Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
74 changes: 74 additions & 0 deletions samples/email-triage-agent/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,74 @@
# Email Triage Agent

A long-running LangGraph agent that watches a UiPath Integration Services Outlook connection for new emails matching a subject filter, classifies each one with an LLM, and replies to the original email with a polite acknowledgement drafted by the LLM.

The graph has no terminal node — once started, the agent stays SUSPENDED on the Outlook trigger forever, briefly waking to triage and reply to each matching email and then re-suspending. Cancel the job manually when you're done.

## What this sample demonstrates

- **`WaitIntegrationEvent`** — the agent suspends until an external IS connector event fires. The Connections-service registers a remote subscription on the user's behalf; when a matching email arrives, Orchestrator resumes the job and the SDK enriches the IS event metadata into the actual Microsoft Graph `Message`.
- An LLM call with strict structured output (Pydantic schema for the triage result).
- A direct Microsoft Graph call to send the reply, authenticated with the OAuth token issued for the same UiPath connection that received the trigger.

## Flow

```
START
└─► wait_for_email (suspend on Outlook EMAIL_RECEIVED, resume with Graph Message)
└─► triage_email (LLM → severity / category / summary / suggested_response)
└─► send_reply (Graph POST /me/messages/{id}/reply with the LLM draft)
└─► finalize (log result, clear transient state, increment counter)
└─► wait_for_email (loop)
```

## Input

The agent takes four required inputs at job start:

```json
{
"subject": "Issue",
"connection_name": "support@example.com",
"connection_folder": "Support",
"connector": "uipath-microsoft-outlook365"
}
```

| Field | Description |
|---|---|
| `subject` | Exact email subject to watch for. The IS trigger registers a server-side filter `(subject=='<value>')` so only matching emails fire it. |
| `connection_name` | Name of the Outlook 365 connection in UiPath Integration Services. |
| `connection_folder` | Folder where the connection lives. |
| `connector` | Connector key in the IS catalog (typically `uipath-microsoft-outlook365`). |

All four are persisted in state across loop iterations — set once at job start.

## Connection requirements

The Outlook connection passed in via input must be authorized to **read AND send** mail (`Mail.Read` + `Mail.Send` Graph scopes), and the `EMAIL_RECEIVED` IS trigger must be enabled on the connector. Re-authorize the connection from the UiPath Connections UI if either scope is missing.

## Running locally

```bash
uv sync
uipath run agent '{"subject": "Issue", "connection_name": "support@example.com", "connection_folder": "Support", "connector": "uipath-microsoft-outlook365"}'
```

The agent suspends waiting for the first matching email. Send (or have someone send) a message with subject `Issue` to the inbox the connection is bound to. When it arrives, the agent resumes, triages, replies, logs the result, and re-suspends on the next email.

Sample iteration log:

```
[INFO] Waiting for next email on 'support@example.com' (folder='Support') with subject='Issue' (triaged so far: 0)...
[INFO] Received email from alice@example.com: Issue
[INFO] Triage: severity=P0_critical category=bug
[INFO] Reply sent.
[INFO] Triaged email #1 from alice@example.com (subject='Issue', severity=P0_critical, category=bug, reply_sent=True)
[INFO] Waiting for next email on 'support@example.com' (folder='Support') with subject='Issue' (triaged so far: 1)...
```

## Notes

- **All four inputs are set once at job start.** Persisted in state across loop iterations — to change any of them (subject, connection, folder, connector), cancel the job and start a new one.
- **Long-running pattern.** This sample is deliberately a single long-lived job to demo `WaitIntegrationEvent` cleanly. The idiomatic UiPath production pattern for "react to many emails" is the inverse: configure an Orchestrator event trigger that starts a fresh, one-shot agent job per matching email. That gives you a finite lifecycle per email, parallel processing, and no recursion-limit concerns. Use whichever shape fits your operational model.
- **Adding human-in-the-loop later.** The sample previously included a `route_to_human` step using `CreateTask` to escalate high-severity emails through Action Center. That branch was removed for shipping simplicity. Adding it back when you have a working Action Center app or wrapper process is straightforward — the LLM's structured output already supports a `requires_human` field if you need to re-introduce conditional escalation.
15 changes: 15 additions & 0 deletions samples/email-triage-agent/agent.mermaid
Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
%% AUTO-GENERATED by `uipath init`. Do not edit manually.
%% Regenerated on every `uipath init`.
flowchart TB
__start__(__start__)
wait_for_email(wait_for_email)
triage_email(triage_email)
send_reply(send_reply)
finalize(finalize)
__end__(__end__)
__start__ --> wait_for_email
finalize --> wait_for_email
send_reply --> finalize
triage_email --> send_reply
wait_for_email --> triage_email
triage_email --> __end__
272 changes: 272 additions & 0 deletions samples/email-triage-agent/graph.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,272 @@
"""Long-running support inbox triage agent.

Watches a UiPath Integration Services Outlook connection for emails whose
subject matches the value passed in as agent input. Each match:

1. Resumes the suspended job with the enriched Microsoft Graph `Message`
as the resume value of `WaitIntegrationEvent`.
2. The LLM classifies the email into severity, category, a one-sentence
summary, and a polite acknowledgement draft.
3. The agent replies to the original email with the LLM-drafted
acknowledgement (via Microsoft Graph, using the connection's OAuth token).
4. The result is logged, transient state is cleared, and the agent loops
back to suspend on the next matching email.

The graph has no terminal node — the agent stays SUSPENDED on the Outlook
trigger forever, briefly waking to triage and reply to each matching email
and then re-suspending. Cancel the job manually when you're done with it.

Demonstrates one suspend/resume primitive in a long-running agent:
- `WaitIntegrationEvent` — suspend until an external IS connector event fires.
"""

import logging
from enum import Enum
from typing import Any, Optional

import httpx
from langchain_core.messages import HumanMessage, SystemMessage
from langgraph.graph import START, StateGraph
from langgraph.types import interrupt
from pydantic import BaseModel, Field
from uipath.platform import UiPath
from uipath.platform.common import WaitIntegrationEvent
from uipath_langchain.chat import UiPathChat

logger = logging.getLogger(__name__)

GRAPH_API_BASE = "https://graph.microsoft.com/v1.0"


class Severity(str, Enum):
P0_CRITICAL = "P0_critical"
P1_HIGH = "P1_high"
P2_NORMAL = "P2_normal"
P3_LOW = "P3_low"


class Category(str, Enum):
BUG = "bug"
FEATURE_REQUEST = "feature_request"
HOWTO = "howto"
BILLING = "billing"
SPAM = "spam"
OTHER = "other"


class Triage(BaseModel):
severity: Severity = Field(
description=(
"P0 = production outage / data loss, "
"P1 = major workflow impact, "
"P2 = normal request or single-user impact, "
"P3 = low / cosmetic / general question."
)
)
category: Category
summary: str = Field(description="One-sentence summary in the customer's voice.")
suggested_response: str = Field(
description="Polite acknowledgement reply confirming receipt and next steps."
)


class GraphInput(BaseModel):
subject: str = Field(
description="The exact email subject to watch for. The IS trigger filters incoming emails by this value."
)
connection_name: str = Field(
description="Name of the Outlook 365 connection in UiPath Integration Services."
)
connection_folder: str = Field(
description="Folder where the Outlook connection lives."
)
connector: str = Field(
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can't we hardcode this value?

description="Connector key in the IS catalog (typically 'uipath-microsoft-outlook365')."
)


class GraphState(BaseModel):
subject: str = ""
connection_name: str = ""
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

let s have the connection configured as a binding

connection_folder: str = ""
connector: str = ""
email: Optional[dict[str, Any]] = None
triage: Optional[Triage] = None
reply_sent: Optional[bool] = None
reply_body: Optional[str] = None
triage_count: int = 0


llm = UiPathChat(model="gpt-4o-mini-2024-07-18")


def _email_str(email: dict[str, Any], *path: str, default: str = "") -> str:
cur: Any = email
for p in path:
if not isinstance(cur, dict):
return default
cur = cur.get(p)
return cur if isinstance(cur, str) else default


async def _send_outlook_reply(
message_id: str,
body: str,
connection_name: str,
connection_folder: str,
connector: str,
) -> None:
"""Reply to an Outlook message via Microsoft Graph, using the OAuth token
issued for the UiPath Outlook connection that received the trigger.
"""
sdk = UiPath()
connections = await sdk.connections.list_async(
name=connection_name,
folder_path=connection_folder,
connector_key=connector,
)
connection = next(
(c for c in connections if c.name == connection_name),
None,
)
if connection is None or connection.id is None:
raise RuntimeError(
f"Outlook connection {connection_name!r} not found in "
f"folder {connection_folder!r}."
)

token = await sdk.connections.retrieve_token_async(connection.id)

async with httpx.AsyncClient(timeout=30) as client:
response = await client.post(
f"{GRAPH_API_BASE}/me/messages/{message_id}/reply",
headers={
"Authorization": f"Bearer {token.access_token}",
"Content-Type": "application/json",
},
json={"comment": body},
)
response.raise_for_status()


async def wait_for_email(state: GraphState) -> dict[str, Any]:
logger.info(
"Waiting for next email on '%s' (folder='%s') with subject=%r (triaged so far: %d)...",
state.connection_name,
state.connection_folder,
state.subject,
state.triage_count,
)
email = interrupt(
WaitIntegrationEvent(
connector=state.connector,
connection_name=state.connection_name,
connection_folder_path=state.connection_folder,
operation="EMAIL_RECEIVED",
object_name="Message",
filter_expression=f"(subject=='{state.subject}')",
)
)
sender = _email_str(email, "from", "emailAddress", "address", default="?")
logger.info("Received email from %s: %s", sender, _email_str(email, "subject"))
return {"email": email}


async def triage_email(state: GraphState) -> dict[str, Any]:
email = state.email or {}
sender = _email_str(email, "from", "emailAddress", "address", default="unknown")
subject = _email_str(email, "subject")
body = _email_str(email, "bodyPreview") or _email_str(email, "body", "content")

triage_llm = llm.with_structured_output(Triage)
result: Triage = await triage_llm.ainvoke(
[
SystemMessage(
"You are a support triage assistant. Read the customer email and "
"produce a structured triage result.\n\n"
"Severity guidelines:\n"
"- P0: production outage, data loss, or anything blocking critical work.\n"
"- P1: major workflow impact; affects many users.\n"
"- P2: normal request or single-user impact.\n"
"- P3: low priority, cosmetic, or general question.\n\n"
"Always draft a polite acknowledgement confirming receipt and "
"setting expectations for next steps."
),
HumanMessage(f"From: {sender}\nSubject: {subject}\n\n{body}"),
]
)
logger.info(
"Triage: severity=%s category=%s",
result.severity.value,
result.category.value,
)
return {"triage": result}


async def send_reply(state: GraphState) -> dict[str, Any]:
email = state.email or {}
triage = state.triage
message_id = email.get("id") if isinstance(email, dict) else None
body = triage.suggested_response if triage else None

if not body:
logger.warning("No reply body resolved — skipping send.")
return {"reply_sent": False, "reply_body": None}

if not message_id:
logger.warning("Email payload had no 'id' field — cannot send reply.")
return {"reply_sent": False, "reply_body": body}

try:
await _send_outlook_reply(
message_id,
body,
connection_name=state.connection_name,
connection_folder=state.connection_folder,
connector=state.connector,
)
logger.info("Reply sent.")
return {"reply_sent": True, "reply_body": body}
except Exception:
logger.exception("Failed to send Outlook reply.")
return {"reply_sent": False, "reply_body": body}


async def finalize(state: GraphState) -> dict[str, Any]:
triage = state.triage
assert triage is not None
email = state.email or {}
sender = _email_str(email, "from", "emailAddress", "address", default="unknown")
subject = _email_str(email, "subject")

logger.info(
"Triaged email #%d from %s (subject=%r, severity=%s, category=%s, reply_sent=%s)",
state.triage_count + 1,
sender,
subject,
triage.severity.value,
triage.category.value,
bool(state.reply_sent),
)
return {
"triage_count": state.triage_count + 1,
"email": None,
"triage": None,
"reply_sent": None,
"reply_body": None,
}


builder = StateGraph(GraphState, input_schema=GraphInput)
builder.add_node("wait_for_email", wait_for_email)
builder.add_node("triage_email", triage_email)
builder.add_node("send_reply", send_reply)
builder.add_node("finalize", finalize)

builder.add_edge(START, "wait_for_email")
builder.add_edge("wait_for_email", "triage_email")
builder.add_edge("triage_email", "send_reply")
builder.add_edge("send_reply", "finalize")
builder.add_edge("finalize", "wait_for_email")

graph = builder.compile()
5 changes: 5 additions & 0 deletions samples/email-triage-agent/langgraph.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
{
"graphs": {
"agent": "./graph.py:graph"
}
}
22 changes: 22 additions & 0 deletions samples/email-triage-agent/pyproject.toml
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
[project]
name = "email-triage-agent"
version = "0.0.1"
description = "Wait for new emails via UiPath Integration Services, triage them with an LLM, and optionally escalate to a human via Action Center."
authors = [{ name = "John Doe", email = "john.doe@myemail.com" }]
requires-python = ">=3.11"
dependencies = [
"httpx>=0.27",
"langgraph>=1.0.4",
"uipath-langchain"
]

[dependency-groups]
dev = [
"uipath-dev",
]

[[tool.uv.index]]
name = "testpypi"
url = "https://test.pypi.org/simple/"
publish-url = "https://test.pypi.org/legacy/"
explicit = true
Loading
Loading