Spaces:
Sleeping
Sleeping
Upload app.py with huggingface_hub
Browse files
app.py
CHANGED
|
@@ -1,18 +1,48 @@
|
|
| 1 |
-
|
| 2 |
-
import streamlit as st
|
| 3 |
import re
|
|
|
|
| 4 |
from openai import OpenAI
|
| 5 |
|
| 6 |
# Configure page style
|
| 7 |
-
st.set_page_config(page_title="Clinical ReAct Agent", layout="wide")
|
| 8 |
-
st.title("🩺 Clinical ReAct Agent Dashboard")
|
| 9 |
-
st.markdown(
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
# Sidebar - Secrets & Settings
|
| 12 |
-
st.sidebar.header("⚙️ Settings")
|
| 13 |
-
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
max_steps = st.sidebar.slider("Step Budget (Max Steps)", 2, 8, 6)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
# Clinic data structures
|
| 18 |
REFERENCE = {
|
|
@@ -29,6 +59,10 @@ CHARTS = {
|
|
| 29 |
"bianchi": "Bianchi, F, 81. Type 2 diabetes, frail. eGFR 39. A1c 8.4 percent. Current: metformin 500 mg twice daily.",
|
| 30 |
}
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
def lookup(topic: str) -> str:
|
| 33 |
hit = REFERENCE.get(topic.strip().lower())
|
| 34 |
return hit if hit else f"'{topic}' is not in the clinic reference."
|
|
@@ -55,7 +89,22 @@ def parse_action(text: str):
|
|
| 55 |
return None, None
|
| 56 |
return tool.group(1).strip(), (arg.group(1).strip() if arg else "")
|
| 57 |
|
| 58 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
SYSTEM_PROMPT = '''You answer clinical questions by looking things up. You have two tools:
|
| 60 |
|
| 61 |
lookup: look up one topic in the clinic reference. Input: a short topic, e.g. "metformin".
|
|
@@ -76,16 +125,27 @@ Thought: I now know the final answer
|
|
| 76 |
Final Answer: your answer, in one or two sentences
|
| 77 |
/no_think'''
|
| 78 |
|
| 79 |
-
|
| 80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
|
| 82 |
-
if st.button("Run Agent"):
|
| 83 |
if not api_key:
|
| 84 |
-
st.error("Please enter an OpenAI API Key in the sidebar to run the agent.")
|
| 85 |
else:
|
| 86 |
try:
|
| 87 |
-
|
|
|
|
| 88 |
messages = [{"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": query}]
|
|
|
|
| 89 |
|
| 90 |
st.subheader("🧠 Agent Execution Trace")
|
| 91 |
|
|
@@ -95,20 +155,23 @@ if st.button("Run Agent"):
|
|
| 95 |
model=model_name,
|
| 96 |
messages=messages,
|
| 97 |
temperature=0,
|
| 98 |
-
max_tokens=
|
| 99 |
stop=["Observation:"]
|
| 100 |
)
|
| 101 |
reply = r.choices[0].message.content or ""
|
| 102 |
-
reply = re.sub(r"<think>.*?</think>", "", reply, flags=re.S)
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
-
# Render model thinking step
|
| 105 |
with st.chat_message("assistant"):
|
| 106 |
-
st.markdown(f"**Step {step} Reasoning**")
|
| 107 |
-
st.text(reply)
|
| 108 |
|
| 109 |
if "Final Answer:" in reply:
|
| 110 |
answer = reply.split("Final Answer:", 1)[1].strip()
|
| 111 |
-
st.success(f"✅ **Final Answer:** {answer}")
|
|
|
|
| 112 |
break
|
| 113 |
|
| 114 |
tool, arg = parse_action(reply)
|
|
@@ -117,9 +180,10 @@ if st.button("Run Agent"):
|
|
| 117 |
else:
|
| 118 |
obs, ok = run_tool(tool, arg)
|
| 119 |
|
| 120 |
-
|
|
|
|
| 121 |
with st.chat_message("user"):
|
| 122 |
-
st.markdown(f"**Step {step} Observation (Tool: {tool})**")
|
| 123 |
if ok:
|
| 124 |
st.info(obs)
|
| 125 |
else:
|
|
@@ -128,5 +192,8 @@ if st.button("Run Agent"):
|
|
| 128 |
messages.append({"role": "assistant", "content": reply})
|
| 129 |
messages.append({"role": "user", "content": f"Observation: {obs}"})
|
| 130 |
|
|
|
|
|
|
|
|
|
|
| 131 |
except Exception as e:
|
| 132 |
-
st.error(f"An error occurred: {e}")
|
|
|
|
| 1 |
+
import os
|
|
|
|
| 2 |
import re
|
| 3 |
+
import streamlit as st
|
| 4 |
from openai import OpenAI
|
| 5 |
|
| 6 |
# Configure page style
|
| 7 |
+
st.set_page_config(page_title="Clinical ReAct Agent — MIS 752", layout="wide")
|
| 8 |
+
st.title("🩺 Clinical ReAct Agent & Supervisor Dashboard")
|
| 9 |
+
st.markdown(
|
| 10 |
+
"**MIS 752 · Lab 9: Build the Paperwork Machine** — Custom ReAct loop (`Thought → Action → Observation`) "
|
| 11 |
+
"with `stop=['Observation:']` hallucination prevention and automated human-in-the-loop escalation policies."
|
| 12 |
+
)
|
| 13 |
|
| 14 |
# Sidebar - Secrets & Settings
|
| 15 |
+
st.sidebar.header("⚙️ Agent & Supervisor Settings")
|
| 16 |
+
|
| 17 |
+
default_key = os.environ.get("OPENROUTER_API_KEY") or os.environ.get("OPENAI_API_KEY") or ""
|
| 18 |
+
api_key = st.sidebar.text_input(
|
| 19 |
+
"API Key (OpenRouter or OpenAI)",
|
| 20 |
+
value=default_key,
|
| 21 |
+
type="password",
|
| 22 |
+
placeholder="sk-or-v1-... or sk-..."
|
| 23 |
+
)
|
| 24 |
+
|
| 25 |
+
is_openrouter = api_key.strip().startswith("sk-or-")
|
| 26 |
+
provider_label = "OpenRouter (https://openrouter.ai/api/v1)" if is_openrouter else "OpenAI Default"
|
| 27 |
+
st.sidebar.caption(f"🔌 Active Endpoint: **{provider_label}**")
|
| 28 |
+
|
| 29 |
+
if is_openrouter:
|
| 30 |
+
model_options = [
|
| 31 |
+
"openai/gpt-4o-mini",
|
| 32 |
+
"qwen/qwen3-8b",
|
| 33 |
+
"meta-llama/llama-3.3-70b-instruct",
|
| 34 |
+
"anthropic/claude-3.5-haiku"
|
| 35 |
+
]
|
| 36 |
+
else:
|
| 37 |
+
model_options = ["gpt-4o-mini", "gpt-4o", "gpt-3.5-turbo"]
|
| 38 |
+
|
| 39 |
+
model_name = st.sidebar.selectbox("LLM Model Name", model_options)
|
| 40 |
max_steps = st.sidebar.slider("Step Budget (Max Steps)", 2, 8, 6)
|
| 41 |
+
supervision_policy = st.sidebar.selectbox(
|
| 42 |
+
"Supervision Policy (Step 9.2)",
|
| 43 |
+
["flagged", "every", "audit"],
|
| 44 |
+
help="Controls when the human supervisor is interrupted."
|
| 45 |
+
)
|
| 46 |
|
| 47 |
# Clinic data structures
|
| 48 |
REFERENCE = {
|
|
|
|
| 59 |
"bianchi": "Bianchi, F, 81. Type 2 diabetes, frail. eGFR 39. A1c 8.4 percent. Current: metformin 500 mg twice daily.",
|
| 60 |
}
|
| 61 |
|
| 62 |
+
with st.sidebar.expander("📚 View Reference & Patient Charts"):
|
| 63 |
+
st.markdown("**Clinic Reference Topics:** `" + "`, `".join(REFERENCE.keys()) + "`")
|
| 64 |
+
st.markdown("**Patient Charts:** `" + "`, `".join(CHARTS.keys()) + "`")
|
| 65 |
+
|
| 66 |
def lookup(topic: str) -> str:
|
| 67 |
hit = REFERENCE.get(topic.strip().lower())
|
| 68 |
return hit if hit else f"'{topic}' is not in the clinic reference."
|
|
|
|
| 89 |
return None, None
|
| 90 |
return tool.group(1).strip(), (arg.group(1).strip() if arg else "")
|
| 91 |
|
| 92 |
+
def escalate_step(stage: dict, policy: str = "flagged") -> str:
|
| 93 |
+
where = f"step {stage.get('step', '?')} ({stage.get('stage', '?')})"
|
| 94 |
+
tool = stage.get("tool") or "no tool"
|
| 95 |
+
ok = bool(stage.get("ok", True))
|
| 96 |
+
|
| 97 |
+
if policy == "every":
|
| 98 |
+
return f"INTERRUPT - {where}, {tool}: policy is review-everything."
|
| 99 |
+
elif policy == "audit":
|
| 100 |
+
return f"LOG - {where}, {tool}: recorded for later review, nobody paged."
|
| 101 |
+
elif policy == "flagged":
|
| 102 |
+
if not ok:
|
| 103 |
+
return f"INTERRUPT - {where}, {tool}: the agent reported a failure here."
|
| 104 |
+
return f"SILENT - {where}, {tool}: nothing flagged."
|
| 105 |
+
else:
|
| 106 |
+
return f"INTERRUPT - unknown policy '{policy}'. Defaulting to interrupt."
|
| 107 |
+
|
| 108 |
SYSTEM_PROMPT = '''You answer clinical questions by looking things up. You have two tools:
|
| 109 |
|
| 110 |
lookup: look up one topic in the clinic reference. Input: a short topic, e.g. "metformin".
|
|
|
|
| 125 |
Final Answer: your answer, in one or two sentences
|
| 126 |
/no_think'''
|
| 127 |
|
| 128 |
+
preset = st.selectbox(
|
| 129 |
+
"Choose a preset lab question or customize below:",
|
| 130 |
+
[
|
| 131 |
+
"Is Alvarez still OK on her metformin? Check her chart first.",
|
| 132 |
+
"What share of drug-interaction alerts get overridden, and why does that matter for a unit that fires 240 alerts a shift?",
|
| 133 |
+
"What does our hospital night shift override policy say about 240 alerts in one night?",
|
| 134 |
+
"Check Bianchi's chart and our a1c target and metformin rules. Does her regimen need adjustment?"
|
| 135 |
+
]
|
| 136 |
+
)
|
| 137 |
+
|
| 138 |
+
query = st.text_input("Ask the clinical agent a question:", value=preset)
|
| 139 |
|
| 140 |
+
if st.button("▶️ Run Agent", type="primary"):
|
| 141 |
if not api_key:
|
| 142 |
+
st.error("Please enter an OpenRouter (`sk-or-v1-...`) or OpenAI (`sk-...`) API Key in the sidebar to run the agent.")
|
| 143 |
else:
|
| 144 |
try:
|
| 145 |
+
base_url = "https://openrouter.ai/api/v1" if is_openrouter else None
|
| 146 |
+
client = OpenAI(api_key=api_key.strip(), base_url=base_url)
|
| 147 |
messages = [{"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": query}]
|
| 148 |
+
converged = False
|
| 149 |
|
| 150 |
st.subheader("🧠 Agent Execution Trace")
|
| 151 |
|
|
|
|
| 155 |
model=model_name,
|
| 156 |
messages=messages,
|
| 157 |
temperature=0,
|
| 158 |
+
max_tokens=1000,
|
| 159 |
stop=["Observation:"]
|
| 160 |
)
|
| 161 |
reply = r.choices[0].message.content or ""
|
| 162 |
+
reply = re.sub(r"<think>.*?</think>", "", reply, flags=re.S)
|
| 163 |
+
reply = re.sub(r"<think>.*$", "", reply, flags=re.S).strip()
|
| 164 |
+
|
| 165 |
+
model_stage = {"step": step, "stage": "model", "text": reply, "tool": None, "ok": bool(reply)}
|
| 166 |
|
|
|
|
| 167 |
with st.chat_message("assistant"):
|
| 168 |
+
st.markdown(f"**Step {step} Reasoning** · `{escalate_step(model_stage, supervision_policy)}`")
|
| 169 |
+
st.text(reply or "[empty reply]")
|
| 170 |
|
| 171 |
if "Final Answer:" in reply:
|
| 172 |
answer = reply.split("Final Answer:", 1)[1].strip()
|
| 173 |
+
st.success(f"✅ **Final Answer (Converged in {step} steps):** {answer}")
|
| 174 |
+
converged = True
|
| 175 |
break
|
| 176 |
|
| 177 |
tool, arg = parse_action(reply)
|
|
|
|
| 180 |
else:
|
| 181 |
obs, ok = run_tool(tool, arg)
|
| 182 |
|
| 183 |
+
obs_stage = {"step": step, "stage": "observation", "text": obs, "tool": tool, "ok": ok}
|
| 184 |
+
|
| 185 |
with st.chat_message("user"):
|
| 186 |
+
st.markdown(f"**Step {step} Observation (Tool: `{tool}`)** · `{escalate_step(obs_stage, supervision_policy)}`")
|
| 187 |
if ok:
|
| 188 |
st.info(obs)
|
| 189 |
else:
|
|
|
|
| 192 |
messages.append({"role": "assistant", "content": reply})
|
| 193 |
messages.append({"role": "user", "content": f"Observation: {obs}"})
|
| 194 |
|
| 195 |
+
if not converged:
|
| 196 |
+
st.warning(f"⚠️ Step budget exhausted (`max_steps={max_steps}`) before reaching a Final Answer.")
|
| 197 |
+
|
| 198 |
except Exception as e:
|
| 199 |
+
st.error(f"An error occurred: {e}")
|