app.py

import os
from typing import TypedDict, Annotated, Union, List, Tuple
from typing_extensions import List, TypedDict
from dotenv import load_dotenv
import chainlit as cl
import nest_asyncio
import getpass
from uuid import uuid4
from langchain_community.document_loaders import DirectoryLoader
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_qdrant import QdrantVectorStore
from qdrant_client import QdrantClient
from qdrant_client.http.models import Distance, VectorParams
from langchain.retrievers.contextual_compression import ContextualCompressionRetriever
from langchain_cohere import CohereRerank
from langchain.prompts import ChatPromptTemplate
from langgraph.graph import START, StateGraph
from typing_extensions import List, TypedDict
from langchain_core.runnables import Runnable
from langchain_core.tools import tool
from langgraph.graph.message import add_messages
import operator
from langchain_core.messages import BaseMessage
from langchain_core.documents import Document
from langgraph.prebuilt import ToolNode
from langgraph.graph import StateGraph, END
from langchain_core.messages import HumanMessage
from ragas.embeddings import LangchainEmbeddingsWrapper
from langchain_openai import ChatOpenAI
from ragas.testset import TestsetGenerator
from operator import itemgetter
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough, RunnableParallel
from ragas import EvaluationDataset
from ragas.llms import LangchainLLMWrapper
from ragas.metrics import LLMContextRecall, Faithfulness, FactualCorrectness, ResponseRelevancy, ContextEntityRecall, NoiseSensitivity
from ragas import evaluate, RunConfig
from langchain_community.document_loaders import BSHTMLLoader
import tqdm
import asyncio
import json
from sentence_transformers import SentenceTransformer
from torch.utils.data import DataLoader
from torch.utils.data import Dataset
from sentence_transformers import InputExample
from sentence_transformers.losses import MatryoshkaLoss, MultipleNegativesRankingLoss
from sentence_transformers.evaluation import InformationRetrievalEvaluator
from huggingface_hub import notebook_login
import pandas as pd

load_dotenv()


with open("./data_new/combined_data.json", "r") as f:
    raw_data = json.load(f)

all_docs = [
    Document(page_content=entry["content"], metadata=entry["metadata"])
    for entry in raw_data]


tavily_tool = TavilySearchResults(max_results=5)


text_splitter = RecursiveCharacterTextSplitter(chunk_size=750, chunk_overlap=100, length_function = len)
split_documents = text_splitter.split_documents(all_docs)


embeddings = HuggingFaceEmbeddings(model_name = "bsmith3715/legal-ft-demo_final")


client = QdrantClient(":memory:")

client.create_collection(
    collection_name="reformer",
    vectors_config=VectorParams(size=768, distance=Distance.COSINE),
)

vector_store_ft = QdrantVectorStore(
    client=client,
    collection_name="reformer",
    embedding=embeddings,
)

_ = vector_store_ft.add_documents(documents=split_documents)

retriever_finetune = vector_store_ft.as_retriever(search_kwargs={"k": 5})

def retrieve_adjusted(state):
  compressor = CohereRerank(model="rerank-v3.5", top_n=10)
  compression_retriever = ContextualCompressionRetriever(
    base_compressor=compressor, base_retriever=retriever_finetune, search_kwargs={"k": 5}
  )
  retrieved_docs = compression_retriever.invoke(state["question"])
  return {"context" : retrieved_docs}

RAG_PROMPT = """\
You are a helpful pilates expert who answers questions based on provided context. You must only use the provided context, and cannot use your own knowledge.

### Question
{question}

### Context
{context}
"""

rag_prompt = ChatPromptTemplate.from_template(RAG_PROMPT)

ft_llm = ChatOpenAI(model="gpt-4o-mini")

def generate(state):
  docs_content = "\n\n".join(doc.page_content for doc in state["context"])
  messages = rag_prompt.format_messages(question=state["question"], context=docs_content)
  response = ft_llm.invoke(messages)
  return {"response" : response.content}

class State(TypedDict):
  question: str
  context: List[Document]
  response: str

graph_builder = StateGraph(State).add_sequence([retrieve_adjusted, generate])
graph_builder.add_edge(START, "retrieve_adjusted")
graph = graph_builder.compile()

cert_model = ChatOpenAI(
    model="gpt-4o",
    temperature=0
)

tool_belt = [
    tavily_tool,
]

cert_model = cert_model.bind_tools(tool_belt)

class AgentState(TypedDict):
    messages: Annotated[list, add_messages]
    context: List[Document]

def call_model(state):
  messages = state["messages"]
  response = cert_model.invoke(messages)
  return {"messages" : [response],
          "context" : state.get("context", [])
    }

tool_node = ToolNode(tool_belt)

uncompiled_graph = StateGraph(AgentState)

uncompiled_graph.add_node("agent", call_model)
uncompiled_graph.add_node("action", tool_node)

uncompiled_graph.set_entry_point("agent")

def should_continue(state):
  last_message = state["messages"][-1]

  if last_message.tool_calls:
    return "action"

  return END

uncompiled_graph.add_conditional_edges(
    "agent",
    should_continue
)

uncompiled_graph.add_edge("action", "agent")

compiled_graph_w_fine_tune = uncompiled_graph.compile()

@cl.on_chat_start
async def on_chat_start():
    await cl.Message(content=
"""👋 Welcome to your Reformer Pilates AI!
Here's what you can do:
• Ask questions about Reformer Pilates
• Get individualized workouts based on your level, goals, and equipment
• Get instant exercise modifications based on injuries or limitations
Let's get started! 🚀""").send()
    cl.user_session.set("graph", compiled_graph_w_fine_tune)
    
@cl.on_message
async def handle(message: cl.Message):
    graph = cl.user_session.get("graph")       
    state = {"messages" : [HumanMessage(message.content)]}
    response = await graph.ainvoke(state)
    await cl.Message(content=response["messages"][-1].content, stream = True).send()