arcade-mcp/toolkits/slack/evals/chat/eval_get_users_in_conversation.py
Renato Byrro ed89af4b4d
Slack Toolkit Refactoring (#453)
# Backwards-compatible refactoring of the Slack toolkit

Several performance improvements, streamlined tool set, and easier to
understand tool interfaces.

All "old" tools were kept for backwards compatibility, with the same
interfaces and response structure (but using the new and more performant
tools under the hood).

Full revision of unit-tests and evals.

##  Streamlined tool set

Multiple groups of tools were streamlined into a single one:

* "get conversation metadata" from 5 tools to one;
* "send message" from 2 tools to one;
* "get users in conversation" from 3 tools to one;
* "get messages" from 4 tools to one

## New capabilities

* Messages retrieved are now populated with the users' names, apart from
ID: makes it easier for LLMs to reference who sent a message, were
mentioned, or reacted to a message
* Retrieve users by username, email, and/or ID (before we only supported
ID)
* Retrieve multiple users in a single tool call

## Concurrency controls

All operations issuing multiple requests concurrently now have a
`Semaphore` to limit the concurrency level. The limit can be controlled
through the `SLACK_MAX_CONCURRENT_REQUESTS` env var (defaults to 3).

## Networking performance improvement

Various operations that used to make multiple API calls are now executed
more efficiently:

### Find users by username

* Before: a full scan of `users_list` was required (potentially multiple
pages for large workspaces);
* Now it stops as soon as we have all users needed (yes, it was dumb
before)

### Get multiple users by their IDs

* Before: for each user ID, we made one API call to the `users_info`
endpoint
* Now: we call `list_users` and scan the results to match the user IDs
(an estimate of 99.5% of Slack workspaces have < 200 users; for large
workspaces, we may need to paginate `list_users`)

### Get a conversation by its users

* Before:
  * Call to `list_conversations` (potentially paginating)
* For each conversation, one call to `conversations_members`
(potentially paginating)
  * Then loop and find which conversation matches the users' IDs
* Now:
  * A single call to `conversations_open`
2025-07-10 13:50:28 -03:00

81 lines
2.4 KiB
Python

from arcade_evals import (
BinaryCritic,
EvalRubric,
EvalSuite,
ExpectedToolCall,
tool_eval,
)
from arcade_tdk import ToolCatalog
import arcade_slack
from arcade_slack.tools.chat import get_users_in_conversation
# Evaluation rubric
rubric = EvalRubric(
fail_threshold=0.8,
warn_threshold=0.9,
)
catalog = ToolCatalog()
# Register the Slack tools
catalog.add_module(arcade_slack)
@tool_eval()
def get_users_in_conversation_eval_suite() -> EvalSuite:
"""Create an evaluation suite for tools getting conversations members."""
suite = EvalSuite(
name="Slack Tools Evaluation",
system_message="You are an AI assistant that can interact with Slack to send messages and get information from conversations, users, etc.",
catalog=catalog,
rubric=rubric,
)
user_messages = [
"Get the members of the #general channel",
"Get the users in the #general channel",
"Get a list of people in the #general channel",
"Get a list of people in the general channel",
"Show me who's in the #general channel",
"Who is in the general channel?",
]
for user_message in user_messages:
suite.add_case(
name=f"Get users in channel by channel name: {user_message}",
user_message=user_message,
expected_tool_calls=[
ExpectedToolCall(
func=get_users_in_conversation,
args={
"conversation_id": None,
"channel_name": "general",
},
),
],
critics=[
BinaryCritic(critic_field="conversation_id", weight=0.4),
BinaryCritic(critic_field="channel_name", weight=0.6),
],
)
suite.add_case(
name="Get users in conversation by conversation id",
user_message="Get the users in the conversation with id '1234567890'",
expected_tool_calls=[
ExpectedToolCall(
func=get_users_in_conversation,
args={
"conversation_id": "1234567890",
"channel_name": None,
},
),
],
critics=[
BinaryCritic(critic_field="conversation_id", weight=0.6),
BinaryCritic(critic_field="channel_name", weight=0.4),
],
)
return suite