src/streamlit_app.py

import os
import re
import json
import traceback
import streamlit as st
from pathlib import Path
from typing import List, Annotated, Any
import chromadb
import operator
import tempfile
from tqdm import tqdm
from pydantic import BaseModel
from langchain.embeddings.cohere import CohereEmbeddings
from langchain_cohere import ChatCohere
from langchain.document_loaders import DirectoryLoader, TextLoader, PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
import cohere
from langchain_core.messages import AnyMessage, SystemMessage, HumanMessage, AIMessage
from langgraph.graph import StateGraph, START, END, add_messages
from langgraph.constants import Send
from langgraph.checkpoint.memory import MemorySaver

chromadb.api.client.SharedSystemClient.clear_system_cache()

COHERE_API_KEY = os.environ["COHERE_API_KEY"]
co = cohere.Client(COHERE_API_KEY)

documents_path = Path(__file__).parent / "documents"
persist_dir = tempfile.mkdtemp()

def prepare_vectorstore(uploaded_files=None):
    documents = []

    if uploaded_files and any(file.size > 0 for file in uploaded_files):
        st.write("📁 Uploaded files:")
        for file in uploaded_files:
            st.write(f"• {file.name} ({file.size} bytes)")
            file_path = Path(tempfile.gettempdir()) / file.name
            try:
                with open(file_path, "wb") as f:
                    f.write(file.getbuffer())
                st.write(f"✅ Saved to: {file_path}")

                if file.name.endswith(".pdf"):
                    st.write(f"📄 Loading PDF: {file.name}")
                    loader = PyPDFLoader(str(file_path))
                elif file.name.endswith(".txt"):
                    st.write(f"📃 Loading TXT: {file.name}")
                    loader = TextLoader(str(file_path))
                else:
                    st.warning(f"Unsupported file type: {file.name}")
                    continue

                loaded = loader.load()
                st.write(f"Loaded {len(loaded)} pages from {file.name}")
                documents.extend(loaded)

            except Exception as e:
                st.error(f"Error loading {file.name}:")
                st.exception(e)
                st.text(traceback.format_exc())

    else:
        st.warning("No uploaded files found or all were empty.")
        st.stop()

    if not documents:
        st.error("No content could be loaded from the uploaded files.")
        st.stop()

    st.write("Splitting documents into chunks...")
    splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=50)
    docs = splitter.split_documents(documents)
    st.write(f"Total chunks created: {len(docs)}")

    if not docs:
        st.error("No content found in the documents after splitting.")
        st.stop()

    st.write("Embedding documents...")
    embedding = CohereEmbeddings(
        model="embed-multilingual-light-v3.0",
        cohere_api_key=COHERE_API_KEY,
        user_agent="langgraph-app"
    )

    try:
        vectorstore = Chroma.from_documents(
            documents=tqdm(docs, desc="Embedding"),
            embedding=embedding,
            persist_directory=persist_dir
        )
        vectorstore.persist()
        st.success("Document embedding complete.")
        return vectorstore
    except Exception as e:
        st.error("Embedding failed:")
        st.exception(e)
        st.text(traceback.format_exc())
        st.stop()

class State(BaseModel):
    state: List[str] = []
    messages: Annotated[list[AnyMessage], add_messages]
    topic: List[str] = []
    context: List[str] = []
    sub_topic_list: List[str] = []
    sub_topics: Annotated[list[AnyMessage], add_messages]
    stories: Annotated[list[AnyMessage], add_messages]
    stories_lst: Annotated[list, operator.add]

class StoryState(BaseModel):
    retrieved_docs: List[Any] = []
    stories: Annotated[list[AnyMessage], add_messages]
    reranked_docs: List[str] = []
    story_topic: str = ""
    stories_lst: Annotated[list, operator.add]

def extract_topics(messages):
    topics = []
    for message in messages:
        topics.extend(re.findall(r'- \*\*(.*?)\*\*', message.content))
    return topics

embedding_llm = CohereEmbeddings(
    model="embed-multilingual-light-v3.0",
    cohere_api_key=COHERE_API_KEY,
    user_agent="langgraph-app"
)

llm = ChatCohere(
    api_version="2024-02-15-preview",
    temperature=0.7,
    model="command-r-plus-08-2024",
    cohere_api_key=COHERE_API_KEY
)

beginner_topic_sys_msg = SystemMessage(content="Suppose you are a middle grader who wants to learn constantly about new topics to get a good score in exams.")
middle_topic_sys_msg = SystemMessage(content="Suppose you are a college student who wants to learn constantly about new topics to get a good score in exams.")
advanced_topic_sys_msg = SystemMessage(content="Suppose you are a teacher who wants to learn constantly about new topics to teach your students.")

def retrieve_node(state):
    topic = state.story_topic
    query = f"information about {topic}"
    retriever = Chroma(persist_directory=persist_dir, embedding_function=embedding_llm).as_retriever(search_kwargs={"k": 20})
    docs = retriever.get_relevant_documents(query)
    return {"retrieved_docs": docs, "question": query}

def rerank_node(state):
    topic = state.story_topic
    query = f"Rerank documents based on how good they explain the topic {topic}"
    docs = state.retrieved_docs
    texts = [doc.page_content for doc in docs]
    rerank_results = co.rerank(query=query, documents=texts, top_n=5, model="rerank-v3.5")
    top_docs = [texts[result.index] for result in rerank_results.results]
    return {"reranked_docs": top_docs, "question": query}

def generate_story_node(state):
    context = "\n\n".join(state.reranked_docs)
    topic = state.story_topic
    system_message = """
    Suppose you're an amazing story writer and scientific thinker. 
    You've written hundreds of story books explaining scientific topics in a childlike manner.
    You add a subtle humor to your stories to make them more engaging.
    """
    prompt = f"""
    Use the following context to generate a simple engaging story that explains {topic} in such a way a middle schooler can understand it.

    Context:
    {context}

    Story:
    """
    response = llm.invoke([SystemMessage(system_message), HumanMessage(prompt)])
    return {"stories": response}

def beginner_topic(state: State):
    prompt = f"What are the beginner-level topics you can learn about {', '.join(state.topic)} in {', '.join(state.context)}?"
    sub_topics = [llm.invoke([beginner_topic_sys_msg] + [prompt])]
    return {"message": sub_topics[0], "sub_topics": sub_topics[0]}

def middle_topic(state: State):
    prompt = f"What are the middle-level topics you can learn about {', '.join(state.topic)} in {', '.join(state.context)}? Don't include the topics below:\n\n{(state.sub_topics)}"
    sub_topics = [llm.invoke([middle_topic_sys_msg] + [prompt])]
    return {"message": sub_topics, "sub_topics": sub_topics}

def advanced_topic(state: State):
    prompt = f"What are the advanced-level topics you can learn about {', '.join(state.topic)} in {', '.join(state.context)}? Don't include the topics below:\n\n{(state.sub_topics)}"
    sub_topics = [llm.invoke([advanced_topic_sys_msg] + [prompt])]
    return {"message": sub_topics, "sub_topics": sub_topics}

def topic_extractor(state: State):
    return {"sub_topic_list": extract_topics(state.sub_topics)}

def dynamic_topic_edges(state: State):
    return [Send("story_generator", {"story_topic": topic}) for topic in state.sub_topic_list]

story_builder = StateGraph(StoryState)
story_builder.add_node("Retrieve", retrieve_node)
story_builder.add_node("Rerank", rerank_node)
story_builder.add_node("Generate", generate_story_node)
story_builder.set_entry_point("Retrieve")
story_builder.add_edge("Retrieve", "Rerank")
story_builder.add_edge("Rerank", "Generate")
story_builder.set_finish_point("Generate")

story_graph = story_builder.compile()

main_builder = StateGraph(State)
main_builder.add_node("beginner_topic", beginner_topic)
main_builder.add_node("middle_topic", middle_topic)
main_builder.add_node("advanced_topic", advanced_topic)
main_builder.add_node("topic_extractor", topic_extractor)
main_builder.add_node("story_generator", story_graph)
main_builder.add_edge(START, "beginner_topic")
main_builder.add_edge("beginner_topic", "middle_topic")
main_builder.add_edge("middle_topic", "advanced_topic")
main_builder.add_edge("advanced_topic", "topic_extractor")
main_builder.add_conditional_edges("topic_extractor", dynamic_topic_edges, ["story_generator"])
main_builder.add_edge("story_generator", END)

memory = MemorySaver()
react_graph = main_builder.compile(checkpointer=memory, interrupt_after=["topic_extractor"])

st.title("LangGraph Topic Story Generator")

uploaded_files = st.file_uploader(
    "Upload .txt or .pdf files",
    type=["txt", "pdf"],
    accept_multiple_files=True,
    key="file_uploader"
)

if uploaded_files:
    st.session_state["files"] = uploaded_files
    st.success(f"{len(uploaded_files)} file(s) uploaded:")
    for file in uploaded_files:
        st.write(f"• {file.name} ({file.size} bytes)")
elif "files" in st.session_state:
    st.info("Using previously uploaded files:")
    for file in st.session_state["files"]:
        st.write(f"• {file.name} ({file.size} bytes)")
else:
    st.info("No files uploaded yet.")

topic = st.text_input("Enter a topic", "Human Evolution")
context = st.text_input("Enter a context", "Science")

if st.button("Generate Stories"):
    uploaded = st.session_state.get("files")
    if not uploaded or all(file.size == 0 for file in uploaded):
        st.warning("You uploaded files, but they appear to be empty.")
        st.stop()

    try:
        prepare_vectorstore(uploaded)
        thread = {"configurable": {"thread_id": "1"}}
        react_graph.invoke({"topic": [topic], "context": [context]}, thread)
        react_graph.update_state(thread, {"sub_topic_list": ['Early Hominins', 'Fossil Evidence', "Darwin's Theory of Evolution"]})
        result = react_graph.invoke(None, thread, stream_mode="values")
        for story in result["stories"]:
            st.markdown(story.content)
    except Exception as e:
        st.error("Something went wrong during story generation.")
        st.exception(e)
        st.text(traceback.format_exc())