Upload 5 files

app.py

@@ -0,0 +1,53 @@
+import gradio as gr
+
+from setup_database import get_document_store, add_data
+from setup_modules import create_retriever, create_readers_and_pipeline, text_reader_types, table_reader_types
+
+document_index = "document"
+document_store = get_document_store(document_index)
+filenames = ["processed_website_tables","processed_website_text","processed_schedule_tables"]
+document_store, data = add_data(filenames, document_store, document_index)
+document_store, retriever = create_retriever(document_store)
+text_reader_type = text_reader_types['deberta-large']
+table_reader_type = table_reader_types['tapas']
+pipeline = create_readers_and_pipeline(retriever, text_reader_type, table_reader_type, True, True)
+
+title = "Welcome to the BounWiki: The Question Answering Enginge for Bogazici Students!"
+
+head = '''
+This engine uses information from the Bogazici University Website to answer questions about different areas such as:
+
+ - Semester Dates (e.g. Registration Period, Add/Dropp Period...)
+ - Campus buildings and their locations
+ - General Uni Information, like Busses from Uni, Taxi-Numbers
+ - Schedule Information for all courses
+ 
+It returns the top 
+'''
+
+
+article = '''
+# How does this work?
+
+This App uses an "MPNet" sentence-transformer to encode information from the website into an embedding space.
+When faced with a query, the semantically most similar document is retrieved.
+A language model ("deberta-large" here) extracts the answer to the original question from this document and returns it to the interface 
+'''
+
+
+examples = [
+    ["When is the add/dropp period?"],
+    ["What does it mean if instructor consent is required?"],
+    ["Where is the english preparatory unit located?"],
+]
+
+label = gr.outputs.Label(num_top_classes=3)
+
+def predict(input):
+    prediction = pipeline.run(
+        query=input, params={"top_k": 3}
+        )
+    return {a.answer: float(a.score) for a in prediction["answers"]}
+
+interface = gr.Interface(fn=predict, inputs=gr.Textbox(lines=5, max_lines=6, label="Input Text"), outputs=label, title=title, description=head, article=article, examples=examples)
+interface.launch(server_name="0.0.0.0", server_port=8080)

eval_helper.py

@@ -0,0 +1,80 @@
+from haystack import Label, MultiLabel, Answer
+import json
+import re
+
+def read_labels(labels, tables):
+    processed_labels = []
+    for table in tables:
+        if table.id not in labels:
+            continue
+        doc_labels = labels[table.id]
+        for label in doc_labels:
+            label = Label(
+                query=label["question"],
+                document=table,
+                is_correct_answer=True,
+                is_correct_document=True,
+                answer=Answer(answer=label["answers"][0]["text"]),
+                origin="gold-label",
+            )
+            processed_labels.append(MultiLabel(labels=[label]))
+    return processed_labels
+
+def create_labels(labels_file, data, seperate_eval):
+  eval_labels = []
+  with open(labels_file) as labels_file:
+    labels = json.load(labels_file)
+  if seperate_eval:
+    use_labels = filter_labels(labels)
+  else:
+    use_labels = [labels]
+  for l in use_labels:
+    labels = []
+    for d in data:
+      labels += read_labels(l, d)
+    print(f"Number of Labels: {len(labels)}")
+    eval_labels.append(labels)
+  return eval_labels
+
+def get_processed_squad_labels(squad_labels):
+  with open(f'./data/validation_data/{squad_labels}') as fp:
+    squad_labels = json.load(fp)
+  # Process Squad File by aligning the right document IDs for the course schedules
+  processed_squad_labels = {}
+  for paragraph in squad_labels["data"]:
+    context = paragraph["paragraphs"][0]["context"]
+    if context[:43] == "Code\tName\tEcts\tInstructor\tDays\tHours\tRooms\n":
+      faculty_abb = re.search(r"[a-z]*", context[43:], re.IGNORECASE).group()
+      if faculty_abb in processed_squad_labels:
+        processed_squad_labels[faculty_abb].extend(paragraph["paragraphs"][0]["qas"])
+      else:
+        processed_squad_labels[faculty_abb] = paragraph["paragraphs"][0]["qas"]
+    else:
+      processed_squad_labels[str(paragraph["paragraphs"][0]["document_id"])] = paragraph["paragraphs"][0]["qas"]
+
+  with open("./data/validation_data/processed_qa.json", "w") as outfile:
+    json.dump(processed_squad_labels, outfile)
+  #return processed_squad_labels
+  
+def filter_labels(labels):
+  with open("./data/validation_data/questions_new.txt", "r") as fp:
+    user_questions = fp.read()
+
+  user_questions = user_questions.split("\n")
+  user_questions = [qu.strip() for qu in user_questions]
+  user_squad_labels = {}
+  synthetic_squad_labels = {}
+  for doc, questions in labels.items():
+    for q in questions:
+      if q["question"].strip() in user_questions:
+        if doc in user_squad_labels:
+          user_squad_labels[doc].append(q)
+        else:
+          user_squad_labels[doc] = [q]
+      else:
+        if doc in synthetic_squad_labels:
+          synthetic_squad_labels[doc].append(q)
+        else:
+          synthetic_squad_labels[doc] = [q]
+          
+  return [user_squad_labels, synthetic_squad_labels]

requirements.txt

@@ -0,0 +1,12 @@
+#git+https://github.com/deepset-ai/haystack.git#egg=farm-haystack
+farm-haystack
+fastapi
+requests
+torch==1.13.*
+transformers
+uvicorn[standard]
+grpcio
+numpy
+beautifulsoup4
+gradio
+

setup_database.py

@@ -0,0 +1,22 @@
+from haystack.document_stores import ElasticsearchDocumentStore
+from haystack.document_stores import InMemoryDocumentStore
+import os
+import pickle
+
+# def get_document_store(document_index):
+#     host = os.environ.get("ELASTICSEARCH_HOST", "localhost")
+#     document_store = ElasticsearchDocumentStore(host=host, username="", password="", index=document_index)
+#     return document_store
+
+def add_data(filenames, document_store, document_index):
+    data = []
+    for filename in filenames:
+        with open(f"./data/website_data/{filename}", "rb") as fp:
+            file = pickle.load(fp)
+            data.append(file)
+            document_store.write_documents(file, index=document_index)
+    return document_store, data
+
+def get_document_store(document_index):
+    document_store = InMemoryDocumentStore(index=document_index)
+    return document_store

setup_modules.py

@@ -0,0 +1,48 @@
+from haystack.nodes.retriever import EmbeddingRetriever
+from haystack.nodes import TableReader, FARMReader, RouteDocuments, JoinAnswers
+from haystack import Pipeline
+
+text_reader_types = {
+    "minilm": "deepset/minilm-uncased-squad2", 
+    "distilroberta": "deepset/tinyroberta-squad2", 
+    "electra-base": "deepset/electra-base-squad2", 
+    "bert-base": "deepset/bert-base-cased-squad2", 
+    "deberta-large": "deepset/deberta-v3-large-squad2", 
+    "gpt3": "implement openai answer generator"
+}
+table_reader_types = {
+    "tapas": "deepset/tapas-large-nq-hn-reader",
+    "text": "implement changing tables to text"
+}
+
+
+def create_retriever(document_store):
+    retriever = EmbeddingRetriever(document_store=document_store, embedding_model="deepset/all-mpnet-base-v2-table")
+    document_store.update_embeddings(retriever=retriever)
+    return document_store, retriever
+
+def create_readers_and_pipeline(retriever, text_reader_type = "deepset/roberta-base-squad2", table_reader_type="deepset/tapas-large-nq-hn-reader", use_table=True, use_text=True):
+    both = (use_table and use_text)
+    if use_text or both:
+        print("Initializing Text reader..")
+        text_reader = FARMReader(text_reader_type)
+    if use_table or both:
+        print("Initializing table reader..")
+        table_reader = TableReader(table_reader_type)
+    if both:
+        route_documents = RouteDocuments()
+        join_answers = JoinAnswers()
+    
+    text_table_qa_pipeline = Pipeline()
+    text_table_qa_pipeline.add_node(component=retriever, name="EmbeddingRetriever", inputs=["Query"])
+    if use_table and not use_text:
+        text_table_qa_pipeline.add_node(component=table_reader, name="TableReader", inputs=["EmbeddingRetriever"])
+    elif use_text and not use_table:
+        text_table_qa_pipeline.add_node(component=text_reader, name="TextReader", inputs=["EmbeddingRetriever"])
+    elif both:
+        text_table_qa_pipeline.add_node(component=route_documents, name="RouteDocuments", inputs=["EmbeddingRetriever"])
+        text_table_qa_pipeline.add_node(component=text_reader, name="TextReader", inputs=["RouteDocuments.output_1"])
+        text_table_qa_pipeline.add_node(component=table_reader, name="TableReader", inputs=["RouteDocuments.output_2"])
+        text_table_qa_pipeline.add_node(component=join_answers, name="JoinAnswers", inputs=["TextReader", "TableReader"])
+
+    return text_table_qa_pipeline