LOUIS SANNA
feat(data): add url
fe19632
raw
history blame
20.7 kB
import gradio as gr
from utils import create_user_id
# Langchain
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
# ClimateQ&A imports
from anyqa.config import get_domains
from anyqa.chains import load_qa_chain_with_text, load_reformulation_chain
from anyqa.embeddings import EMBEDDING_MODEL_NAME
from anyqa.llm import get_llm
from anyqa.prompts import audience_prompts
from anyqa.qa_logging import log
from anyqa.retriever import QARetriever
from anyqa.source_table import generate_source_table
from anyqa.vectorstore import get_vectorstore
# Load environment variables in local mode
try:
from dotenv import load_dotenv
load_dotenv()
except Exception as e:
pass
# Set up Gradio Theme
theme = gr.themes.Base(
primary_hue="blue",
secondary_hue="red",
font=[gr.themes.GoogleFont("Poppins"), "ui-sans-serif", "system-ui", "sans-serif"],
)
init_prompt = ""
system_template = {
"role": "system",
"content": init_prompt,
}
user_id = create_user_id()
# ---------------------------------------------------------------------------
# ClimateQ&A core functions
# ---------------------------------------------------------------------------
from langchain.callbacks.base import BaseCallbackHandler
from queue import Empty
from threading import Thread
from langchain.schema import LLMResult
from typing import Any, Union, Dict, List
from queue import SimpleQueue
# # Create a Queue
# Q = Queue()
import re
def parse_output_llm_with_sources(output):
# Split the content into a list of text and "[Doc X]" references
content_parts = re.split(r"\[(Doc\s?\d+(?:,\s?Doc\s?\d+)*)\]", output)
parts = []
for part in content_parts:
if part.startswith("Doc"):
subparts = part.split(",")
subparts = [
subpart.lower().replace("doc", "").strip() for subpart in subparts
]
subparts = [
f"<span class='doc-ref'><sup>{subpart}</sup></span>"
for subpart in subparts
]
parts.append("".join(subparts))
else:
parts.append(part)
content_parts = "".join(parts)
return content_parts
job_done = object() # signals the processing is done
class StreamingGradioCallbackHandler(BaseCallbackHandler):
def __init__(self, q: SimpleQueue):
self.q = q
def on_llm_start(
self, serialized: Dict[str, Any], prompts: List[str], **kwargs: Any
) -> None:
"""Run when LLM starts running. Clean the queue."""
while not self.q.empty():
try:
self.q.get(block=False)
except Empty:
continue
def on_llm_new_token(self, token: str, **kwargs: Any) -> None:
"""Run on new LLM token. Only available when streaming is enabled."""
self.q.put(token)
def on_llm_end(self, response: LLMResult, **kwargs: Any) -> None:
"""Run when LLM ends running."""
self.q.put(job_done)
def on_llm_error(
self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any
) -> None:
"""Run when LLM errors."""
self.q.put(job_done)
# Create embeddings function and LLM
embeddings_function = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME)
# Create vectorstore and retriever
vectorstore = get_vectorstore(embeddings_function)
# ---------------------------------------------------------------------------
# ClimateQ&A Streaming
# From https://github.com/gradio-app/gradio/issues/5345
# And https://stackoverflow.com/questions/76057076/how-to-stream-agents-response-in-langchain
# ---------------------------------------------------------------------------
from threading import Thread
def answer_user(query, query_example, history):
if len(query) <= 2:
raise Exception("Please ask a longer question")
return query, history + [[query, ". . ."]]
def answer_user_example(query, query_example, history):
return query_example, history + [[query_example, ". . ."]]
def fetch_sources(query, domains):
llm_reformulation = get_llm(
max_tokens=512, temperature=0.0, verbose=True, streaming=False
)
print("domains", domains)
retriever = QARetriever(
vectorstore=vectorstore, domains=domains, k_summary=0, k_total=10
)
reformulation_chain = load_reformulation_chain(llm_reformulation)
# Calculate language
output_reformulation = reformulation_chain({"query": query})
question = output_reformulation["question"]
language = output_reformulation["language"]
# Retrieve docs
docs = retriever.get_relevant_documents(question)
if len(docs) > 0:
# Already display the sources
sources_text = []
for i, d in enumerate(docs, 1):
sources_text.append(make_html_source(d, i))
citations_text = "".join(sources_text)
docs_text = "\n\n".join([d.page_content for d in docs])
return "", citations_text, docs_text, question, language
else:
sources_text = (
"⚠️ No relevant passages found in the scientific reports (IPCC and IPBES)"
)
citations_text = "**⚠️ No relevant passages found in the sources, you may want to ask a more specific question.**"
docs_text = ""
return "", citations_text, docs_text, question, language
def answer_bot(query, history, docs, question, language, audience):
if audience == "Children":
audience_prompt = audience_prompts["children"]
elif audience == "General public":
audience_prompt = audience_prompts["general"]
elif audience == "Experts":
audience_prompt = audience_prompts["experts"]
else:
audience_prompt = audience_prompts["experts"]
# Prepare Queue for streaming LLMs
Q = SimpleQueue()
llm_streaming = get_llm(
max_tokens=1024,
temperature=0.0,
verbose=True,
streaming=True,
callbacks=[StreamingGradioCallbackHandler(Q), StreamingStdOutCallbackHandler()],
)
qa_chain = load_qa_chain_with_text(llm_streaming)
def threaded_chain(question, audience, language, docs):
try:
response = qa_chain(
{
"question": question,
"audience": audience,
"language": language,
"summaries": docs,
}
)
Q.put(response)
Q.put(job_done)
except Exception as e:
print(e)
history[-1][1] = ""
textbox = gr.Textbox(
placeholder=". . .", show_label=False, scale=1, lines=1, interactive=False
)
if len(docs) > 0:
# Start thread for streaming
thread = Thread(
target=threaded_chain,
kwargs={
"question": question,
"audience": audience_prompt,
"language": language,
"docs": docs,
},
)
thread.start()
while True:
next_item = Q.get(block=True) # Blocks until an input is available
if next_item is job_done:
break
elif isinstance(next_item, str):
new_paragraph = history[-1][1] + next_item
new_paragraph = parse_output_llm_with_sources(new_paragraph)
history[-1][1] = new_paragraph
yield textbox, history
else:
pass
thread.join()
log(question=question, history=history, docs=docs, user_id=user_id)
else:
complete_response = "**⚠️ No relevant passages found in the sources, you may want to ask a more specific question.**"
history[-1][1] += complete_response
yield "", history
# ---------------------------------------------------------------------------
# ClimateQ&A core functions
# ---------------------------------------------------------------------------
def make_html_source(source, i):
meta = source.metadata
content = source.page_content.split(":", 1)[1].strip()
link = (
f'<a href="{meta["url"]}#page={int(meta["page_number"])}" target="_blank" class="pdf-link"><span role="img" aria-label="Open PDF">🔗</span></a>'
if "url" in meta
else ""
)
return f"""
<div class="card">
<div class="card-content">
<h2>Doc {i} - {meta['short_name']} - Page {int(meta['page_number'])}</h2>
<p>{content}</p>
</div>
<div class="card-footer">
<span>{meta['name']}</span>
{link}
</div>
</div>
"""
def reset_textbox():
return gr.update(value="")
# --------------------------------------------------------------------
# Gradio
# --------------------------------------------------------------------
init_prompt = """
Hello, I'm a conversational assistant. I will answer your questions by **sifting through trusted data sources**.
💡 How to use
- **Language**: You can ask me your questions in any language.
- **Audience**: You can specify your audience (children, general public, experts) to get a more adapted answer.
- **Sources**: You can choose to search in which sources you want me to look for answers. By default, I will search in all sources.
⚠️ Limitations
*Please note that the AI is not perfect and may sometimes give irrelevant answers. If you are not satisfied with the answer, please ask a more specific question or report your feedback to help us improve the system.*
❓ What do you want to learn ?
"""
def vote(data: gr.LikeData):
if data.liked:
print(data.value)
else:
print(data)
def change_tab():
return gr.Tabs.update(selected=1)
with gr.Blocks(title="❓ Q&A", css="style.css", theme=theme) as demo:
# user_id_state = gr.State([user_id])
with gr.Tab("❓ Q&A"):
with gr.Row(elem_id="chatbot-row"):
with gr.Column(scale=2):
# state = gr.State([system_template])
bot = gr.Chatbot(
value=[[None, init_prompt]],
show_copy_button=True,
show_label=False,
elem_id="chatbot",
layout="panel",
avatar_images=("assets/bot_avatar.png", None),
)
# bot.like(vote,None,None)
with gr.Row(elem_id="input-message"):
textbox = gr.Textbox(
placeholder="Ask me anything here!",
show_label=False,
scale=1,
lines=1,
interactive=True,
)
# submit_button = gr.Button(">",scale = 1,elem_id = "submit-button")
with gr.Column(scale=1, variant="panel", elem_id="right-panel"):
with gr.Tabs() as tabs:
with gr.TabItem("📝 Examples", elem_id="tab-examples", id=0):
examples_hidden = gr.Textbox(elem_id="hidden-message")
questions = [
"How does Daoism view our dependence on modern technology?",
"From a Confucian perspective, what is the role of tradition in modern society?",
"How might Daoism influence sustainable economic practices?",
"Does Confucianism advocate for a particular economic model?",
"How does Daoism interpret the dynamics of modern relationships?",
"From a Confucian viewpoint, what are the responsibilities of individuals in a family?",
"How might Daoism guide our approach to mental and physical health?",
"Does Confucianism offer insights into educational methods?",
"How does Daoism view the purpose and methods of modern education?",
"From a Confucian perspective, what is the importance of social harmony?",
]
examples_questions = gr.Examples(
questions,
[examples_hidden],
examples_per_page=10,
run_on_click=False,
# cache_examples=True,
)
with gr.Tab("📚 Citations", elem_id="tab-citations", id=1):
sources_textbox = gr.HTML(
show_label=False, elem_id="sources-textbox"
)
docs_textbox = gr.State("")
with gr.Tab("⚙️ Configuration", elem_id="tab-config", id=2):
gr.Markdown(
"Reminder: You can talk in any language, this tool is multi-lingual!"
)
domains = get_domains()
dropdown_domains = gr.CheckboxGroup(
domains,
label="Select source types",
value=[],
interactive=True,
)
dropdown_audience = gr.Dropdown(
["Children", "General public", "Experts"],
label="Select audience",
value="Experts",
interactive=True,
)
output_query = gr.Textbox(
label="Query used for retrieval",
show_label=True,
elem_id="reformulated-query",
lines=2,
interactive=False,
)
output_language = gr.Textbox(
label="Language",
show_label=True,
elem_id="language",
lines=1,
interactive=False,
)
(
textbox.submit(
answer_user,
[textbox, examples_hidden, bot],
[textbox, bot],
queue=False,
)
.success(change_tab, None, tabs)
.success(
fetch_sources,
[textbox, dropdown_domains],
[
textbox,
sources_textbox,
docs_textbox,
output_query,
output_language,
],
)
.success(
answer_bot,
[
textbox,
bot,
docs_textbox,
output_query,
output_language,
dropdown_audience,
],
[textbox, bot],
queue=True,
)
.success(lambda x: textbox, [textbox], [textbox])
)
(
examples_hidden.change(
answer_user_example,
[textbox, examples_hidden, bot],
[textbox, bot],
queue=False,
)
.success(change_tab, None, tabs)
.success(
fetch_sources,
[textbox, dropdown_domains],
[
textbox,
sources_textbox,
docs_textbox,
output_query,
output_language,
],
)
.success(
answer_bot,
[
textbox,
bot,
docs_textbox,
output_query,
output_language,
dropdown_audience,
],
[textbox, bot],
queue=True,
)
.success(lambda x: textbox, [textbox], [textbox])
)
# ---------------------------------------------------------------------------------------
# OTHER TABS
# ---------------------------------------------------------------------------------------
with gr.Tab("ℹ️ About", elem_classes="max-height"):
gr.Markdown(
"""
<div class="tip-box">
<div class="tip-box-title">
<span class="light-bulb" role="img" aria-label="Light Bulb">💡</span>
How does this tool work?
</div>
This tool harnesses modern OCR techniques to parse and preprocess documents. By leveraging state-of-the-art question-answering algorithms, <i>our tool is able to sift through the extensive collection of trusted sources and identify relevant passages in response to user inquiries</i>. Furthermore, the integration of the ChatGPT API allows Q&A to present complex data in a user-friendly manner, summarizing key points and facilitating communication to a wider audience.
</div>
"""
)
gr.Markdown("## How to use")
gr.Markdown(
"""
### 💪 Getting started
- In the chatbot section, simply type your question, and the app will provide an answer with references to relevant sources.
- the app retrieves specific passages to help answer your question accurately.
- Source information, including page numbers and passages, is displayed on the right side of the screen for easy verification.
- Feel free to ask follow-up questions within the chatbot for a more in-depth understanding.
- You can ask question in any language, the tool is multi-lingual !
"""
)
gr.Markdown(
"""
### ⚠️ Limitations
<div class="warning-box">
<ul>
<li>Please note that, like any AI, the model may occasionally generate an inaccurate or imprecise answer. Always refer to the provided sources to verify the validity of the information given. If you find any issues with the response, kindly provide feedback to help improve the system.</li>
</div>
"""
)
with gr.Tab("👩‍💻 Community"):
gr.Markdown(
"""
We welcomes community contributions.
To participate, head over to the Community Tab and create a "New Discussion" to ask questions and share your insights.
*This tool is a fork from the work done by the R&D lab at **Ekimetrics** for Climate Q&A: https://climateqa.com/.*
"""
)
with gr.Tab("📚 Sources", elem_classes="max-height"):
gr.Markdown(generate_source_table())
with gr.Tab("🛢️ Carbon Footprint"):
gr.Markdown(
"""
Carbon emissions were measured during the development and inference process using CodeCarbon [https://github.com/mlco2/codecarbon](https://github.com/mlco2/codecarbon)
| Phase | Description | Emissions | Source |
| --- | --- | --- | --- |
| Development | OCR and parsing all pdf documents with AI | 28gCO2e | CodeCarbon |
| Development | Question Answering development | 114gCO2e | CodeCarbon |
| Inference | Question Answering | ~0.102gCO2e / call | CodeCarbon |
| Inference | API call to turbo-GPT | ~0.38gCO2e / call | https://medium.com/@chrispointon/the-carbon-footprint-of-chatgpt-e1bc14e4cc2a |
Carbon Emissions are **relatively low but not negligible** compared to other usages: one question asked is around 0.482gCO2e - equivalent to 2.2m by car (https://datagir.ademe.fr/apps/impact-co2/)
Or around 2 to 4 times more than a typical Google search.
"""
)
with gr.Tab("🪄 Changelog"):
gr.Markdown(
"""
##### v1.0.0 - 2023-10-25
- Forked ClimateQ&A
- Added Chroma as vector store
- Added support for OpenAI api
- Added support for other topics
"""
)
demo.queue(concurrency_count=16)
demo.launch()