ed89af4b4d
# 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`
175 lines
5 KiB
Python
175 lines
5 KiB
Python
from arcade_evals import (
|
|
BinaryCritic,
|
|
EvalRubric,
|
|
EvalSuite,
|
|
ExpectedToolCall,
|
|
tool_eval,
|
|
)
|
|
from arcade_tdk import ToolCatalog
|
|
|
|
import arcade_slack
|
|
from arcade_slack.models import ConversationType
|
|
from arcade_slack.tools.chat import list_conversations
|
|
|
|
# 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 list_conversations_eval_suite() -> EvalSuite:
|
|
"""Create an evaluation suite for tools listing conversations."""
|
|
suite = EvalSuite(
|
|
name="Slack Messaging 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,
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List all conversations I am a member of",
|
|
user_message="List all conversations I am a member of",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": None,
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List 10 conversations I am a member of",
|
|
user_message="List 10 conversations I am a member of",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": None,
|
|
"limit": 10,
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=0.5),
|
|
BinaryCritic(critic_field="limit", weight=0.5),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List all public channels",
|
|
user_message="List all public channels",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [ConversationType.PUBLIC_CHANNEL.value],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List all private channels",
|
|
user_message="List all private channels",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [ConversationType.PRIVATE_CHANNEL.value],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List all public and private channels",
|
|
user_message="List all public and private channels",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [
|
|
ConversationType.PUBLIC_CHANNEL.value,
|
|
ConversationType.PRIVATE_CHANNEL.value,
|
|
],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List direct message channels",
|
|
user_message="List direct message channels",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [
|
|
ConversationType.DIRECT_MESSAGE.value,
|
|
],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List group direct message channels",
|
|
user_message="List group direct message channels",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [
|
|
ConversationType.MULTI_PERSON_DIRECT_MESSAGE.value,
|
|
],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
suite.add_case(
|
|
name="List my multi-person conversations",
|
|
user_message="List my multi-person conversations",
|
|
expected_tool_calls=[
|
|
ExpectedToolCall(
|
|
func=list_conversations,
|
|
args={
|
|
"conversation_types": [
|
|
ConversationType.MULTI_PERSON_DIRECT_MESSAGE.value,
|
|
],
|
|
},
|
|
)
|
|
],
|
|
critics=[
|
|
BinaryCritic(critic_field="conversation_types", weight=1.0),
|
|
],
|
|
)
|
|
|
|
return suite
|