Midterm proj

app.py

@@ -3,7 +3,7 @@ import chainlit as cl
 from dotenv import load_dotenv
 from operator import itemgetter
 from langchain_huggingface import HuggingFaceEndpoint
-from langchain_community.document_loaders import PyMuPDFLoader
+from PyPDF2 import PdfReader
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import Qdrant
 from langchain_openai.embeddings import OpenAIEmbeddings
@@ -35,21 +35,36 @@ VECTOR_STORE_PATH = "./data/vectorstore"
 3. Load HuggingFace Embeddings (remember to use the URL we set above)
 4. Index Files if they do not exist, otherwise load the vectorstore
 """
-document_loader = PyMuPDFLoader("./data/Airbnb_Q1_Filings.pdf")
-documents = document_loader.load()
-
-text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=30)
-split_documents = text_splitter.split_documents(documents)
+document_loader = PdfReader("./data/Airbnb_Q1_Filings.pdf")
+
+raw_text = ''
+for i, page in enumerate(document_loader.pages):
+    content = page.extract_text()
+    if content:
+        raw_text += content
+
+text_splitter = RecursiveCharacterTextSplitter(
+    chunk_size = 850,
+    chunk_overlap  = 50,
+    length_function = len,
+)
+texts = text_splitter.split_text(raw_text)
 
 # Note: Uses OPENAI_API_KEY env variable to make api calls
-openai_embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
-
-os.makedirs(VECTOR_STORE_PATH, exist_ok=True)
-vectorstore = Qdrant.from_documents(documents=split_documents,embedding=openai_embeddings,path=VECTOR_STORE_PATH,collection_name="airbnb_financials",batch_size=32,
-)
+hf_embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
+
+for i in range(0, len(texts), 32):
+    if i == 0:
+        vectorstore = Qdrant.from_texts(
+                texts[i:i+32],
+                hf_embeddings,
+                force_recreate=True,
+                path=VECTOR_STORE_PATH,
+                collection_name="Airbnb Filings")
+        continue
+    vectorstore.add_texts(texts[i:i+32])
 
 retriever = vectorstore.as_retriever()
-
 # -- AUGMENTED -- #
 """
 1. Define a String Template

backup.py

@@ -0,0 +1,138 @@
+import os
+import chainlit as cl
+from dotenv import load_dotenv
+from operator import itemgetter
+from langchain_huggingface import HuggingFaceEndpoint
+from langchain_community.document_loaders import PyMuPDFLoader
+from langchain_text_splitters import RecursiveCharacterTextSplitter
+from langchain_community.vectorstores import Qdrant
+from langchain_openai.embeddings import OpenAIEmbeddings
+from langchain_core.prompts import PromptTemplate
+from langchain.schema.runnable.config import RunnableConfig
+
+# GLOBAL SCOPE - ENTIRE APPLICATION HAS ACCESS TO VALUES SET IN THIS SCOPE #
+# ---- ENV VARIABLES ---- # 
+"""
+This function will load our environment file (.env) if it is present.
+
+NOTE: Make sure that .env is in your .gitignore file - it is by default, but please ensure it remains there.
+"""
+load_dotenv()
+
+"""
+We will load our environment variables here.
+"""
+HF_LLM_ENDPOINT = os.environ["HF_LLM_ENDPOINT"]
+HF_TOKEN = os.environ["HF_TOKEN"]
+VECTOR_STORE_PATH = "./data/vectorstore"
+
+# ---- GLOBAL DECLARATIONS ---- #
+
+# -- RETRIEVAL -- #
+"""
+1. Load Documents from PDF File
+2. Split Documents into Chunks
+3. Load HuggingFace Embeddings (remember to use the URL we set above)
+4. Index Files if they do not exist, otherwise load the vectorstore
+"""
+document_loader = PyMuPDFLoader("./data/Airbnb_Q1_Filings.pdf")
+documents = document_loader.load()
+
+text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=30)
+split_documents = text_splitter.split_documents(documents)
+
+# Note: Uses OPENAI_API_KEY env variable to make api calls
+openai_embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
+
+os.makedirs(VECTOR_STORE_PATH, exist_ok=True)
+vectorstore = Qdrant.from_documents(documents=split_documents,embedding=openai_embeddings,path=VECTOR_STORE_PATH,collection_name="airbnb_financials",batch_size=32,
+)
+
+retriever = vectorstore.as_retriever()
+
+# -- AUGMENTED -- #
+"""
+1. Define a String Template
+2. Create a Prompt Template from the String Template
+"""
+RAG_PROMPT_TEMPLATE = """\
+<|start_header_id|>system<|end_header_id|>
+You are a helpful assistant. You answer user questions based on provided context. If you can't answer the question with the provided context, say you don't know.<|eot_id|>
+
+<|start_header_id|>user<|end_header_id|>
+User Query:
+{query}
+
+Context:
+{context}<|eot_id|>
+
+<|start_header_id|>assistant<|end_header_id|>
+"""
+
+rag_prompt = PromptTemplate.from_template(RAG_PROMPT_TEMPLATE)
+
+# -- GENERATION -- #
+"""
+1. Create a HuggingFaceEndpoint for the LLM
+"""
+hf_llm = HuggingFaceEndpoint(
+    endpoint_url=HF_LLM_ENDPOINT,
+    max_new_tokens=512,
+    top_k=10,
+    top_p=0.95,
+    temperature=0.3,
+    repetition_penalty=1.15,
+    huggingfacehub_api_token=HF_TOKEN,
+)
+
+@cl.author_rename
+def rename(original_author: str):
+    """
+    This function can be used to rename the 'author' of a message. 
+
+    In this case, we're overriding the 'Assistant' author to be 'Paul Graham Essay Bot'.
+    """
+    rename_dict = {
+        "Assistant" : "AirBnB Auditor"
+    }
+    return rename_dict.get(original_author, original_author)
+
+@cl.on_chat_start
+async def start_chat():
+    """
+    This function will be called at the start of every user session. 
+
+    We will build our LCEL RAG chain here, and store it in the user session. 
+
+    The user session is a dictionary that is unique to each user session, and is stored in the memory of the server.
+    """
+
+    lcel_rag_chain = (
+        {"context": itemgetter("query") | retriever, "query": itemgetter("query")}
+        | rag_prompt | hf_llm
+    )
+
+    cl.user_session.set("lcel_rag_chain", lcel_rag_chain)
+
+@cl.on_message  
+async def main(message: cl.Message):
+    """
+    This function will be called every time a message is recieved from a session.
+
+    We will use the LCEL RAG chain to generate a response to the user query.
+
+    The LCEL RAG chain is stored in the user session, and is unique to each user session - this is why we can access it here.
+    """
+    lcel_rag_chain = cl.user_session.get("lcel_rag_chain")
+
+    msg = cl.Message(content="")
+
+    for chunk in await cl.make_async(lcel_rag_chain.stream)(
+        {"query": message.content},
+        config=RunnableConfig(callbacks=[cl.LangchainCallbackHandler()]),
+    ):
+        # Note: Skip printing eot_id token at the end of response. A more elegant solution would be to fix the model's generator config. 
+        if chunk != "<|eot_id|>":
+            await msg.stream_token(chunk)
+
+    await msg.send()