import streamlit as st import os import sys import io import re import pandas as pd import uuid import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt from typing import Annotated, TypedDict, Literal # LangGraph & LangChain from langgraph.graph import StateGraph , START,END from langgraph.graph.message import add_messages from langgraph.checkpoint.memory import MemorySaver from langgraph.types import interrupt, Command from langchain_google_genai import ChatGoogleGenerativeAI from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, SystemMessage # Supabase from supabase import create_client, Client # --- CONFIGURATION --- # Required for server-side plotting st.set_page_config(page_title="Autonomous Data Scientist", layout="wide") st.title("🤖 Autonomous Data Scientist Agent") # --- DATABASE SETUP --- @st.cache_resource def init_supabase(): url = os.environ.get("SUPABASE_URL") key = os.environ.get("SUPABASE_KEY") return create_client(url, key) if url and key else None supabase = init_supabase() def get_all_sessions(): """Fetch unique session IDs for the sidebar.""" if not supabase: return [] try: res = supabase.table("chat_history").select("session_id").execute() unique_ids = list(set([row['session_id'] for row in res.data])) return sorted(unique_ids, reverse=True) except: return [] def save_message(session_id, role, content): """Persist a single message to Supabase.""" if supabase: supabase.table("chat_history").insert({ "session_id": session_id, "role": role, "content": content }).execute() def get_messages_from_db(session_id): """Convert Supabase rows into LangChain Message objects for rehydration.""" if not supabase: return [] res = supabase.table("chat_history").select("role, content").eq("session_id", session_id).order("created_at", desc=False).execute() messages = [] for row in res.data: if row['role'] == 'user': messages.append(HumanMessage(content=row['content'])) elif row['role'] == 'assistant': messages.append(AIMessage(content=row['content'])) return messages # --- SIDEBAR & SESSION MANAGEMENT --- with st.sidebar: st.header("🗄️ Chat History") # New Chat Button if st.button("➕ New Chat", type="primary"): st.session_state["session_id"] = str(uuid.uuid4())[:8] st.session_state["messages"] = [] st.rerun() existing_sessions = get_all_sessions() # Ensure session_id exists if "session_id" not in st.session_state: st.session_state["session_id"] = existing_sessions[0] if existing_sessions else str(uuid.uuid4())[:8] # Handle Ghost IDs (ensure current ID is in list) current_id = st.session_state["session_id"] if current_id not in existing_sessions: existing_sessions.insert(0, current_id) # Session Selector selected = st.selectbox( "Load Conversation:", options=existing_sessions, index=existing_sessions.index(current_id) ) if selected != st.session_state["session_id"]: st.session_state["session_id"] = selected st.rerun() st.caption(f"ID: {st.session_state['session_id']}") st.divider() # Credentials & File Upload api_key = st.text_input("Gemini API Key", type="password") if api_key: os.environ["GOOGLE_API_KEY"] = api_key uploaded_file = st.file_uploader("Upload CSV", type=["csv"]) if uploaded_file: path = f"/tmp/{st.session_state['session_id']}_{uploaded_file.name}" with open(path, "wb") as f: f.write(uploaded_file.getbuffer()) st.session_state["dataset_path"] = path st.success(f"File Uploaded: {uploaded_file.name}") # --- AGENT LOGIC --- def local_executor(code): """Executes Python code in a restricted/sanitized environment.""" old_out = sys.stdout new_out = io.StringIO() sys.stdout = new_out # ---1. SECURITY FIX START --- if "agent_globals" not in st.session_state: # 1. Create a whitelist of safe libraries safe_globals = { "pd": pd, "plt": plt, "io": io, "np": __import__("numpy"), "re": re, "math": __import__("math"), "__builtins__": __builtins__ # Required for basic python to work } # 2. Add specific session variables they explicitly NEED if "dataset_path" in st.session_state: safe_globals["dataset_path"] = st.session_state["dataset_path"] st.session_state["agent_globals"] = safe_globals # --- 2. SECURITY FIX: STRING GUARDRAILS --- # Even with restricted globals, 'exec' allows imports. We must ban them explicitly. forbidden_terms = ["import os", "import sys", "from os", "from sys", "subprocess", "shutil"] if any(term in code for term in forbidden_terms): sys.stdout = old_out # Reset stdout before returning return "Security Error: Direct system imports (os, sys, subprocess) are not allowed.", None try: # Run code in the restricted 'safe_globals' instead of full 'globals()' exec(code, st.session_state["agent_globals"]) # (Rest of your plot capturing logic remains the same) if plt.get_fignums(): plot_buf = io.BytesIO() plt.savefig(plot_buf, format='png') plot_buf.seek(0) plt.close() return new_out.getvalue(), plot_buf return new_out.getvalue(), None except Exception as e: return f"Error: {e}", None finally: sys.stdout = old_out class State(TypedDict): messages: Annotated[list[BaseMessage], add_messages] # --- NODE 1: THE AGENT (THINKER) --- def agent_node(state: State): """ Responsibilities: 1. Check API Key. 2. Contextualize (System Prompt). 3. Generate the next message (Text or Code). """ if "GOOGLE_API_KEY" not in os.environ: return {"messages": [AIMessage("Please provide a Gemini API Key in the sidebar.")]} llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash", temperature=0) sys_msg = SystemMessage(content=f""" You are an Autonomous Data Scientist. 1. Output code strictly in ```python ... ``` blocks. 2. Always ask for permission before running model training (e.g., .fit(), joblib.dump). 3. Save models to '/tmp/model_{st.session_state['session_id']}.joblib'. 4. Dataset path is: {st.session_state.get('dataset_path', 'None')}. 5. If the user asks to plot, use matplotlib.pyplot. """) # Invoke LLM response = llm.invoke([sys_msg] + state["messages"]) return {"messages": [response]} # --- NODE 2: THE TOOL (DOER) --- def tool_node(state: State): """ Responsibilities: 1. Extract code from the last AIMessage. 2. Check Safety/Interrupts (HITL). 3. Execute code. 4. Return Output. """ last_message = state["messages"][-1] content = last_message.content # 1. Extract Code code_match = re.search(r"```python\n(.*?)```", content, re.DOTALL) if not code_match: return {"messages": [HumanMessage(content="Error: No code block found to execute.")]} code = code_match.group(1) # 2. Safety Check (Human-in-the-loop) # We interrupt strictly inside the Tool Node before execution if any(x in code for x in [".fit", "joblib.dump", "RandomForest", "sklearn"]): # This pauses the graph. The user must provide a value via Command(resume="...") decision = interrupt({"question": "Approve training code?", "code": code}) # Validate decision if str(decision).lower() not in ["yes", "approve"]: return {"messages": [HumanMessage(content=f"Action Blocked by User. Decision: {decision}")]} # 3. Execution text_out, plot_out = local_executor(code) # 4. Handle Plot Side-effect if plot_out: st.session_state["last_plot"] = plot_out return {"messages": [HumanMessage(content=f"Code Executed. Output: {text_out}\n[System: A plot was generated]")]} return {"messages": [HumanMessage(content=f"Code Executed. Output: {text_out}")]} # --- CONDITIONAL LOGIC (ROUTER) --- def should_continue(state: State) -> Literal["tools", "__end__"]: """ Determines if we need to run tools or stop. """ last_message = state["messages"][-1] # If the agent generated a Python block, go to tools if "```python" in last_message.content: return "tools" # Otherwise, stop and wait for user input return "__end__" # --- GRAPH BUILD --- if "graph" not in st.session_state: workflow = StateGraph(State) # Add Nodes workflow.add_node("agent", agent_node) workflow.add_node("tools", tool_node) # Set Entry Point workflow.set_entry_point("agent") # Add Edges # 1. From Agent, decide where to go (Tool or End) # THE CORRECT CODE workflow.add_conditional_edges( "agent", # 1. Start at 'agent' node should_continue, # 2. Run this function to get a string output { # 3. Map the string output to the next node "tools": "tools", "__end__": END } ) # 2. From Tool, always go back to Agent (to interpret results) workflow.add_edge("tools", "agent") st.session_state["graph"] = workflow.compile(checkpointer=MemorySaver()) app = st.session_state["graph"] config = {"configurable": {"thread_id": st.session_state["session_id"]}} # --- REHYDRATION LOGIC (The Fix) --- # Check if LangGraph memory is empty but DB has data current_graph_state = app.get_state(config) if not current_graph_state.values: db_msgs = get_messages_from_db(st.session_state["session_id"]) if db_msgs: # Inject DB history into LangGraph RAM app.update_state(config, {"messages": db_msgs}) # --- UI RENDER --- # Load display history from DB hist_rows = supabase.table("chat_history").select("*").eq("session_id", st.session_state["session_id"]).order("created_at", desc=False).execute().data if supabase else [] for m in hist_rows: with st.chat_message("user" if m["role"]=="user" else "assistant"): st.write(m["content"]) # If the message indicates a plot was made, we can't fetch the old plot from DB # (unless we stored it). For now, we only show live plots. # Show live plot if just generated if "last_plot" in st.session_state and st.session_state["last_plot"]: st.image(st.session_state["last_plot"], caption="Generated Plot") # Clear it so it doesn't persist forever on screen st.session_state["last_plot"] = None # --- CHAT INPUT --- if prompt := st.chat_input(): save_message(st.session_state["session_id"], "user", prompt) st.chat_message("user").write(prompt) with st.spinner("Agent is thinking..."): try: # Run the graph for event in app.stream({"messages": [HumanMessage(prompt)]}, config): if "agent" in event: msg = event["agent"]["messages"][-1] if isinstance(msg, AIMessage): save_message(st.session_state["session_id"], "assistant", msg.content) st.chat_message("assistant").write(msg.content) # If a plot was generated in this step, show it immediately if "last_plot" in st.session_state and st.session_state["last_plot"]: st.image(st.session_state["last_plot"]) st.session_state["last_plot"] = None except Exception as e: st.error(f"Execution Error: {str(e)}") # --- INTERRUPTS (APPROVALS) --- snap = app.get_state(config) if snap.next and snap.tasks[0].interrupts: intr = snap.tasks[0].interrupts[0].value st.warning("⚠️ APPROVAL NEEDED") st.write(f"**Question:** {intr.get('question')}") if intr.get('code'): st.code(intr.get('code'), language='python') col1, col2 = st.columns(2) if col1.button("✅ Approve"): with st.spinner("Resuming..."): for ev in app.stream(Command(resume="yes"), config): if "agent" in ev and isinstance(ev["agent"]["messages"][-1], AIMessage): m = ev["agent"]["messages"][-1] save_message(st.session_state["session_id"], "assistant", m.content) st.chat_message("assistant").write(m.content) st.rerun() if col2.button("❌ Reject"): with st.spinner("Rejecting..."): list(app.stream(Command(resume="no"), config)) st.rerun() # --- DOWNLOAD MODEL --- model_path = f"/tmp/model_{st.session_state['session_id']}.joblib" if os.path.exists(model_path): with open(model_path, "rb") as f: st.download_button("⬇️ Download Trained Model", f, f"model_{st.session_state['session_id']}.joblib")