skillpacks 0.1.22

Pluggable skills for AI agents

SkillPacks

Pluggable skillsets for AI agents
Explore the docs »

View Demo · Report Bug · Request Feature

Skillpacks provide a means of fine tuning agents on tools, and the ability to hotswap learned skills at inference time.

Teach a model how to use a website | code base | API | database | application | ...   then swap in that learned layer the moment you need it.

Install

pip install skillpacks

Quick Start

Create an episode to record agent events

from skillpacks import Episode

episode = Episode(remote="https://foo.bar")

Take an action

from mllm import Router, RoleThread
from skillpacks import V1Action
from agentdesk import Desktop

router = Router.from_env()
desktop = Desktop.local()

thread = RoleThread()
msg = f"""
I need to open Google to search, your available action are {desktop.json_schema()}
please return your selection as {V1Action.model_json_schema()}
"""
thread.post(role="user", msg=msg)

response = router.chat(thread, expect=V1Action)
v1action = response.parsed

action = desktop.find_action(name=v1action.name)
result = desktop.use(action, **v1action.parameters)

Record the action in the episode

event = episode.record(
    prompt=response.prompt,
    action=v1action,
    tool=desktop.ref(),
    result=result,
)

Mark actions as approved

# approve one
episode.approve_one(event.id)

# approve the event and all actions prior to it
episode.approve_prior(event.id)

# approve all
episode.approve_all()

Get all approved actions in an episode

episode = Episode.find(id="123")[0]
actions = episode.approved_actions()

Get all approved actions in a namespace

from skillpacks import ActionEvent

actions = ActionEvent.find(namespace="foo", approved=True)

Get all approved actions for a tool

actions = ActionEvent.find(tool=desktop.ref(), approved=True)

Tune a model on the actions (In progress)

from skillpacks.model import InternVLChat
from skillpacks.runtime import KubernetesRuntime

runtime = KubernetesRuntime()
model = InternVLChat(runtime=runtime)

result = model.train(actions=actions, follow=True, publish=True)

Backends

Thread and prompt storage can be backed by:

• Sqlite
• Postgresql

Sqlite will be used by default. To use postgres simply configure the env vars:

DB_TYPE=postgres
DB_NAME=skills
DB_HOST=localhost
DB_USER=postgres
DB_PASS=abc123