import os import sys import pandas as pd import streamlit as st from annotated_text import annotated_text # streamlit does not support any states out of the box. On every button click, streamlit reload the whole page # 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 from utils import retrieve_doc from utils import upload_doc # 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?" 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) 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 # Define state state_question = SessionState.get( random_question=DEFAULT_QUESTION_AT_STARTUP, random_answer="", next_question="false", run_query="false" ) # Initialize variables eval_mode = False random_question = DEFAULT_QUESTION_AT_STARTUP eval_labels = os.getenv("EVAL_FILE", "eval_labels_example.csv") # UI search bar and sidebar st.write("# Haystack Demo") 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 ) eval_mode = st.sidebar.checkbox("Evaluation mode") debug = st.sidebar.checkbox("Show debug info") st.sidebar.write("## File Upload:") data_file = st.sidebar.file_uploader("", type=["pdf", "txt", "docx"]) # 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) # load csv into pandas dataframe if eval_mode: try: 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 # 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" # 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 = "" # Get results for query if run_query: with st.spinner( "Performing neural search on documents... 🧠 \n " "Do you want to optimize speed or accuracy? \n" "Check out the docs: https://haystack.deepset.ai/docs/latest/optimizationmd " ): results, raw_json = retrieve_doc(question, top_k_reader=top_k_reader, top_k_retriever=top_k_retriever) # Show if we use a question of the given set if question == random_question and eval_mode: st.write("## Correct answers:") random_answer st.write("## Results:") # Make every button key unique count = 0 for result in results: if result["answer"]: annotate_answer(result["answer"], result["context"]) else: show_plain_documents(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.beta_columns([1, 1, 1, 6]) if button_col1.button("👍", key=(result["context"] + str(count)), help="Correct answer"): raw_json_feedback = feedback_doc( question, "true", result["document_id"], 1, "true", result["answer"], result["offset_start_in_doc"] ) st.success("Thanks for your feedback") if button_col2.button("👎", key=(result["context"] + str(count)), 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("Thanks for your feedback!") if button_col3.button("👎👍", key=(result["context"] + str(count)), 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("Thanks for your feedback!") count += 1 st.write("___") if debug: st.subheader("REST API JSON response") st.write(raw_json)