Haystack

In a pipeline, NiadraContext goes between the prompt builder and the chat generator, and NiadraReply after the generator. With an Agent, NiadraAgentHooks(conversation).hooks and history_tools(conversation) do the same. Python only; needs Haystack 3.2 or newer.

Install

pip install 'niadra[haystack]'   # haystack-ai 3.2 or newer, below 4

The five primitives

Primitive	How the adapter wires it
Context	The pack (after the agent’s own notes, with `agent_memory=`) goes right after the leading system messages and the `turn_block` at the end, as system messages. Messages Niadra put in an earlier turn are taken out first, so a history passed back in never holds them twice, and the agent’s `messages` output never carries them
Turns	`NiadraContext` and the `before_run` hook record the last user message as the customer’s turn; `NiadraReply` and the `after_run` hook record the assistant’s text as the agent’s, with its usage and model
Tools	`history_tools()` gives Haystack `Tool`s whose parameters are the kit’s JSON Schemas, bound to the customer, with a sync and an async function
Verification	`conversation.verify()` before the run
Handoff	Outside the components: `conversation.handoff()` where your flow transfers

Minimal example

"""A Haystack Agent with the customer's memory as hooks and tools."""

from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIChatGenerator
from haystack.dataclasses import ChatMessage

from niadra import Niadra, phone
from niadra.integrations.haystack import NiadraAgentHooks, history_tools

niadra = Niadra(channel="chat")
conversation = niadra.conversation("thread-81", subject=phone("+5511912345678"))
agent = Agent(
    chat_generator=OpenAIChatGenerator(model="gpt-4.1"),
    system_prompt="You are Acme's agent.",
    tools=history_tools(conversation),
    hooks=NiadraAgentHooks(conversation).hooks,
)
result = agent.run(messages=[ChatMessage.from_user("Where is my replacement lid?")])
print(result["last_message"].text)
niadra.close()

The same code is in examples/haystack_agent.py.

In a pipeline

from haystack import Pipeline
from haystack.components.builders import ChatPromptBuilder
from haystack.components.generators.chat import OpenAIChatGenerator

from niadra.integrations.haystack import NiadraContext, NiadraReply

pipeline = Pipeline()
pipeline.add_component("prompt", ChatPromptBuilder())
pipeline.add_component("niadra", NiadraContext(conversation))
pipeline.add_component("llm", OpenAIChatGenerator())
pipeline.add_component("reply", NiadraReply(conversation))
pipeline.connect("prompt.prompt", "niadra.messages")
pipeline.connect("niadra.messages", "llm.messages")
pipeline.connect("llm.replies", "reply.replies")

Agent memory

NiadraContext(conversation, agent_memory=True), NiadraAgentHooks(conversation, agent_memory=True) and history_tools(conversation, agent_memory=True) put the agent’s own notes before the customer’s context and add the two tools. See Agent memory.

Limits

The components and the hooks take Niadra and AsyncNiadra conversations; with AsyncNiadra, run the pipeline or the agent with run_async.
Nothing here stops the pipeline: with Niadra slow or down, the messages go on without the pack and a tool answers that the history is unavailable.
Tested against haystack-ai 3.2.0 with the generator replaced by a fake and Niadra on the emulator.

Get started

Voice

Channels

Agent frameworks

Model SDKs

Automation platforms

MCP clients

Install

The five primitives

Minimal example

In a pipeline

Agent memory

Limits

Next steps

History navigation

Context and views

​Install

​The five primitives

​Minimal example

​In a pipeline

​Agent memory

​Limits

​Next steps

History navigation

Context and views

Install

The five primitives

Minimal example

In a pipeline

Agent memory

Limits

Next steps