streamlit_app.py

# -*- coding: utf-8 -*-
# app.py

import os
import json
import pickle
import streamlit as st
from typing import List, Any, Literal
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
from langchain_core.messages import AIMessage

from pydantic import BaseModel, Field
from typing import Literal as TypingLiteral

from langchain_core.documents import Document
from langchain_core.messages import SystemMessage, HumanMessage

from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_openai import OpenAIEmbeddings, ChatOpenAI

from langchain.chat_models import init_chat_model
from langchain.tools import tool

from langchain_core.vectorstores import InMemoryVectorStore

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.prebuilt import ToolNode, tools_condition
from langgraph.checkpoint.memory import MemorySaver



# -----------------------------
# Streamlit page config
# -----------------------------
AGILE_ICON_URL = "https://raw.githubusercontent.com/RaziehAkbari2020/Agentic_RAG_Agile_Task_Management.Streamlit/main/Agile.png"

st.set_page_config(
    page_title="Agentic RAG for Agile Task Management",
    page_icon=AGILE_ICON_URL,
    layout="wide"
)

st.markdown(
    f"""
    <h1 style="display:flex; align-items:center; gap:12px; font-size:32px;">
        <img src="{AGILE_ICON_URL}" width="170">
        Agentic RAG for Agile Task Management
    </h1>
    """,
    unsafe_allow_html=True
)

st.caption(
    "Upload your preprocessed Agile project data (Taiga, Jira, GitHub, etc.), "
    "and chat with the AI Scrum Master assistant."
)


# -----------------------------
# Helpers: data loading
# -----------------------------
def _ensure_documents(obj: Any) -> List[Document]:
    if obj is None:
        return []

    if isinstance(obj, list) and obj and all(isinstance(x, Document) for x in obj):
        return obj

    if isinstance(obj, dict) and "documents" in obj:
        return _ensure_documents(obj["documents"])

    if isinstance(obj, list) and (len(obj) == 0 or isinstance(obj[0], dict)):
        docs: List[Document] = []
        for d in obj:
            page = d.get("page_content") or d.get("content") or d.get("text") or ""
            meta = d.get("metadata") or {}
            docs.append(Document(page_content=str(page), metadata=meta))
        return docs

    raise ValueError(
        "Uploaded data format not recognized. Expected a list of LangChain Documents, "
        "or a list of dicts with page_content/metadata, or a dict with key 'documents'."
    )


def load_uploaded_data(uploaded_file) -> List[Document]:
    name = uploaded_file.name.lower()

    if name.endswith(".pkl") or name.endswith(".pickle"):
        obj = pickle.load(uploaded_file)
        return _ensure_documents(obj)

    if name.endswith(".json"):
        obj = json.load(uploaded_file)
        return _ensure_documents(obj)

    if name.endswith(".jsonl"):
        lines = uploaded_file.getvalue().decode("utf-8").splitlines()
        arr = []
        for ln in lines:
            ln = ln.strip()
            if not ln:
                continue
            arr.append(json.loads(ln))
        return _ensure_documents(arr)

    raise ValueError("Unsupported file type. Please upload .pkl, .json, or .jsonl")


# -----------------------------
# Prompts
# -----------------------------
SYSTEM_PROMPT = """
You are an Agile assistant answering questions about Taiga project data.

You have access to a tool that retrieves user stories and tasks.

Use the tool when:

- the user asks about a specific user story
- the user asks about tasks, effort, assignees, sprint, story points
- the answer is not fully available in the conversation memory
- the question introduces new entities

Do NOT use the tool when:

- the question clearly refers to the immediately previous answer
- and all necessary information is already present in the conversation

Always prefer grounded answers based on retrieved context when uncertain.
""".strip()

GRADE_PROMPT = (
    "You are a grader assessing relevance of a retrieved Taiga Agile project document to a user question.\n\n"
    "The document is a Taiga user story and may contain:\n"
    "- User Story description\n"
    "- Tasks\n"
    "- Story points\n"
    "- Estimated effort\n"
    "- Actual effort\n"
    "- Assignees\n"
    "- Sprint order\n\n"
    "Retrieved document:\n"
    "{context}\n\n"
    "User question:\n"
    "{question}\n\n"
    "If the document contains information useful for answering the question "
    "(especially about tasks, effort, story points, assignment, sprint, or task management), "
    "grade it as relevant.\n\n"
    "Respond ONLY with 'yes' or 'no'."
)

REWRITE_PROMPT = (
    "You are rewriting a user question ONLY to improve retrieval over Taiga user stories and tasks.\n\n"
    "STRICT RULES:\n"
    "1) DO NOT change the meaning.\n"
    "2) DO NOT introduce new user story names, new entities, or new IDs.\n"
    "3) If the question mentions a specific user story, you MUST keep it exactly.\n"
    "4) If the question is already clear and retrieval-ready, return it unchanged.\n"
    "5) Keep it as ONE sentence. No lists, no extra commentary.\n\n"
    "Original question:\n"
    "{question}\n\n"
    "Rewritten question:"
)

GENERATE_PROMPT = (
    "You are an Agile assistant specialized in task planning and analysis.\n"
    "Use ONLY the retrieved context to answer.\n"
    "If the question involves tasks, provide a clear and structured plan when appropriate.\n"
    "NEVER infer missing fields.\n"
    "If the question asks for a field that is not explicitly present in the context, say 'Not provided in the data.'\n"
    "Do NOT treat Sprint Order as Priority unless the context explicitly says so.\n"
    "Provide a concise but structured answer.\n\n"
    "Question: {question}\n"
    "Context: {context}"
)

DIRECT_GENERATE_PROMPT = (
    "You are an Agile assistant specialized in task planning and analysis.\n"
    "Answer the user question directly.\n"
    "If the question requires project-specific data and no retrieved context is available, "
    "say that project data is needed.\n"
    "Do NOT invent project-specific details.\n"
    "Provide a concise but structured answer.\n\n"
    "Question: {question}"
)


class GradeDocuments(BaseModel):
    binary_score: TypingLiteral["yes", "no"] = Field(
        description="Relevance score: 'yes' if relevant, or 'no' if not relevant"
    )


def get_last_human_text(messages) -> str:
    for m in reversed(messages):
        if isinstance(m, HumanMessage):
            return m.content
    return messages[0].content if messages else ""

BASE_MODEL = "Qwen/Qwen3-0.6B"
LORA_MODEL = "Razieh87/AgileTaskGen-Agent-Qwen3-0.6B"


@st.cache_resource(show_spinner=True)
def load_qwen_ft_model():
    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)

    base_model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
        device_map="auto",
        trust_remote_code=True,
    )

    model = PeftModel.from_pretrained(base_model, LORA_MODEL)
    model.eval()

    return tokenizer, model


class QwenFTAnswerModel:
    def invoke(self, messages):
        tokenizer, model = load_qwen_ft_model()

        if isinstance(messages[-1], dict):
            user_prompt = messages[-1]["content"]
        else:
            user_prompt = messages[-1].content

        inputs = tokenizer(user_prompt, return_tensors="pt").to(model.device)

        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=512,
                do_sample=False,
                pad_token_id=tokenizer.eos_token_id,
            )

        text = tokenizer.decode(outputs[0], skip_special_tokens=True)

        if user_prompt in text:
            answer = text.replace(user_prompt, "").strip()
        else:
            answer = text.strip()

        return AIMessage(content=answer)
@st.cache_resource(show_spinner=False)
def build_graph_from_documents(docs: List[Document]):

    text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
        chunk_size=2500,
        chunk_overlap=100,
    )
    doc_splits = text_splitter.split_documents(docs)

    vectorstore = InMemoryVectorStore.from_documents(
        documents=doc_splits,
        embedding=OpenAIEmbeddings(),
    )
    retriever = vectorstore.as_retriever(search_kwargs={"k": 8})

    @tool
    def retrieve_user_stories(query: str) -> str:
        """Search and return relevant Taiga user stories and their tasks."""
        found = retriever.invoke(query)
        return "\n\n".join(d.page_content for d in found)

    retriever_tool = retrieve_user_stories

    # -----------------------------
    # Models
    # -----------------------------
    # OpenAI controls routing, tool calling, grading, and rewriting
    control_model = init_chat_model("gpt-4o-mini", temperature=0)
    answer_model = QwenFTAnswerModel()
    response_model = control_model
    grader_model = control_model

    # -----------------------------
    # Nodes
    # -----------------------------
    def generate_query_or_respond(state: MessagesState):
        messages = state["messages"]
        messages_with_system = [SystemMessage(content=SYSTEM_PROMPT), *messages]

        msg = response_model.bind_tools([retriever_tool]).invoke(messages_with_system)
        return {"messages": [msg]}

    def grade_documents(state: MessagesState) -> Literal["generate_answer", "rewrite_question"]:
        question = get_last_human_text(state["messages"])
        context = state["messages"][-1].content
        prompt = GRADE_PROMPT.format(question=question, context=context)

        response = grader_model.with_structured_output(GradeDocuments).invoke(
            [{"role": "user", "content": prompt}]
        )

        return "generate_answer" if response.binary_score == "yes" else "rewrite_question"

    def rewrite_question(state: MessagesState):
        question = get_last_human_text(state["messages"]).strip()
        prompt = REWRITE_PROMPT.format(question=question)

        rewritten = response_model.invoke(
            [{"role": "user", "content": prompt}]
        ).content.strip()

        if not rewritten:
            rewritten = question

        return {"messages": [HumanMessage(content=rewritten)]}

    def generate_answer(state: MessagesState):
        question = get_last_human_text(state["messages"])
        context = state["messages"][-1].content
        prompt = GENERATE_PROMPT.format(question=question, context=context)

        response = answer_model.invoke(
            [{"role": "user", "content": prompt}]
        )

        return {"messages": [response]}

    def generate_answer_without_context(state: MessagesState):
        question = get_last_human_text(state["messages"])
        prompt = DIRECT_GENERATE_PROMPT.format(question=question)

        response = answer_model.invoke(
            [{"role": "user", "content": prompt}]
        )

        return {"messages": [response]}

    # -----------------------------
    # Graph
    # -----------------------------
    workflow = StateGraph(MessagesState)

    workflow.add_node("generate_query_or_respond", generate_query_or_respond)
    workflow.add_node("retrieve", ToolNode([retriever_tool]))
    workflow.add_node("rewrite_question", rewrite_question)
    workflow.add_node("generate_answer", generate_answer)
    workflow.add_node("generate_answer_without_context", generate_answer_without_context)

    workflow.add_edge(START, "generate_query_or_respond")

    workflow.add_conditional_edges(
        "generate_query_or_respond",
        tools_condition,
        {
            "tools": "retrieve",
            END: "generate_answer_without_context",
        },
    )

    workflow.add_conditional_edges(
        "retrieve",
        grade_documents,
        {
            "generate_answer": "generate_answer",
            "rewrite_question": "rewrite_question",
        },
    )

    workflow.add_edge("generate_answer", END)
    workflow.add_edge("generate_answer_without_context", END)
    workflow.add_edge("rewrite_question", "generate_query_or_respond")

    memory = MemorySaver()
    graph = workflow.compile(checkpointer=memory)

    return graph


# -----------------------------
# Sidebar: API key + upload + settings
# -----------------------------
with st.sidebar:
    st.header("⚙️ Setup")

    import os

    try:
        api_key = st.secrets["OPENAI_API_KEY"]
        os.environ["OPENAI_API_KEY"] = api_key
    except Exception:
        api_key = ""

    if not api_key:
        st.error("OpenAI API key is not configured.")
        st.stop()



    uploaded = st.file_uploader(
        "Upload your preprocessed RAG dataset",
        type=["pkl", "pickle", "json", "jsonl"],
        help="Upload the dataset that is already prepared for RAG documents.",
    )

    show_retrieved = st.toggle("Show retrieved context", value=True)

    tool_preview_chars = st.slider(
        "Retrieved context preview chars",
        200,
        3000,
        900,
        50
    )

    thread_id = st.text_input(
        "Chat session",
        value=st.session_state.get("thread_id", "Task-thread-1"),
        help="Keeps conversation memory. Change it to start a new session."
    )

    st.session_state["thread_id"] = thread_id

    if st.button("🧹 Reset chat"):
        st.session_state.pop("chat_messages", None)
        st.session_state.pop("last_tool_context", None)
        st.session_state.pop("thread_id", None)
        st.rerun()


# -----------------------------
# Load docs + build graph
# -----------------------------
graph = None
docs = None
error = None

if not api_key:
    st.warning("Please enter your OpenAI API key in the sidebar.")
else:
    if uploaded is None:
        st.info("Upload your preprocessed Taiga dataset to start.")
    else:
        try:
            docs = load_uploaded_data(uploaded)
            st.success(f"Loaded {len(docs)} documents.")
            graph = build_graph_from_documents(docs)
        except Exception as e:
            error = str(e)
            st.error(f"Failed to load/build: {error}")


# -----------------------------
# Chat UI
# -----------------------------
if "chat_messages" not in st.session_state:
    st.session_state["chat_messages"] = []

if "last_tool_context" not in st.session_state:
    st.session_state["last_tool_context"] = ""

for m in st.session_state["chat_messages"]:
    with st.chat_message(m["role"]):
        st.markdown(m["content"])

prompt = st.chat_input(
    "Ask about tasks, effort estimation, assignment, workload, or progress..."
)

if prompt and graph is not None: 
    st.session_state["chat_messages"].append(
        {"role": "user", "content": prompt}
    )

    with st.chat_message("user"):
        st.markdown(prompt)

    config = {"configurable": {"thread_id": st.session_state["thread_id"]}}
    tool_context_accum = ""
    assistant_text = ""

    with st.chat_message("assistant"):
        placeholder = st.empty()

        for step in graph.stream(
            {"messages": [HumanMessage(content=prompt)]},
            stream_mode="values",
            config=config,
        ):
            last_msg = step["messages"][-1]
            msg_type = getattr(last_msg, "type", None)

            if msg_type == "tool":
                content = last_msg.content or ""

                if len(content) > tool_preview_chars:
                    content = content[:tool_preview_chars] + "\n… [truncated]"

                tool_context_accum = content
                continue

            if msg_type == "ai":
                assistant_text = last_msg.content or ""
                placeholder.markdown(assistant_text)

        if show_retrieved and tool_context_accum:
            with st.expander("🔎 Retrieved context (tool output)", expanded=False):
                st.code(tool_context_accum)

    st.session_state["chat_messages"].append(
        {"role": "assistant", "content": assistant_text}
    )

    st.session_state["last_tool_context"] = tool_context_accum