feat: multi-agent AI XBlock #9
Conversation
ArturGaspar
force-pushed
the
artur/multiagent
branch
2 times, most recently
from
e6fc51e to
0a58220
Compare
ArturGaspar
force-pushed
the
artur/multiagent
branch
from
0a58220 to
f6f0470
Compare
|
|
||
|
|
||
| DEFAULT_SUPERVISOR_PROMPT = textwrap.dedent(""" | ||
| You are a supervisor managing an interaction between the following agents: Coach, Character. |
There was a problem hiding this comment.
@ArturGaspar Was this prompt provided by WGU?
| role: str | ||
|
|
||
|
|
||
| class CharacterData(pydantic.BaseModel): |
There was a problem hiding this comment.
These pydantic models seem to be used for validation. Any reason for not using django serializers?
There was a problem hiding this comment.
@pkulkark Is there any reason to prefer them?
There was a problem hiding this comment.
@ArturGaspar Just for consistency with the rest of edx-platform xblocks. Since the xblock uses Django models, using a Django serializer would be more idiomatic.
| ValidationMessage.ERROR, | ||
| f"{self._get_field_display_name('agent_prompt')}: {e}", | ||
| )) | ||
| else: |
There was a problem hiding this comment.
Why do we need else here? Couldn't we just continue without it?
There was a problem hiding this comment.
The part under the else checks that the template will render correctly for each character. The part above that it will render at all with the scenario data.
If the agent prompt has errors with the scenario data, it would be pointlessly noisy to repeat it for every character.
| <textarea class="user-input" rows="1" placeholder="Type your answer here" maxlength="1000"></textarea> | ||
| <span id="submit-button">Submit <i class="fa fa-paper-plane"></i></span> | ||
| {% if has_finish_button %} | ||
| <span class="chat-button" id="finish-button">Submit for evaluation</span> |
There was a problem hiding this comment.
@ArturGaspar It might be better to disable this button until there's a user response. What do you think?
| <span class="chat-button" id="finish-button">Submit for evaluation</span> | ||
| {% endif %} | ||
| <textarea id="user-input" rows="1" placeholder="Type your answer here" maxlength="1000"></textarea> | ||
| <span class="chat-button" id="submit-button">Submit <i class="fa fa-paper-plane"></i></span> |
| else: | ||
| message = self._get_agent_response(agent, user_input) | ||
| character_name = self._get_character_name(agent) | ||
| character_data = self._get_character_data(character_name) |
There was a problem hiding this comment.
Isn't the character_name and character_data being fetched within _get_agent_response function too? Couldn't we just pass that info back instead of calling them again?
Description
Adds an XBlock similar to the short answer one, but which can select between different instructions for the LLM depending on context.
Testing instructions
The below example with the default settings shows how to test this functionality.
Other information
Private-ref: https://tasks.opencraft.com/browse/BB-9325