app.py

import os
import streamlit as st
from together import Together
from langchain_community.vectorstores import Chroma
from langchain_huggingface import HuggingFaceEmbeddings

# --- Configuration ---
# TogetherAI API key (env var name pilotikval)
TOGETHER_API_KEY = os.environ.get("pilotikval")
if not TOGETHER_API_KEY:
    st.error("Missing pilotikval environment variable.")
    st.stop()

# Initialize TogetherAI client
client = Together(api_key=TOGETHER_API_KEY)

# Embeddings setup
EMBED_MODEL_NAME = "BAAI/bge-base-en"
embeddings = HuggingFaceEmbeddings(
    model_name=EMBED_MODEL_NAME,
    encode_kwargs={"normalize_embeddings": True},
)

# Sidebar: select collection
st.sidebar.title("DocChatter RAG")
collection = st.sidebar.selectbox(
    "Choose a document collection:",
    ['General Medicine', 'RespiratoryFishman', 'RespiratoryMurray', 'MedMRCP2', 'OldMedicine']
)

dirs = {
    'General Medicine': './oxfordmedbookdir/',
    'RespiratoryFishman': './respfishmandbcud/',
    'RespiratoryMurray': './respmurray/',
    'MedMRCP2': './medmrcp2store/',
    'OldMedicine': './mrcpchromadb/'
}
cols = {
    'General Medicine': 'oxfordmed',
    'RespiratoryFishman': 'fishmannotescud',
    'RespiratoryMurray': 'respmurraynotes',
    'MedMRCP2': 'medmrcp2notes',
    'OldMedicine': 'mrcppassmednotes'
}

persist_directory = dirs[collection]
collection_name = cols[collection]

# Load Chroma vector store
vectorstore = Chroma(
    collection_name=collection_name,
    persist_directory=persist_directory,
    embedding_function=embeddings
)
retriever = vectorstore.as_retriever(search_kwargs={"k": 20})  # k=20

# System prompt template

def build_system(context: str) -> dict:
    """
    Build a comprehensive system prompt:
    - Act as an expert medical assistant and attentive listener.
    - Leverage retrieved context to craft detailed, accurate, and empathetic responses.
    - Ask clarifying follow-up questions if the user's query is ambiguous.
    - Structure answers clearly with headings, bullet points, and step-by-step explanations.
    - Cite relevant context sections when appropriate.
    - Maintain conversational memory for follow-up continuity.
    """
    prompt = f"""
You are a world-class medical assistant and conversational partner.

Listen carefully to the user’s questions, reference the context below, and provide a thorough, evidence-based response.
If any part of the question is unclear, ask a clarifying question before proceeding.
Organize your answer with clear headings or bullet points, and refer back to specific context snippets as needed.
Always be empathetic, concise, and precise in your medical explanations.
Retain memory of previous user messages to support follow-up interactions.

=== Retrieved Context Start ===
{context}
=== Retrieved Context End ===
"""
    return {"role": "system", "content": prompt}

st.title("🩺 DocChatter RAG (Streaming & Memory)")

# Initialize chat history
if 'chat_history' not in st.session_state:
    st.session_state.chat_history = []  # list of dicts {role, content}

# Get user input at top level
user_prompt = st.chat_input("Ask anything about your docs…")

# Tabs for UI
chat_tab, clear_tab = st.tabs(["Chat", "Clear History"])

with chat_tab:
    # Display existing chat
    for msg in st.session_state.chat_history:
        st.chat_message(msg['role']).write(msg['content'])

    # Handle new user input
    if user_prompt:
        # Echo user
        st.chat_message("user").write(user_prompt)
        st.session_state.chat_history.append({"role": "user", "content": user_prompt})

        # Retrieve top-k documents
        try:
            docs = retriever.invoke({"query": user_prompt})
        except Exception:
            docs = retriever.get_relevant_documents(user_prompt)
        context = "\n---\n".join([d.page_content for d in docs])

        # Build TogetherAI message sequence
        messages = [build_system(context)]
        for m in st.session_state.chat_history:
            messages.append(m)

        # Stream assistant response
        response_container = st.chat_message("assistant")
        stream_placeholder = response_container.empty()
        answer = ""

        for token in client.chat.completions.create(
            model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
            messages=messages,
            max_tokens=22048,
            temperature=0.1,
            stream=True
        ):
            try:
                choice = token.choices[0]
                delta = getattr(choice.delta, 'content', '')
                if delta:
                    answer += delta
                    stream_placeholder.write(answer)
            except (IndexError, AttributeError):
                continue

        # Save assistant response
        st.session_state.chat_history.append({"role": "assistant", "content": answer})

with clear_tab:
    if st.button("🗑️ Clear chat history"):
        st.session_state.chat_history = []
        st.experimental_rerun()

# (Optional) persist new docs
# vectorstore.persist()