Spaces:
Sleeping
Sleeping
| import os | |
| import re | |
| import streamlit as st | |
| from openai import OpenAI | |
| # Configure page style | |
| st.set_page_config(page_title="Clinical ReAct Agent — MIS 752", layout="wide") | |
| st.title("🩺 Clinical ReAct Agent & Supervisor Dashboard") | |
| st.markdown( | |
| "**MIS 752 · Lab 9: Build the Paperwork Machine** — Custom ReAct loop (`Thought → Action → Observation`) " | |
| "with `stop=['Observation:']` hallucination prevention and automated human-in-the-loop escalation policies." | |
| ) | |
| # Sidebar - Secrets & Settings | |
| st.sidebar.header("⚙️ Agent & Supervisor Settings") | |
| default_key = os.environ.get("OPENROUTER_API_KEY") or os.environ.get("OPENAI_API_KEY") or "" | |
| api_key = st.sidebar.text_input( | |
| "API Key (OpenRouter or OpenAI)", | |
| value=default_key, | |
| type="password", | |
| placeholder="sk-or-v1-... or sk-..." | |
| ) | |
| is_openrouter = api_key.strip().startswith("sk-or-") | |
| provider_label = "OpenRouter (https://openrouter.ai/api/v1)" if is_openrouter else "OpenAI Default" | |
| st.sidebar.caption(f"🔌 Active Endpoint: **{provider_label}**") | |
| if is_openrouter: | |
| model_options = [ | |
| "openai/gpt-4o-mini", | |
| "qwen/qwen3-8b", | |
| "meta-llama/llama-3.3-70b-instruct", | |
| "anthropic/claude-3.5-haiku" | |
| ] | |
| else: | |
| model_options = ["gpt-4o-mini", "gpt-4o", "gpt-3.5-turbo"] | |
| model_name = st.sidebar.selectbox("LLM Model Name", model_options) | |
| max_steps = st.sidebar.slider("Step Budget (Max Steps)", 2, 8, 6) | |
| supervision_policy = st.sidebar.selectbox( | |
| "Supervision Policy (Step 9.2)", | |
| ["flagged", "every", "audit"], | |
| help="Controls when the human supervisor is interrupted." | |
| ) | |
| # Clinic data structures | |
| REFERENCE = { | |
| "metformin": "Contraindicated when eGFR is below 30. Between 30 and 45, do not start it; if the patient is already taking it, halve the maximum to 1000 mg/day and recheck eGFR every 3 months.", | |
| "a1c target": "General adult target is 7.0 percent. For an older or frail patient, 8.0 percent is reasonable: a tighter target buys little and costs hypoglycaemia risk.", | |
| "ckd stage": "G1-G2 is eGFR 60 or above. G3a is 45-59. G3b is 30-44. G4 is 15-29. G5 is below 15, which is kidney failure.", | |
| "alert fatigue": "Degraded human response to a high volume of alarms, caused by the alarms themselves. Around 90 percent of drug-interaction alerts are overridden.", | |
| "lisinopril": "An ACE inhibitor. Monitor potassium and renal function. Hold it if the patient becomes acutely unwell or dehydrated.", | |
| } | |
| CHARTS = { | |
| "alvarez": "Alvarez, F, 68. Type 2 diabetes. eGFR 28. A1c 9.1 percent. Current: metformin 1000 mg twice daily, lisinopril 10 mg daily.", | |
| "okafor": "Okafor, M, 54. Hypertension. eGFR 71. A1c 5.8 percent. Current: lisinopril 20 mg daily.", | |
| "bianchi": "Bianchi, F, 81. Type 2 diabetes, frail. eGFR 39. A1c 8.4 percent. Current: metformin 500 mg twice daily.", | |
| } | |
| with st.sidebar.expander("📚 View Reference & Patient Charts"): | |
| st.markdown("**Clinic Reference Topics:** `" + "`, `".join(REFERENCE.keys()) + "`") | |
| st.markdown("**Patient Charts:** `" + "`, `".join(CHARTS.keys()) + "`") | |
| def lookup(topic: str) -> str: | |
| hit = REFERENCE.get(topic.strip().lower()) | |
| return hit if hit else f"'{topic}' is not in the clinic reference." | |
| def get_patient(name: str) -> str: | |
| hit = CHARTS.get(name.strip().lower()) | |
| return hit if hit else f"There is no patient called '{name}' on this list." | |
| TOOLS = {"lookup": lookup, "get_patient": get_patient} | |
| def run_tool(name, arg): | |
| fn = TOOLS.get(name) | |
| if fn is None: | |
| return f"There is no tool called '{name}'. Available: {', '.join(TOOLS)}.", False | |
| try: | |
| return fn(arg), True | |
| except Exception as e: | |
| return f"The tool failed: {type(e).__name__}: {e}", False | |
| def parse_action(text: str): | |
| tool = re.search(r"Action:\s*([a-zA-Z_]+)", text) | |
| arg = re.search(r"Action Input:\s*(.+)", text) | |
| if not tool: | |
| return None, None | |
| return tool.group(1).strip(), (arg.group(1).strip() if arg else "") | |
| def escalate_step(stage: dict, policy: str = "flagged") -> str: | |
| where = f"step {stage.get('step', '?')} ({stage.get('stage', '?')})" | |
| tool = stage.get("tool") or "no tool" | |
| ok = bool(stage.get("ok", True)) | |
| if policy == "every": | |
| return f"INTERRUPT - {where}, {tool}: policy is review-everything." | |
| elif policy == "audit": | |
| return f"LOG - {where}, {tool}: recorded for later review, nobody paged." | |
| elif policy == "flagged": | |
| if not ok: | |
| return f"INTERRUPT - {where}, {tool}: the agent reported a failure here." | |
| return f"SILENT - {where}, {tool}: nothing flagged." | |
| else: | |
| return f"INTERRUPT - unknown policy '{policy}'. Defaulting to interrupt." | |
| SYSTEM_PROMPT = '''You answer clinical questions by looking things up. You have two tools: | |
| lookup: look up one topic in the clinic reference. Input: a short topic, e.g. "metformin". | |
| get_patient: look up a patient chart by surname. Input: a surname, e.g. "Alvarez". | |
| ALWAYS answer in this exact format, ONE step at a time: | |
| Thought: what you know so far and what you need next | |
| Action: lookup | |
| Action Input: the topic to look up | |
| Then STOP and wait. You will be given an Observation with the result. | |
| Never write the Observation yourself. | |
| When you have enough to answer, write instead: | |
| Thought: I now know the final answer | |
| Final Answer: your answer, in one or two sentences | |
| /no_think''' | |
| preset = st.selectbox( | |
| "Choose a preset lab question or customize below:", | |
| [ | |
| "Is Alvarez still OK on her metformin? Check her chart first.", | |
| "What share of drug-interaction alerts get overridden, and why does that matter for a unit that fires 240 alerts a shift?", | |
| "What does our hospital night shift override policy say about 240 alerts in one night?", | |
| "Check Bianchi's chart and our a1c target and metformin rules. Does her regimen need adjustment?" | |
| ] | |
| ) | |
| query = st.text_input("Ask the clinical agent a question:", value=preset) | |
| if st.button("▶️ Run Agent", type="primary"): | |
| if not api_key: | |
| st.error("Please enter an OpenRouter (`sk-or-v1-...`) or OpenAI (`sk-...`) API Key in the sidebar to run the agent.") | |
| else: | |
| try: | |
| base_url = "https://openrouter.ai/api/v1" if is_openrouter else None | |
| client = OpenAI(api_key=api_key.strip(), base_url=base_url) | |
| messages = [{"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": query}] | |
| converged = False | |
| st.subheader("🧠 Agent Execution Trace") | |
| for step in range(1, max_steps + 1): | |
| with st.spinner(f"Thinking (Step {step})..."): | |
| r = client.chat.completions.create( | |
| model=model_name, | |
| messages=messages, | |
| temperature=0, | |
| max_tokens=1000, | |
| stop=["Observation:"] | |
| ) | |
| reply = r.choices[0].message.content or "" | |
| reply = re.sub(r"<think>.*?</think>", "", reply, flags=re.S) | |
| reply = re.sub(r"<think>.*$", "", reply, flags=re.S).strip() | |
| model_stage = {"step": step, "stage": "model", "text": reply, "tool": None, "ok": bool(reply)} | |
| with st.chat_message("assistant"): | |
| st.markdown(f"**Step {step} Reasoning** · `{escalate_step(model_stage, supervision_policy)}`") | |
| st.text(reply or "[empty reply]") | |
| if "Final Answer:" in reply: | |
| answer = reply.split("Final Answer:", 1)[1].strip() | |
| st.success(f"✅ **Final Answer (Converged in {step} steps):** {answer}") | |
| converged = True | |
| break | |
| tool, arg = parse_action(reply) | |
| if tool is None: | |
| obs, ok = "Could not parse action format. Please follow Action/Action Input syntax.", False | |
| else: | |
| obs, ok = run_tool(tool, arg) | |
| obs_stage = {"step": step, "stage": "observation", "text": obs, "tool": tool, "ok": ok} | |
| with st.chat_message("user"): | |
| st.markdown(f"**Step {step} Observation (Tool: `{tool}`)** · `{escalate_step(obs_stage, supervision_policy)}`") | |
| if ok: | |
| st.info(obs) | |
| else: | |
| st.error(obs) | |
| messages.append({"role": "assistant", "content": reply}) | |
| messages.append({"role": "user", "content": f"Observation: {obs}"}) | |
| if not converged: | |
| st.warning(f"⚠️ Step budget exhausted (`max_steps={max_steps}`) before reaching a Final Answer.") | |
| except Exception as e: | |
| st.error(f"An error occurred: {e}") |