompt, for the AI model.
HTTP Timeout	The maximum time in seconds to wait for a response. Default is 10. Raise it, for example to 120, when you enable local tools.
Max Output Tokens	The maximum number of tokens the model can generate per response. Leave empty for the model default.
Max Message History	The maximum number of messages to keep in the conversation history. Default is 20, which is 10 exchanges. Older messages are dropped when the limit is reached.
File Download Timeout	The timeout in seconds for downloading files referenced in responses. Default is 30.
Advanced Thinking / Reasoning	Enables extended thinking mode where the model supports it.
Enable Web Search	Allows the model to search the web during generation.
Enable Code Execution	Allows the model to run code during generation.
Custom Tools	A JSON array of custom tool descriptors appended to the request tools list, for MCP servers and Anthropic skill, memory, and bash tools. The descriptor shape is vendor specific. Leave empty for none.
Skills	AI Suite skill files (.md, .txt) bundled into the prompt on every event, selected by checkmark in the UI. Leave empty to disable.
Instruction Files	AI Suite instruction files prepended to the prompt on every event as always-on guidance, selected by checkmark in the UI. Leave empty to disable.
Context Files	AI Suite image and PDF files attached as multimodal content blocks, sent inline as base64. Leave empty to disable.
Enable Litmus CLI Tool	Exposes a litmus_cli tool that runs the Litmus Edge SDK CLI locally with the arguments the model provides.
Litmus CLI Path	The path to the litmus-sdk-cli binary. Default is /ftp-data/bin/litmus-sdk-cli.
Litmus CLI Environment	The environment for the CLI, one KEY=VALUE per line (EDGE_URL, EDGE_API_CLIENT_ID, EDGE_API_CLIENT_SECRET). Lines starting with # are ignored.
Enable DataHub Read Tool	Exposes a read-only datahub_read tool that subscribes to a DataHub topic and returns live values.
Enable TSDB Query Tool	Exposes a read-only tsdb_query tool that queries historical data from the time-series database.
Max Tool Iterations	The maximum number of tool-call round trips per event. Default is 5.
Tools	AI Suite tool definition files (.json) exposed to the model, re-read on every event. Each targets a Webhook hook_path in this instance or an external url. Leave empty to disable.
MCP Servers	AI Suite MCP server registrations whose tools are exposed to the model, selected by checkmark in the UI. Leave empty to disable.
Enable Memory	Gives the model a persistent memory folder it manages itself with the memory_* tools, plus a memory index injected into the prompt on every event.
Enable Knowledge Graph	Adds a knowledge_sync tool that maintains a Knowledge Graph of observed entities in the memory folder. Requires memory.
Shared Memory Write	The shared memory namespace this instance writes, which makes it the namespace's only writer. Leave empty for a private agent.
Shared Memory Read	The shared memory namespaces this instance reads but cannot change, as a JSON array of names. Leave empty for none.
Timer Interval	Timer in milliseconds in which to resend the last payload. Enter 0 to disable this timer.
Pass Through Value	Joins the input with the output of the processor.