File size: 26,262 Bytes
900edd7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 |
import copy
import logging
from typing import Optional, Union
import gradio as gr
import pandas as pd
import src.cfg as cfg
from buster.completers import Completion
from src.app_utils import add_sources, get_session_id, get_utc_time
from src.cfg import setup_buster
from src.feedback import FeedbackForm, Interaction
logger = logging.getLogger(__name__)
logging.basicConfig(level=logging.INFO)
# Typehint for chatbot history
ChatHistory = list[list[Optional[str], Optional[str]]]
app_name = cfg.app_name
example_questions = cfg.example_questions
disclaimer = cfg.disclaimer
mongo_db = cfg.mongo_db
buster_cfg = copy.deepcopy(cfg.buster_cfg)
buster = setup_buster(buster_cfg=buster_cfg)
max_sources = cfg.max_sources
data_dir = cfg.data_dir
# link to the terms and conditions to be rendered in markdown blocks
path_to_tncs = "file=src/buster/assets/index.html"
md_link_to_tncs = f"[terms and conditions]({path_to_tncs})"
# get documents metadata
documents_metadata_file = str(data_dir / "documents_metadata.csv")
documents_metadata = pd.read_csv(documents_metadata_file)
css = """
.source {
max-height: 250px; /* Set the maximum height for the textboxes */
overflow: auto; /* Enable scrollbars when content exceeds dimensions */
outline: 1px solid gray; /* Add a gray outline */
border-radius: 5px; /* Add rounded corners to the outline */
}
"""
def add_disclaimer(completion: Completion, chat_history: ChatHistory, disclaimer: str = disclaimer):
"""Add a disclaimer response if the answer was relevant."""
if completion.question_relevant:
chat_history.append([None, disclaimer])
return chat_history
def hide_about_panel(accept_checkbox):
# Stay open while not accepted
open = not bool(accept_checkbox)
return {about_panel: gr.update(open=open)}
def set_relevant_sources_selection(num_sources: int):
relevant_sources_selection = gr.CheckboxGroup(
choices=[f"Source {i+1}" for i in range(num_sources)],
label="Check all relevant sources (if any)",
)
return relevant_sources_selection
def setup_feedback_form(num_sources: int):
# Feedback
feedback_elems = {}
with gr.Row():
with gr.Column():
gr.Markdown(
f""" ## We would love your feedback!
Please submit feedback for each question asked.
Your feedback is anonymous and will help us make the tool as useful as possible for the community!
"""
)
with gr.Row():
overall_experience = gr.Radio(choices=["π", "π"], label=f"Did {app_name} help answer your question?")
# Currently, we show all feedback, but also support having a small portion of it display at first
show_additional_feedback = gr.Group(visible=True)
with show_additional_feedback:
with gr.Column():
clear_answer = gr.Radio(
choices=["π", "π"], label="Was the generated answer clear and understandable?"
)
accurate_answer = gr.Radio(choices=["π", "π"], label="Was the generated answer accurate?")
relevant_sources = gr.Radio(
choices=["π", "π"],
label="Were the retrieved sources generally relevant to your query?",
)
relevant_sources_selection = set_relevant_sources_selection(num_sources=num_sources)
relevant_sources_order = gr.Radio(
choices=["π", "π"],
label="Were the sources ranked appropriately, in order of relevance?",
)
extra_info = gr.Textbox(
label="Any other comments?",
lines=3,
placeholder="Please enter other feedback for improvement here...",
)
expertise = gr.Radio(
choices=["Beginner", "Intermediate", "Expert"],
label="How would you rate your knowledge of AI policy",
interactive=True,
)
submit_feedback_btn = gr.Button("Submit feedback", variant="primary", interactive=True)
with gr.Column(visible=False) as submitted_message:
gr.Markdown("Feedback recorded, thank you π! You can now ask a new question in the search bar.")
# fmt: off
submit_feedback_btn.click(
toggle_visibility,
inputs=gr.State(False),
outputs=submitted_message,
).then(
submit_feedback,
inputs=[
overall_experience,
clear_answer,
accurate_answer,
relevant_sources,
relevant_sources_order,
relevant_sources_selection,
expertise,
extra_info,
last_completion,
session_id,
],
).success(
toggle_visibility,
inputs=gr.State(True),
outputs=submitted_message,
).success(
toggle_interactivity,
inputs=gr.State(False),
outputs=submit_feedback_btn,
)
# fmt: on
feedback_elems = {
"overall_experience": overall_experience,
"clear_answer": clear_answer,
"accurate_answer": accurate_answer,
"relevant_sources": relevant_sources,
"relevant_sources_selection": relevant_sources_selection,
"relevant_sources_order": relevant_sources_order,
"submit_feedback_btn": submit_feedback_btn,
"submitted_message": submitted_message,
"show_additional_feedback": show_additional_feedback,
"expertise": expertise,
"extra_info": extra_info,
}
return feedback_elems
def to_md_link(title: str, link: str) -> str:
"""Converts a title and link to markown link format"""
return f"[{title}]({link})"
def get_metadata_markdown(df) -> str:
"""Converts the content from a dataframe to a markdown table string format."""
metadata = []
# Order articles by year, with latest first
df = df.sort_values(["Country", "Year"], ascending=True)
for _, item in df.iterrows():
# source = item["Source"]
link = item["Link"]
title = item["Title"]
year = item["Year"]
country = item["Country"]
metadata.append(f"{year} | {country} | {to_md_link(title, link)} ")
metadata_str = "\n".join(metadata)
markdown_text = f"""
| Year | Country | Report |
| --- | --- | --- |
{metadata_str}
"""
return markdown_text
def add_user_question(user_question: str, chat_history: Optional[ChatHistory] = None) -> ChatHistory:
"""Adds a user's question to the chat history.
If no history is provided, the first element of the history will be the user conversation.
"""
if chat_history is None:
chat_history = []
chat_history.append([user_question, None])
return chat_history
def chat(chat_history: ChatHistory, reformulate_question: bool, top_k: Optional[int] = None):
"""Answer a user's question using retrieval augmented generation."""
# Make sure top k is an int between 1 and 15
top_k = int(top_k)
top_k = max(top_k, 1)
top_k = min(top_k, max_sources)
# We assume that the question is the user's last interaction
user_input = chat_history[-1][0]
completion = buster.process_input(user_input, reformulate_question=reformulate_question, top_k=top_k)
if completion.question_relevant and not completion.error:
if reformulate_question and user_input not in cfg.example_questions:
assert completion.user_inputs.reformulated_input is not None
chat_history.append(
[
None,
f"{cfg.message_before_reformulation}{completion.user_inputs.reformulated_input}{cfg.message_after_reformulation}",
]
)
chat_history.append([None, None])
# Stream tokens one at a time
chat_history[-1][1] = ""
for token in completion.answer_generator:
chat_history[-1][1] += token
yield chat_history, completion
def log_completion(
completion: Union[Completion, list[Completion]],
collection: str,
session_id: str,
request: gr.Request,
instance_type: Optional[str] = cfg.INSTANCE_TYPE,
instance_name: Optional[str] = cfg.INSTANCE_NAME,
mongo_db=cfg.mongo_db,
):
"""
Log user completions in a specified collection for analytics.
Parameters:
completion (Union[Completion, list[Completion]]): A single completion or a list of completions
to log. Completions can be instances of the Completion class.
collection (str): The name of the MongoDB collection where the interactions will be stored.
session_id (str): A unique identifier for the current session. In gradio this is reset every time a page is refreshed.
request (gr.Request): The gradio request object containing request metadata.
instance_type (str, optional): The type of instance where the completion took place.
Defaults to cfg.INSTANCE_TYPE.
instance_name (str, optional): The name of the instance where the completion took place.
Defaults to cfg.INSTANCE_NAME.
"""
# TODO: add UID for each page visitor instead of username
# make sure it's always a list
if isinstance(completion, Completion):
user_completions = [completion]
else:
user_completions = completion
interaction = Interaction(
user_completions=user_completions,
time=get_utc_time(),
username=request.username,
session_id=session_id,
instance_name=instance_name,
instance_type=instance_type,
data_version=cfg.MONGO_DATABASE_DATA,
)
interaction.send(mongo_db, collection=collection)
def submit_feedback(
overall_experience: str,
clear_answer: str,
accuracte_answer: str,
relevant_sources: str,
relevant_sources_order: list[str],
relevant_sources_selection: str,
expertise: list[str],
extra_info: str,
completion: Union[Completion, list[Completion]],
session_id: str,
request: gr.Request,
instance_type: Optional[str] = cfg.INSTANCE_TYPE,
instance_name: Optional[str] = cfg.INSTANCE_NAME,
):
feedback_form = FeedbackForm(
overall_experience=overall_experience,
clear_answer=clear_answer,
accurate_answer=accuracte_answer,
relevant_sources=relevant_sources,
relevant_sources_order=relevant_sources_order,
relevant_sources_selection=relevant_sources_selection,
expertise=expertise,
extra_info=extra_info,
)
# make sure it's always a list
if isinstance(completion, Completion):
user_completions = [completion]
else:
user_completions = completion
feedback = Interaction(
user_completions=user_completions,
form=feedback_form,
time=get_utc_time(),
username=request.username,
session_id=session_id,
instance_name=instance_name,
instance_type=instance_type,
)
feedback.send(mongo_db, collection=cfg.MONGO_COLLECTION_FEEDBACK)
def toggle_visibility(visible: bool):
"""Toggles the visibility of the gradio element."""
return gr.update(visible=visible)
def toggle_interactivity(interactive: bool):
"""Toggles the visibility of the gradio element."""
return gr.update(interactive=interactive)
def clear_user_input():
"""Clears the contents of the user_input box."""
return gr.Textbox(value="")
def clear_sources():
"""Clears all the documents in the tabs"""
return ["" for _ in range(max_sources)]
def clear_feedback_form():
"""Clears the contents of the feedback form."""
return {
feedback_elems["overall_experience"]: gr.update(value=None),
feedback_elems["clear_answer"]: gr.update(value=None),
feedback_elems["accurate_answer"]: gr.update(value=None),
feedback_elems["relevant_sources"]: gr.update(value=None),
feedback_elems["relevant_sources_selection"]: gr.update(value=None),
feedback_elems["relevant_sources_order"]: gr.update(value=None),
feedback_elems["expertise"]: gr.update(value=None),
feedback_elems["extra_info"]: gr.update(value=None),
}
def reveal_app(choice: gr.SelectData):
return (
gr.Group(visible=False),
gr.Textbox(interactive=True, value=""),
gr.Button(interactive=True),
)
def display_sources():
with gr.Column():
gr.Markdown(
"""## Relevant sources
All retrieved documents will be listed here in order of importance.
"""
)
sources_textboxes = []
for i in range(max_sources):
t = gr.Markdown(latex_delimiters=[], elem_classes="source", visible=False)
sources_textboxes.append(t)
return sources_textboxes
def setup_about_panel():
with gr.Accordion(label=f"About {app_name}", open=False) as about_panel:
with gr.Row(variant="panel"):
gr.Markdown(
f"""
## Welcome
Artificial intelligence is a field that's developing fast! In response, policy makers from around the world are creating guidelines, rules and regulations to keep up.
Finding accurate and up-to-date information about regulatory changes can be difficult but crucial to share best practices, ensure interoperability and promote adherence to local laws and regulations. That's why we've created {app_name}.
{app_name} is a Q&A search engine designed to provide relevant and high quality information about AI policies from around the world. Using this tool, your AI policy questions will be answered, accompanied by relevant analyses by the OECD's AI Observatory!
## How it works (and doesn't)
{app_name} uses Large Language Models (AI algorithms that work with text) to pinpoint sections of policy documents that are relevant to your question. Rather than presenting you with the specific policy section verbatim, {app_name} has been designed to summarize the information in a digestible format, so that the response you receive more naturally fits with the question you've posed.
It's helpful to keep in mind that {app_name} is entirely restricted to our database (see βAvailable Sourcesβ below). These sources are from the [OECD.AI](http://oecd.ai/) Database (containing national AI policies) and AI-related reports from the OECD iLibrary. If the answer to your question is not contained in these policy documents, the model won't be able to respond.
Since we restrict the model to information found in the documentation, it has a hard time with questions that require more generalized knowledge. Therefore, if you ask the model for information about AI policies in Asia, the model won't necessarily show you Japanese policy documentation. To overcome this limitation, it's best to be as specific as possible in your question, referencing the particular country you're looking for information on.
For more information about the tool's strengths and limitations, please see our website [here](https://mila.quebec/en/project/sai/).
"""
)
gr.Markdown(
f"""
## Risks
We have done our best to make sure that the AI algorithms are __only__ taking information from what is available in the OECD AI Observatory's Database; but, of course, Large Language Models (LLMs) are prone to fabrication. This means LLMs can make things up and present this made up information as if it were real, making it seem as if the information was found in a policy document. We therefore advise you to check the sources provided by the model to validate that the answer is in fact true. If you'd like to know exactly which documents the model can reference in its response, please see below.
## Recommended usage
{app_name} can only answer specific types of questions, for example:
* Questions about policy documents that are currently in the OECD AI Observatory's database
* Questions that are posed in English and target English language documents;
* Questions for which the answer can be found in the text (i.e. the thinking has already been done by the author) these AI models are not able to write their own research report combining information across policy documents and analyzing them itself).
If your question is outside the scope of the recommended use, the model has been instructed not to answer.
We are looking to create a tool that is as inclusive as possible.
While currently the tool only works with English language questions and documents we will continue assessing {app_name}'s capacity to perform as intended for users with different levels of fluency in English and plan to expand the functionality to ensure accessibility and impact across countries and user groups.
"""
)
return about_panel
def setup_terms_and_conditions():
with gr.Group(visible=True) as accept_terms_group:
with gr.Column(scale=1):
gr.Markdown(
f"""
By using this tool you agree to our {md_link_to_tncs}
""",
)
accept_checkbox = gr.Checkbox(value=0, label="I accept", interactive=True, container=False, scale=1)
return accept_terms_group, accept_checkbox
def setup_additional_sources():
# Display additional sources
gr.Markdown(f"")
gr.Markdown(
f"""## π Available sources
{app_name} has access to dozens of AI policy documents from various sources.
Below we list all of the sources that {app_name} has access to.
"""
)
with gr.Accordion(open=False, label="Click to list all available sources π"):
with gr.Column():
# Display the sources using a dataframe table
documents_metadata["Report"] = documents_metadata.apply(
lambda row: to_md_link(row["Title"], row["Link"]), axis=1
)
sub_df = documents_metadata[["Country", "Year", "Report"]]
gr.DataFrame(
sub_df, headers=list(sub_df.columns), interactive=False, datatype=["number", "str", "markdown"]
)
# Uncomment to display the sources instead as a simple markdown table
# gr.Markdown(get_metadata_markdown(documents_metadata))
def raise_flagging_message():
"""Raises a red banner indicating that the content has been flagged."""
gr.Info(
"Thank you for flagging the content. Our moderation team will look closely at these samples. We apologize for any harm this might have caused you."
)
def setup_flag_button():
"""Sets up a flag button with some accompanying text explaining why we have it."""
with gr.Column(variant="compact"):
gr.Markdown(
"""# Report bugs and harmful content
While we took many steps to ensure the tool is safe, we still rely on third parties for some of the model's capabilities. Please let us know if any harmful content shows up by clicking the button below and sending screenshots/concerns to [email protected]"""
)
flag_button = gr.Button(value="Flag content π©")
return flag_button
def setup_user_settings(
reformulate_question: bool, visible: bool, num_sources: int, max_sources: int = 15, min_sources: int = 1
) -> dict:
"""Set up user interface elements for frontend user settings in a web application.
This function creates an accordion containing a slider and a checkbox to configure
the number of sources and the option to reformulate questions, respectively.
The values set here will also be the values used by default by the app.
Args:
reformulate_question (bool): Initial state of the checkbox for reformulating questions.
visible (bool, optional): Visibility state of the settings tab. Defaults to False.
num_sources (int): Initial value for the number of sources slider. Defaults to 3.
max_sources (int, optional): Maximum limit for the number of sources slider. Defaults to 15.
min_sources (int, optional): Minimum limit for the number of sources slider. Defaults to 1.
Returns:
dict: A dictionary containing the UI elements for the reformulate question checkbox and the sources slider.
"""
with gr.Accordion(label=f"Settings βοΈ", open=False, visible=visible):
top_k_slider = gr.Slider(
minimum=min_sources,
maximum=max_sources,
interactive=True,
value=num_sources,
step=1,
label="Number of sources",
info="Number of documents to pass to the language model during its retrieval.",
)
reformulate_question_cbox = gr.Checkbox(
value=reformulate_question,
label="Reformulate Question (Beta)",
info="Reformulates a user's question to enhance source retrieval.",
)
settings_elems = {
"reformulate_question_cbox": reformulate_question_cbox,
"top_k_slider": top_k_slider,
}
return settings_elems
buster_app = gr.Blocks(css=css)
with buster_app:
# State variables are client-side and are reset every time a client refreshes the page
# Store the users' last completion here
last_completion = gr.State()
# A unique identifier that resets every time a page is refreshed
session_id = gr.State(get_session_id)
gr.Markdown(f"<h1><center>AIR: Q&A tool for AI Policy</center></h1>")
about_panel = setup_about_panel()
with gr.Row():
with gr.Column(scale=2, variant="panel"):
gr.Markdown(
f"""
Ask {app_name} your AI policy questions! Keep in mind this tool is a demo and can sometimes provide inaccurate information. Always verify the integrity of the information using the provided sources.
Since this tool is still in its early stages of development, please only engage with it as a demo.
"""
)
accept_terms_group, accept_terms_checkbox = setup_terms_and_conditions()
with gr.Row():
with gr.Column(scale=20):
user_input = gr.Textbox(
label="",
placeholder="Ask your AI policy questions here...",
value="β οΈ Accept the terms and conditions to use the app",
lines=1,
interactive=True,
)
submit = gr.Button(value="Ask", variant="primary", size="lg", interactive=False)
gr.Examples(
examples=example_questions,
inputs=user_input,
label=f"Sample questions to ask {app_name}",
)
chatbot = gr.Chatbot(label="Generated Answer", show_share_button=False)
sources_textboxes = display_sources()
with gr.Column():
settings_elems = setup_user_settings(
reformulate_question=cfg.reformulate_question,
visible=cfg.reveal_user_settings,
num_sources=cfg.buster_cfg.retriever_cfg["top_k"],
max_sources=cfg.max_sources,
)
top_k_slider = settings_elems["top_k_slider"]
feedback_elems = setup_feedback_form(top_k_slider.value)
flag_button = setup_flag_button()
top_k_slider.change(
set_relevant_sources_selection, inputs=top_k_slider, outputs=feedback_elems["relevant_sources_selection"]
)
setup_additional_sources()
gr.HTML(
f"""
<center>
<div style='margin-bottom: 20px;'> <!-- Add margin to the bottom of this div -->
Powered by <a href='https://github.com/jerpint/buster'>Buster</a> π€
</div>
<div>
<a href='{path_to_tncs}'> Terms And Conditions </a>
</div>
</center>
"""
)
# fmt: off
# Allow use of submit button and hide checkbox when accepted
accept_terms_checkbox.select(
reveal_app,
outputs=[accept_terms_group, user_input, submit]
)
gr.on(
triggers=[submit.click, user_input.submit],
fn=add_user_question,
inputs=[user_input],
outputs=[chatbot]
).then(
clear_user_input,
outputs=[user_input]
).then(
clear_sources,
outputs=[*sources_textboxes]
).then(
toggle_visibility,
inputs=gr.State(False),
outputs=feedback_elems["submitted_message"],
).then(
toggle_interactivity,
inputs=gr.State(True),
outputs=feedback_elems["submit_feedback_btn"],
).then(
clear_feedback_form,
outputs=[
feedback_elems["overall_experience"],
feedback_elems["clear_answer"],
feedback_elems["accurate_answer"],
feedback_elems["relevant_sources"],
feedback_elems["relevant_sources_selection"],
feedback_elems["relevant_sources_order"],
feedback_elems["expertise"],
feedback_elems["extra_info"],
]
).then(
chat,
inputs=[chatbot, settings_elems["reformulate_question_cbox"], settings_elems["top_k_slider"]],
outputs=[chatbot, last_completion],
).then(
add_disclaimer,
inputs=[last_completion, chatbot, gr.State(cfg.disclaimer)],
outputs=[chatbot]
).then(
add_sources,
inputs=[last_completion, gr.State(max_sources)],
outputs=[*sources_textboxes]
).then(
log_completion,
inputs=[last_completion, gr.State(cfg.MONGO_COLLECTION_INTERACTION), session_id]
)
flag_button.click(
log_completion,
inputs=[last_completion, gr.State(cfg.MONGO_COLLECTION_FLAGGED), session_id]
).then(
raise_flagging_message,
)
# fmt: on
|