yujunjun/haystack
1
0
Fork 0
mirror of https://github.com/deepset-ai/haystack.git synced 2025-11-01 18:29:32 +00:00
Code Issues Packages Projects Releases Wiki Activity

Public demo (#1747)

Browse Source 
* Queries now run only when pressing RUN. File upload hidden. Question is not sent if the textbox is empty.

* Add latest docstring and tutorial changes

* Tidy up: remove needless state, add comments, fix minor bugs

* Had to add results to the status to avoid some bugs in eval mode

* Added 'credits'

* Add footers, update requirements, some random questions for the evaluation

* Add requested changes

* Temporary rollback the UI to the old GoT dataset

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
This commit is contained in:
Sara Zan 2021-11-19 11:34:32 +01:00 committed by GitHub
parent c0892717a0
commit d81897535e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 158 additions and 118 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
rest_api/controller/search.py
View File

@ -4,6 +4,7 @@ from pathlib import Path
from fastapi import APIRouter
import haystack
from haystack.pipelines.base import Pipeline
from rest_api.config import PIPELINE_YAML_PATH, QUERY_PIPELINE_NAME
from rest_api.config import LOG_LEVEL, CONCURRENT_REQUEST_PER_WORKER
@ -42,6 +43,11 @@ def check_status():
return True
@router.get("/hs_version")
def haystack_version():
return {"hs_version": haystack.__version__}
@router.post("/query", response_model=QueryResponse, response_model_exclude_none=True)
def query(request: QueryRequest):
with concurrency_limiter.run():

4
ui/requirements.txt
View File

@ -1,2 +1,2 @@
streamlit>=0.84.0
st-annotated-text==1.1.0
streamlit>=1.2.0
st-annotated-text==2.0.0

11
ui/utils.py
View File

@ -6,6 +6,7 @@ import streamlit as st
API_ENDPOINT = os.getenv("API_ENDPOINT", "http://localhost:8000")
STATUS = "initialized"
HS_VERSION = "hs_version"
DOC_REQUEST = "query"
DOC_FEEDBACK = "feedback"
DOC_UPLOAD = "file-upload"
@ -20,8 +21,12 @@ def haystack_is_ready():
logging.exception(e)
return False
@st.cache
def haystack_version():
url = f"{API_ENDPOINT}/{HS_VERSION}"
return requests.get(url).json()["hs_version"]
@st.cache(show_spinner=False)
def retrieve_doc(query, filters=None, top_k_reader=5, top_k_retriever=5):
# Query Haystack API
url = f"{API_ENDPOINT}/{DOC_REQUEST}"
@ -31,6 +36,10 @@ def retrieve_doc(query, filters=None, top_k_reader=5, top_k_retriever=5):
# Format response
result = []
if "errors" in response_raw:
raise Exception(", ".join(response_raw["errors"]))
answers = response_raw["answers"]
for i in range(len(answers)):
answer = answers[i]

255
ui/webapp.py
View File

@ -3,6 +3,7 @@ import sys
import logging
import pandas as pd
from pathlib import Path
import streamlit as st
from annotated_text import annotated_text
@ -10,177 +11,201 @@ from annotated_text import annotated_text
# and every value gets lost. To keep track of our feedback state we use the official streamlit gist mentioned
# here https://gist.github.com/tvst/036da038ab3e999a64497f42de966a92
import SessionState
from utils import feedback_doc, haystack_is_ready, retrieve_doc, upload_doc
from utils import HS_VERSION, feedback_doc, haystack_is_ready, retrieve_doc, upload_doc, haystack_version
# Adjust to a question that you would like users to see in the search bar when they load the UI:
DEFAULT_QUESTION_AT_STARTUP = "Who is the father of Arya Stark?"
DEFAULT_QUESTION_AT_STARTUP = "Who's the father of Arya Stark?"
# Labels for the evaluation
EVAL_LABELS = os.getenv("EVAL_FILE", Path(__file__).parent / "eval_labels_example.csv")
def annotate_answer(answer, context):
""" If we are using an extractive QA pipeline, we'll get answers
from the API that we highlight in the given context"""
start_idx = context.find(answer)
end_idx = start_idx + len(answer)
# calculate dynamic height depending on context length
height = int(len(context) * 0.50) + 5
annotated_text(context[:start_idx], (answer, "ANSWER", "#8ef"), context[end_idx:], height=height)
# Whether the file upload should be enabled or not
DISABLE_FILE_UPLOAD = os.getenv("HAYSTACK_UI_DISABLE_FILE_UPLOAD")
def show_plain_documents(text):
""" If we are using a plain document search pipeline, i.e. only retriever, we'll get plain documents
from the API that we just show without any highlighting"""
st.markdown(text)
def random_questions(df):
"""
Helper to get one random question + gold random_answer from the user's CSV 'eval_labels_example'.
This can then be shown in the UI when the evaluation mode is selected. Users can easily give feedback on the
model's results and "enrich" the eval dataset with more acceptable labels
"""
random_row = df.sample(1)
random_question = random_row["Question Text"].values[0]
random_answer = random_row["Answer"].values[0]
return random_question, random_answer
# Retrieve Haystack version from the REST API
HS_VERSION = haystack_version()
def main():
# Define state
state_question = SessionState.get(
random_question=DEFAULT_QUESTION_AT_STARTUP, random_answer="", next_question="false", run_query="false"
# Persistent state
state = SessionState.get(
random_question=DEFAULT_QUESTION_AT_STARTUP,
random_answer="",
results=None,
raw_json=None,
get_next_question=True
)
# Initialize variables
eval_mode = False
random_question = DEFAULT_QUESTION_AT_STARTUP
eval_labels = os.getenv("EVAL_FILE", "eval_labels_example.csv")
# Small callback to reset the interface in case the text of the question changes
def reset_results(*args):
state.results = None
state.raw_json = None
# UI search bar and sidebar
# Title
st.write("# Haystack Demo")
# Sidebar
st.sidebar.header("Options")
top_k_reader = st.sidebar.slider("Max. number of answers", min_value=1, max_value=10, value=3, step=1)
top_k_retriever = st.sidebar.slider(
"Max. number of documents from retriever", min_value=1, max_value=10, value=3, step=1
)
top_k_retriever = st.sidebar.slider("Max. number of documents from retriever", min_value=1, max_value=10, value=3, step=1)
eval_mode = st.sidebar.checkbox("Evaluation mode")
debug = st.sidebar.checkbox("Show debug info")
st.sidebar.write("## File Upload:")
data_files = st.sidebar.file_uploader("", type=["pdf", "txt", "docx"], accept_multiple_files=True)
for data_file in data_files:
# Upload file
if data_file:
raw_json = upload_doc(data_file)
st.sidebar.write(raw_json)
if debug:
st.subheader("REST API JSON response")
st.sidebar.write(raw_json)
# File upload block
if not DISABLE_FILE_UPLOAD:
st.sidebar.write("## File Upload:")
data_files = st.sidebar.file_uploader("", type=["pdf", "txt", "docx"], accept_multiple_files=True)
for data_file in data_files:
# Upload file
if data_file:
raw_json = upload_doc(data_file)
st.sidebar.write(str(data_file.name) + " &nbsp;&nbsp; ✅ ")
if debug:
st.subheader("REST API JSON response")
st.sidebar.write(raw_json)
# load csv into pandas dataframe
st.sidebar.markdown(f"""
<style>
a {{
text-decoration: none;
}}
.haystack-footer {{
text-align: center;
}}
.haystack-footer h4 {{
margin: 0.1rem;
padding:0;
}}
footer {{
opacity: 0;
}}
</style>
<div class="haystack-footer">
<hr />
<h4>Built with <a href="https://www.deepset.ai/haystack">Haystack</a> <small>(v{HS_VERSION})</small></h4>
<p>Get it on <a href="https://github.com/deepset-ai/haystack/">GitHub</a> &nbsp;&nbsp; - &nbsp;&nbsp; Read the <a href="https://haystack.deepset.ai/overview/intro">Docs</a></p>
<small>Data crawled from <a href="https://en.wikipedia.org/wiki/Category:Lists_of_countries_by_continent">Wikipedia</a> in November 2021.<br />See the <a href="https://creativecommons.org/licenses/by-sa/3.0/">License</a> (CC BY-SA 3.0).</small>
</div>
""", unsafe_allow_html=True)
# Load csv into pandas dataframe
if eval_mode:
try:
df = pd.read_csv(eval_labels, sep=";")
df = pd.read_csv(EVAL_LABELS, sep=";")
except Exception:
sys.exit("The eval file was not found. Please check the README for more information.")
if (
state_question
and hasattr(state_question, "next_question")
and hasattr(state_question, "random_question")
and state_question.next_question
):
random_question = state_question.random_question
random_answer = state_question.random_answer
else:
random_question, random_answer = random_questions(df)
state_question.random_question = random_question
state_question.random_answer = random_answer
st.error(f"The eval file was not found. Please check the demo's [README](https://github.com/deepset-ai/haystack/tree/master/ui/README.md) for more information.")
sys.exit(f"The eval file was not found under `{EVAL_LABELS}`. Please check the README (https://github.com/deepset-ai/haystack/tree/master/ui/README.md) for more information.")
# Get next random question from the CSV
if eval_mode:
next_question = st.button("Load new question")
if next_question:
random_question, random_answer = random_questions(df)
state_question.random_question = random_question
state_question.random_answer = random_answer
state_question.next_question = True
state_question.run_query = False
else:
state_question.next_question = False
# Get next random question from the CSV
state.get_next_question = st.button("Load new question")
if state.get_next_question:
reset_results()
new_row = df.sample(1)
while new_row["Question Text"].values[0] == state.random_question: # Avoid picking the same question twice (the change is not visible on the UI)
new_row = df.sample(1)
state.random_question = new_row["Question Text"].values[0]
state.random_answer = new_row["Answer"].values[0]
# Search bar
question = st.text_input("Please provide your query:", value=random_question)
if state_question and state_question.run_query:
run_query = state_question.run_query
st.button("Run")
else:
run_query = st.button("Run")
state_question.run_query = run_query
raw_json_feedback = ""
question = st.text_input(
"Please provide your query:",
value=state.random_question,
max_chars=100,
on_change=reset_results
)
run_query = st.button("Run")
# Check the connection
with st.spinner("⌛️ &nbsp;&nbsp; Haystack is starting..."):
if not haystack_is_ready():
st.error("🚫 &nbsp;&nbsp; Connection Error. Is Haystack running?")
run_query = False
reset_results()
# Get results for query
if run_query:
if run_query and question:
reset_results()
with st.spinner(
"🧠 &nbsp;&nbsp; Performing neural search on documents... \n "
"Do you want to optimize speed or accuracy? \n"
"Check out the docs: https://haystack.deepset.ai/usage/optimization "
):
try:
results, raw_json = retrieve_doc(question, top_k_reader=top_k_reader, top_k_retriever=top_k_retriever)
state.results, state.raw_json = retrieve_doc(question, top_k_reader=top_k_reader, top_k_retriever=top_k_retriever)
except Exception as e:
logging.exception(e)
st.error("🐞 &nbsp;&nbsp; An error occurred during the request. Check the logs in the console to know more.")
if "The server is busy processing requests" in str(e):
st.error("🧑‍🌾 &nbsp;&nbsp; All our workers are busy! Try again later.")
else:
st.error("🐞 &nbsp;&nbsp; An error occurred during the request. Check the logs in the console to know more.")
return
# Show if we use a question of the given set
if question == random_question and eval_mode:
if state.results:
# Show the gold answer if we use a question of the given set
if question == state.random_question and eval_mode:
st.write("## Correct answers:")
random_answer
st.write(state.random_answer)
st.write("## Results:")
count = 0 # Make every button key unique
# Make every button key unique
count = 0
for result in results:
for result in state.results:
if result["answer"]:
annotate_answer(result["answer"], result["context"])
answer, context = result["answer"], result["context"]
start_idx = context.find(answer)
end_idx = start_idx + len(answer)
annotated_text(context[:start_idx], (answer, "ANSWER", "#8ef"), context[end_idx:])
else:
show_plain_documents(result["context"])
st.markdown(result["context"])
st.write("**Relevance:** ", result["relevance"], "**Source:** ", result["source"])
if eval_mode:
# Define columns for buttons
button_col1, button_col2, button_col3, button_col4 = st.columns([1, 1, 1, 6])
if button_col1.button("👍", key=(result["context"] + str(count) + "1"), help="Correct answer"):
raw_json_feedback = feedback_doc(
question, "true", result["document_id"], 1, "true", result["answer"], result["offset_start_in_doc"]
button_col1, button_col2, button_col3, _ = st.columns([1, 1, 1, 6])
if button_col1.button("👍", key=f"{result['context']}{count}1", help="Correct answer"):
feedback_doc(
question=question,
is_correct_answer="true",
document_id=result["document_id"],
model_id=1,
is_correct_document="true",
answer=result["answer"],
offset_start_in_doc=result["offset_start_in_doc"]
)
st.success("Thanks for your feedback")
if button_col2.button("👎", key=(result["context"] + str(count) + "2"), help="Wrong answer and wrong passage"):
raw_json_feedback = feedback_doc(
question,
"false",
result["document_id"],
1,
"false",
result["answer"],
result["offset_start_in_doc"],
st.success("✨ &nbsp;&nbsp; Thanks for your feedback! &nbsp;&nbsp; ✨")
if button_col2.button("👎", key=f"{result['context']}{count}2", help="Wrong answer and wrong passage"):
feedback_doc(
question=question,
is_correct_answer="false",
document_id=result["document_id"],
model_id=1,
is_correct_document="false",
answer=result["answer"],
offset_start_in_doc=result["offset_start_in_doc"]
)
st.success("Thanks for your feedback!")
if button_col3.button("👎👍", key=(result["context"] + str(count) + "3"), help="Wrong answer, but correct passage"):
raw_json_feedback = feedback_doc(
question, "false", result["document_id"], 1, "true", result["answer"], result["offset_start_in_doc"]
st.success("✨ &nbsp;&nbsp; Thanks for your feedback! &nbsp;&nbsp; ✨")
if button_col3.button("👎👍", key=f"{result['context']}{count}3", help="Wrong answer, but correct passage"):
feedback_doc(
question=question,
is_correct_answer="false",
document_id=result["document_id"],
model_id=1,
is_correct_document="true",
answer=result["answer"],
offset_start_in_doc=result["offset_start_in_doc"]
)
st.success("Thanks for your feedback!")
st.success("✨ &nbsp;&nbsp; Thanks for your feedback! &nbsp;&nbsp; ✨")
count += 1
st.write("___")
if debug:
st.subheader("REST API JSON response")
st.write(raw_json)
st.write(state.raw_json)
main()