app.py

from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain.schema import StrOutputParser
from langchain.schema.runnable import Runnable
from langchain.schema.runnable.config import RunnableConfig
from dotenv import load_dotenv
import os
from langchain_community.document_loaders import PyMuPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Qdrant
import chainlit as cl
from sentence_transformers import SentenceTransformer

# Load environment variables
load_dotenv()
os.environ["OPENAI_API_KEY"] = os.getenv("OPENAI_API_KEY")

class SentenceTransformerEmbedding:
    def __init__(self, model_name):
        self.model = SentenceTransformer(model_name)

    def embed_documents(self, texts):
        return self.model.encode(texts, convert_to_tensor=True).tolist()  # Convert to list for compatibility

    def __call__(self, texts):
        return self.embed_documents(texts)  # Make it callable

@cl.on_chat_start
async def on_chat_start():
    model = ChatOpenAI(streaming=True)

    # Load documents
    ai_framework_document = PyMuPDFLoader(file_path="https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf").load()
    ai_blueprint_document = PyMuPDFLoader(file_path="https://www.whitehouse.gov/wp-content/uploads/2022/10/Blueprint-for-an-AI-Bill-of-Rights.pdf").load()

    print("Documents loaded.")

    RAG_PROMPT = """\
    Given a provided context and question, you must answer the question based only on context.

    Context: {context}
    Question: {question}
    """

    rag_prompt = ChatPromptTemplate.from_template(RAG_PROMPT)

    sentence_text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=500,
        chunk_overlap=100,
        separators=["\n\n", "\n", ".", "!", "?"]
    )

    def metadata_generator(document, name, splitter):
        collection = splitter.split_documents(document)
        for doc in collection:
            doc.metadata["source"] = name
        return collection

    sentence_framework = metadata_generator(ai_framework_document, "AI Framework", sentence_text_splitter)
    sentence_blueprint = metadata_generator(ai_blueprint_document, "AI Blueprint", sentence_text_splitter)

    sentence_combined_documents = sentence_framework + sentence_blueprint

    print(f"Total documents to embed: {len(sentence_combined_documents)}")

    # Limit the number of documents processed for debugging
    max_documents = 10
    sentence_combined_documents = sentence_combined_documents[:max_documents]

    # Initialize the embedding model instance
    embedding_model = SentenceTransformerEmbedding('Cheselle/finetuned-arctic-sentence')

    # Create the Qdrant vector store using the embedding instance
    sentence_vectorstore = Qdrant.from_documents(
        documents=sentence_combined_documents,
        embedding=embedding_model,
        location=":memory:",
        collection_name="AI Policy"
    )

    print("Vector store created.")

    # Create retriever from the vector store
    sentence_retriever = sentence_vectorstore.as_retriever()

    # Check if retriever is initialized correctly
    if sentence_retriever is None:
        raise ValueError("Retriever is not initialized correctly.")

    # Set the retriever and prompt into session for reuse
    cl.user_session.set("runnable", model)
    cl.user_session.set("retriever", sentence_retriever)
    cl.user_session.set("prompt_template", rag_prompt)

@cl.on_message
async def on_message(message: cl.Message):
    # Get the stored model, retriever, and prompt
    model = cl.user_session.get("runnable")
    retriever = cl.user_session.get("retriever")
    prompt_template = cl.user_session.get("prompt_template")

    # Log the message content
    print(f"Received message: {message.content}")

    # Retrieve relevant context from documents based on the user's message
    if retriever is None:
        print("Retriever is not available.")
        await cl.Message(content="Sorry, the retriever is not initialized.").send()
        return

    relevant_docs = retriever.get_relevant_documents(message.content)
    print(f"Retrieved {len(relevant_docs)} documents.")

    if not relevant_docs:
        print("No relevant documents found.")
        await cl.Message(content="Sorry, I couldn't find any relevant documents.").send()
        return

    context = "\n\n".join([doc.page_content for doc in relevant_docs])

    # Log the context to check
    print(f"Context: {context}")

    # Construct the final RAG prompt
    final_prompt = prompt_template.format(context=context, question=message.content)
    print(f"Final prompt: {final_prompt}")

    # Initialize a streaming message
    msg = cl.Message(content="")

    # Stream the response from the model
    async for chunk in model.astream(
        final_prompt,
        config=RunnableConfig(callbacks=[cl.LangchainCallbackHandler()]),
    ):
        await msg.stream_token(chunk.content)

    await msg.send()