First Cut

app.py

@@ -0,0 +1,122 @@
+import streamlit as st
+import langchain
+from langchain.document_loaders import OnlinePDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain.vectorstores import Pinecone
+from langchain.embeddings.openai import OpenAIEmbeddings
+import pinecone
+
+st.sidebar.markdown(" # Welcome to Ztudy ")
+
+# ------------------------ PDF ------------------------
+# Hard-coded PDFs (TODO: make this dynamic from Google Drive)
+pdf_dict = {}
+pdf_dict["Field Guide to Data Science"] = "https://wolfpaulus.com/wp-content/uploads/2017/05/field-guide-to-data-science.pdf"
+pdf_dict["2023 GPT-4 Technical Report"] = "https://cdn.openai.com/papers/gpt-4.pdf"
+pdf_dict["Administering Data Centers"] = "https://drive.google.com/file/d/1r3bqHq-ZszXnX6UJLOaeoEEa1plUYXZu"
+pdf_dict["First Aid Reference Guide (Google)"] = "https://drive.google.com/file/d/1fzN2wa_uJ8INUYim88eCymSvJdyDT2fz/"
+pdf_dict["First Aid Reference Guide (Public)"] = "https://www.sja.ca/sites/default/files/2021-05/First%20aid%20reference%20guide_V4.1_Public.pdf"
+pdf_dict["Astronomy 2106"] = "https://drive.google.com/file/d/1XXmjMLENP90-eXEqOaTxQ8O56ZwExsVT"
+pdf_dict["Astronomy 2106 (New)"] = "https://drive.google.com/file/d/1w1S-TY2PzeJ9mjPVb1yLwcYh5EI44oP7"
+pdf_dict["Learning Deep Learning: Chapter 1"] = "https://drive.google.com/file/d/1o7feaKFzXd5-95GffZyynAwY_fzGafhr/view?usp=sharing"
+
+# -------------------- Globals ------------------------
+texts = None
+pinecone_index = "group-1"
+
+if 'exchanges' not in st.session_state:
+    st.session_state.exchanges = []
+
+# -------------------- Functions -----------------------
+def console_log(msg):
+    st.sidebar.write(msg)
+
+def init_pinecone():
+    pinecone.init(
+        api_key=st.secrets["PINECONE_API_KEY"],  # find at app.pinecone.io
+        environment=st.secrets["PINECONE_API_ENV"]  # next to api key in console
+    ) 
+    return
+ 
+def load_vector_database():
+    embeddings = OpenAIEmbeddings(openai_api_key=st.secrets["OPENAI_API_KEY"])
+    init_pinecone()
+    print(f"Number of vectors: {len(texts)} to be upserted to Index: {pinecone_index}")
+    Pinecone.from_texts([t.page_content for t in texts], embeddings, index_name=pinecone_index)
+
+def load_pdf(url):
+    console_log(f"Loading {url}")
+    loader = OnlinePDFLoader(url)
+    data = loader.load()
+    console_log(f'You have {len(data)} document(s) in your data')
+    console_log(f'There are {len(data[0].page_content)} characters in your document')
+    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
+    global texts 
+    texts = text_splitter.split_documents(data)
+    console_log(f'After splitting, you have {len(texts)} documents')
+    load_vector_database()
+
+def chat(query):
+
+    from langchain.llms import OpenAI
+    from langchain.chains.question_answering import load_qa_chain
+
+    llm = OpenAI(temperature=0, openai_api_key=st.secrets["OPENAI_API_KEY"])
+    chain = load_qa_chain(llm, chain_type="stuff")
+   
+    embeddings = OpenAIEmbeddings(openai_api_key=st.secrets["OPENAI_API_KEY"])
+    init_pinecone()
+    vector_store = Pinecone.from_existing_index(pinecone_index, embeddings)
+    docs = vector_store.similarity_search(query, include_metadata=True)
+
+    # Comment/Uncomment to hide/show trace of documents
+    with st.expander("See documents for embedding"):
+        for i in range(len(docs)):
+            st.write(docs[i])
+    
+    return chain.run(input_documents=docs, question=query)
+
+def format_exchanges(exchanges):
+    for i in range(len(exchanges)):
+        if exchanges[i]["role"] == "user":
+            icon, text, blank = st.columns([1,8,1])
+        elif exchanges[i]["role"] == "assistant":
+            blank, text, icon = st.columns([1,8,1])
+        else:
+            st.markdown("*" + exchanges[i]["role"] + ":* " + exchanges[i]["content"]) 
+            continue
+        
+        with icon:
+            st.image("icon_" + exchanges[i]["role"] + ".png", width=50)
+        with text:
+            st.markdown(exchanges[i]["content"])
+        st.markdown("""---""")
+
+def format_prompt(exchanges):
+    # Include the last 6 exchanges
+    prompt = ""
+    for i in range( max(len(exchanges)-7,0), len(exchanges)):
+        prompt += "[Q]" if (exchanges[i]["role"] == "user") else "[A]"
+        prompt += ": " + exchanges[i]["content"] + "\n"
+    with st.expander("See prompt sent to LLM"):
+        st.write(prompt)
+    return prompt
+
+# ------------------------ Load PDF ------------------------
+with st.sidebar:
+    option = st.selectbox("Select a PDF", list(pdf_dict.keys()), key="pdf", on_change=None)
+    st.markdown(f"*Selected*: {option}") 
+    st.button('Click to start loading PDF', key="load_pdf", on_click=load_pdf, args=[pdf_dict[option]])
+
+# ------------------------ Chatbot ------------------------
+st.text_input("Prompt", placeholder="Ask me anything", key="prompt")
+
+if st.session_state.prompt:
+    st.session_state.exchanges.append({"role": "user", "content": st.session_state.prompt})
+    try:
+        response = chat(format_prompt(st.session_state.exchanges))
+    except Exception as e:
+        st.error(e) 
+        st.stop()
+    st.session_state.exchanges.append({"role": "assistant", "content": response})
+    format_exchanges(st.session_state.exchanges)

requirements.txt

@@ -0,0 +1,8 @@
+openai
+langchain
+streamlit
+unstructured
+unstructured[local-inference]
+pinecone-client
+tiktoken
+nltk

run.sh

@@ -0,0 +1,4 @@
+# To run the app, run the following command from the project root directory:
+# sh run.sh
+# For Windonws, need to have a .bat file to run the app
+$PWD/.venv/bin/streamlit run app.py