Semantic Kernel

NiadraKernel brings three pieces: register(kernel) adds the niadra plugin and two filters (prompt rendering and function invocation) to the kernel, thread() is a ChatCompletionAgent thread, and plugin is the kit as a KernelPlugin. Python only.

Install

pip install 'niadra[semantic-kernel]'   # semantic-kernel 1.44.1 or newer, below 2

The five primitives

Primitive	How the adapter wires it
Context	With a `ChatCompletionAgent`, `memory.thread()` is the agent’s thread: on every read the pack (after the agent’s own notes, with `agent_memory=`) comes first, which the agent places right after its instructions; then the stored messages; then the `turn_block` at the end. Only the conversation is stored; the Niadra messages are made fresh on every read. With a prompt function (`kernel.invoke_prompt`, `kernel.invoke(function, ...)`), the prompt rendering filter places the same things in the rendered prompt. When you call a chat service yourself, `await memory.chat_history(history)` gives a copy with them
Turns	The thread records the user’s messages as the customer’s turns and the assistant’s text as the agent’s, with its usage and model; with a prompt function, the filters record the last user message of the prompt and the function’s answer
Tools	`memory.plugin` is a `KernelPlugin` named `niadra` whose functions are the history tools, with the kit’s names, descriptions and JSON Schemas, bound to the customer. Semantic Kernel shows the model a function as `plugin-function`: `niadra-search_customer_history`
Verification	`conversation.verify()` before the first context
Handoff	`transferred_to_agent()` and `transferred_to_human()` record a transfer, in a group chat or an orchestration that hands over, or an escalation to a person

Minimal example

"""A Semantic Kernel ChatCompletionAgent with the customer's memory as a thread, filters and a plugin."""

import asyncio

from semantic_kernel import Kernel
from semantic_kernel.agents import ChatCompletionAgent
from semantic_kernel.connectors.ai import FunctionChoiceBehavior
from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion

from niadra import AsyncNiadra, phone
from niadra.integrations.semantic_kernel import NiadraKernel

niadra = AsyncNiadra(channel="chat")


async def main() -> None:
    memory = NiadraKernel(niadra.conversation("thread-81", subject=phone("+5511912345678")))
    agent = ChatCompletionAgent(
        service=OpenAIChatCompletion(ai_model_id="gpt-4.1"),
        kernel=memory.register(Kernel()),
        name="acme",
        instructions="You are Acme's agent.",
        function_choice_behavior=FunctionChoiceBehavior.Auto(),
    )
    thread = memory.thread()
    response = await agent.get_response(messages="Where is my replacement lid?", thread=thread)
    print(response.content)
    await niadra.close()


asyncio.run(main())

The same code is in examples/semantic_kernel_agent.py.

Agent memory

NiadraKernel(conversation, agent_memory=True) puts the agent’s own notes before the customer’s context and adds the two tools to the plugin. See Agent memory.

Limits

Works with Niadra and AsyncNiadra conversations.
Nothing here stops the kernel: with Niadra slow or down, the prompt goes out without the pack and a tool answers that the history is unavailable.
The extra installs a pre-release of azure-ai-agents, which semantic-kernel 1.44 requires.
Tested against semantic-kernel 1.44.1 with the chat service replaced by a fake and Niadra on the emulator.

Get started

Voice

Channels

Agent frameworks

Model SDKs

Automation platforms

MCP clients

Semantic Kernel

Install

The five primitives

Minimal example

Agent memory

Limits

Next steps

Microsoft Agent Framework

Agent memory

​Install

​The five primitives

​Minimal example

​Agent memory

​Limits

​Next steps

Microsoft Agent Framework

Agent memory

Install

The five primitives

Minimal example

Agent memory

Limits

Next steps