app.py

import os
import sys
import logging
from pathlib import Path
from json import JSONDecodeError
import pandas as pd
import streamlit as st
from annotated_text import annotation
from markdown import markdown
import json
from haystack import Document
import pandas as pd
from haystack.document_stores import PineconeDocumentStore
from haystack.nodes import EmbeddingRetriever, FARMReader
from haystack.pipelines import ExtractiveQAPipeline
import shutil
import uuid
from pathlib import Path
from haystack.pipelines import Pipeline
from haystack.nodes import TextConverter, PreProcessor, FileTypeClassifier, PDFToTextConverter, DocxToTextConverter


preprocessor = PreProcessor(
    clean_empty_lines=True,
    clean_whitespace=True,
    clean_header_footer=False,
    split_by="word",
    split_length=100,
    split_respect_sentence_boundary=True
)
file_type_classifier = FileTypeClassifier()
text_converter = TextConverter()
pdf_converter = PDFToTextConverter()
docx_converter = DocxToTextConverter()


FILE_UPLOAD_PATH= "./data/uploads/"
os.makedirs(FILE_UPLOAD_PATH, exist_ok=True)
# @st.cache
def create_doc_store():
    document_store = PineconeDocumentStore(
        api_key= st.secrets["pinecone_apikey"],
        index='qa_demo',
        similarity="cosine",
        embedding_dim=768
    )
    return document_store

# @st.cache
# def create_pipe(document_store):
    # retriever = EmbeddingRetriever(
    # document_store=document_store,
    # embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
    # model_format="sentence_transformers",
    # )
    # reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2", use_gpu=False)
    # pipe = ExtractiveQAPipeline(reader, retriever)
    # return pipe

def query(pipe, question, top_k_reader, top_k_retriever):
    res = pipe.run(
        query=question, params={"Retriever": {"top_k": top_k_retriever}, "Reader": {"top_k": top_k_reader}}
    )
    answer_df = []
    # for r in res['answers']:
    #     ans_dict = res['answers'][0].meta
    #     ans_dict["answer"] = r.context
    #     answer_df.append(ans_dict)
    # result = pd.DataFrame(answer_df)
    # result.columns = ["Source","Title","Year","Link","Answer"]
    # result[["Answer","Link","Source","Title","Year"]] 
    return res

document_store = create_doc_store()
# pipe = create_pipe(document_store)
retriever = EmbeddingRetriever(
document_store=document_store,
embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
model_format="sentence_transformers",
)
reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2", use_gpu=False)
pipe = ExtractiveQAPipeline(reader, retriever)

indexing_pipeline_with_classification = Pipeline()
indexing_pipeline_with_classification.add_node(
    component=file_type_classifier, name="FileTypeClassifier", inputs=["File"]
)
indexing_pipeline_with_classification.add_node(
    component=text_converter, name="TextConverter", inputs=["FileTypeClassifier.output_1"]
)
indexing_pipeline_with_classification.add_node(
    component=pdf_converter, name="PdfConverter", inputs=["FileTypeClassifier.output_2"]
)
indexing_pipeline_with_classification.add_node(
    component=docx_converter, name="DocxConverter", inputs=["FileTypeClassifier.output_4"]
)
indexing_pipeline_with_classification.add_node(
    component=preprocessor,
    name="Preprocessor",
    inputs=["TextConverter", "PdfConverter", "DocxConverter"],
)
indexing_pipeline_with_classification.add_node(
    component=document_store, name="DocumentStore", inputs=["Preprocessor"]
)

def set_state_if_absent(key, value):
    if key not in st.session_state:
        st.session_state[key] = value

# Adjust to a question that you would like users to see in the search bar when they load the UI:
DEFAULT_QUESTION_AT_STARTUP = os.getenv("DEFAULT_QUESTION_AT_STARTUP", "My blog post discusses remote work. Give me statistics.")
DEFAULT_ANSWER_AT_STARTUP = os.getenv("DEFAULT_ANSWER_AT_STARTUP", "7% more remote workers have been at their current organization for 5 years or fewer")

# Sliders
DEFAULT_DOCS_FROM_RETRIEVER = int(os.getenv("DEFAULT_DOCS_FROM_RETRIEVER", "3"))
DEFAULT_NUMBER_OF_ANSWERS = int(os.getenv("DEFAULT_NUMBER_OF_ANSWERS", "3"))


st.set_page_config(page_title="Haystack Demo", page_icon="https://haystack.deepset.ai/img/HaystackIcon.png")

# Persistent state
set_state_if_absent("question", DEFAULT_QUESTION_AT_STARTUP)
set_state_if_absent("answer", DEFAULT_ANSWER_AT_STARTUP)
set_state_if_absent("results", None)


# Small callback to reset the interface in case the text of the question changes
def reset_results(*args):
    st.session_state.answer = None
    st.session_state.results = None
    st.session_state.raw_json = None

# Title
st.write("# Haystack Search Demo")
st.markdown(
    """
This demo takes its data from two sample data csv with statistics on various topics. \n
Ask any question on this topic and see if Haystack can find the correct answer to your query! \n
*Note: do not use keywords, but full-fledged questions.* The demo is not optimized to deal with keyword queries and might misunderstand you.
""",
    unsafe_allow_html=True,
)

# Sidebar
st.sidebar.header("Options")
st.sidebar.write("## File Upload:")
data_files = st.sidebar.file_uploader(
    "upload", type=["pdf", "txt", "docx"], accept_multiple_files=True, label_visibility="hidden"
)
ALL_FILES = []
for data_file in data_files:
    # Upload file
    if data_file:
        file_path = Path(FILE_UPLOAD_PATH) / f"{uuid.uuid4().hex}_{data_file.name}"
        with file_path.open("wb") as buffer:
            shutil.copyfileobj(data_file.file, buffer)
        ALL_FILES.append(file_path)
        st.sidebar.write(str(data_file.name) + "    ✅ ")
        indexing_pipeline_with_classification.run(file_paths=ALL_FILES)

if len(ALL_FILES) > 0:
    document_store.update_embeddings(retriever, update_existing_embeddings=False)
        
top_k_reader = st.sidebar.slider(
    "Max. number of answers",
    min_value=1,
    max_value=10,
    value=DEFAULT_NUMBER_OF_ANSWERS,
    step=1,
    on_change=reset_results,
)
top_k_retriever = st.sidebar.slider(
    "Max. number of documents from retriever",
    min_value=1,
    max_value=10,
    value=DEFAULT_DOCS_FROM_RETRIEVER,
    step=1,
    on_change=reset_results,
)
# data_files = st.file_uploader(
#         "upload", type=["csv"], accept_multiple_files=True, label_visibility="hidden"
#     )
# for data_file in data_files:
#     # Upload file
#     if data_file:
#         raw_json = upload_doc(data_file)

question = st.text_input(
        value=st.session_state.question,
        max_chars=100,
        on_change=reset_results,
        label="question",
        label_visibility="hidden",
    )
col1, col2 = st.columns(2)
col1.markdown("<style>.stButton button {width:100%;}</style>", unsafe_allow_html=True)
col2.markdown("<style>.stButton button {width:100%;}</style>", unsafe_allow_html=True)

# Run button
run_pressed = col1.button("Run")
if run_pressed:

    run_query = (
        run_pressed or question != st.session_state.question
    )
    # Get results for query
    if run_query and question:
        reset_results()
        st.session_state.question = question

        with st.spinner(
            "🧠 &nbsp;&nbsp; Performing neural search on documents... \n "
        ):
            try:
                st.session_state.results  = query(
                    pipe, question, top_k_reader=top_k_reader, top_k_retriever=top_k_retriever
                )
            except JSONDecodeError as je:
                st.error("👓 &nbsp;&nbsp; An error occurred reading the results. Is the document store working?")
            except Exception as e:
                logging.exception(e)
                if "The server is busy processing requests" in str(e) or "503" in str(e):
                    st.error("🧑‍🌾 &nbsp;&nbsp; All our workers are busy! Try again later.")
                else:
                    st.error(f"🐞 &nbsp;&nbsp; An error occurred during the request. {str(e)}")


if st.session_state.results:

    st.write("## Results:")

    for count, result in enumerate(st.session_state.results['answers']):
        answer, context = result.answer, result.context
        start_idx = context.find(answer)
        end_idx = start_idx + len(answer)
        source = f"[{result.meta['Title']}]({result.meta['link']})"
        # Hack due to this bug: https://github.com/streamlit/streamlit/issues/3190
        st.write(
            markdown(f'**Source:** {source} \n {context[:start_idx] } {str(annotation(answer, "ANSWER", "#8ef"))} {context[end_idx:]} \n '),
            unsafe_allow_html=True,
        )