app.py

import os
import streamlit as st
import nest_asyncio

# ─── 1) PATCH STREAMLIT’S EVENT LOOP ─────────────────────────────
nest_asyncio.apply()                   # allow nested awaits on Tornado’s loop :contentReference[oaicite:3]{index=3}
import asyncio
# No new_event_loop / set_event_loop here!
# We’ll grab the existing loop when we need it.

# ─── 2) LlamaIndex & Parser Imports ──────────────────────────────
from llama_index.core import StorageContext, load_index_from_storage
from llama_index.llms.openai import OpenAI
from llama_parse import LlamaParse
from llama_index.core import VectorStoreIndex
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.core.workflow import Event, StartEvent, StopEvent, Workflow, step, Context
from llama_index.core.memory import ChatMemoryBuffer

# ─── 3) Constants ─────────────────────────────────────────────────
PDF_PATH     = "./data/bank-of-america.pdf"
INDEX_DIR    = "./index_data"
SYSTEM_PROMPT = (
    "You are an expert analyst, who excels in analyzing a company's earnings call deck. "
    "Answer questions ONLY from the indexed document."
)

# ─── 4) Workflow Definition ────────────────────────────────────────
class ChatResponseEvent(Event):
    response: str
    memory: ChatMemoryBuffer

class ChatWorkflow(Workflow):
    @step
    async def answer(self, ev: StartEvent) -> ChatResponseEvent:
        storage = StorageContext.from_defaults(persist_dir=ev.index_dir)
        index   = load_index_from_storage(storage)
        chat_engine = index.as_chat_engine(
            chat_mode="context",
            memory=ev.memory,
            system_prompt=ev.system_prompt,
            llm=ev.llm
        )
        # Still using sync .chat(), but you could switch to an async method if available :contentReference[oaicite:4]{index=4}
        resp = chat_engine.chat(ev.query)
        return ChatResponseEvent(response=resp.response, memory=ev.memory)

    @step
    async def finalize(self, ev: ChatResponseEvent) -> StopEvent:
        return StopEvent(result=ev.response)

# ─── 5) Streamlit UI & Session State ───────────────────────────────
st.set_page_config(page_title="PDF Chatbot", layout="wide")
st.title("📄 Chat with Your PDF")

# Build or load the index
if "index_ready" not in st.session_state:
    os.makedirs(INDEX_DIR, exist_ok=True)
    index_meta = os.path.join(INDEX_DIR, "index_store.json")
    if os.path.isfile(index_meta):
        st.session_state.index_ready = True
        st.success("📚 Loaded existing index!")  # reuse existing index
    else:
        docs = LlamaParse(
            result_type="markdown",
            content_guideline_instruction="You are processing a company’s quarterly earnings-call slide deck. "
                                          "For each slide, produce a clearly sectioned Markdown fragment that includes:\n\n"
                                          "1. **Slide metadata**: slide number, title, and any subtitle or date\n"
                                          "2. **Key bullet points**: preserve existing bullets, but rewrite for clarity\n"
                                          "3. **Tables**: convert any tables into Markdown tables, capturing headers and all rows\n"
                                          "4. **Charts & graphs**: summarize each chart/graph in prose, highlighting axes labels, trends, and top 3 data points or percentage changes\n"
                                          "5. **Figures & images**: if there’s a figure caption, include it verbatim; otherwise, describe the visual in one sentence\n"
                                          "6. **Numeric callouts**: pull out any KPIs (revenue, EPS, growth rates) into a “Metrics” subsection\n"
                                          "7. **Overall slide summary**: a 1–2-sentence plain-English takeaway for the slide’s purpose or conclusion\n\n"
                                          "Keep the output strictly in Markdown, using headings (`##`, `###`), lists (`-`), and tables syntax. "
                                          "Do not include any LLM-specific commentary or markdown outside these rules."
        ).load_data(PDF_PATH)
        idx = VectorStoreIndex.from_documents(
            docs,
            embed_model=OpenAIEmbedding(model_name="text-embedding-3-small")
        )
        idx.storage_context.persist(persist_dir=INDEX_DIR)
        st.session_state.index_ready = True
        st.success("📚 Indexed your document and created index_store.json!")

# Initialize memory & workflow
if "memory" not in st.session_state:
    st.session_state.memory = ChatMemoryBuffer.from_defaults(
        llm=OpenAI(model="gpt-4o"), token_limit=1500
    )
if "workflow" not in st.session_state:
    st.session_state.workflow = ChatWorkflow(timeout=None, verbose=False)

# User input & async scheduling
user_input = st.text_input("Ask a question about the document:")
if user_input:
    # 1) Grab the running loop (patched by nest_asyncio)
    loop = asyncio.get_event_loop()  # returns Tornado’s loop :contentReference[oaicite:5]{index=5}

    # 2) Schedule the workflow.run coroutine on that loop
    future = asyncio.run_coroutine_threadsafe(
        st.session_state.workflow.run(
            index_dir=INDEX_DIR,
            query=user_input,
            system_prompt=SYSTEM_PROMPT,
            memory=st.session_state.memory,
            llm=OpenAI(model="gpt-4o")
        ),
        loop
    )

    # 3) Wait for the result (non-blocking at the loop level)
    stop_evt: StopEvent = future.result()  # avoids run_until_complete errors :contentReference[oaicite:6]{index=6}

    # 4) Update state & display
    st.session_state.memory = stop_evt.memory
    st.markdown(f"**Bot:** {stop_evt.result}")

# End Chat button
if st.button("End Chat"):
    st.write("Chat ended. Refresh to start over.")
    st.stop()