tool using

app.py

@@ -1,4 +1,6 @@
 import os
+import re
+import json
 import gradio as gr
 from openai import OpenAI
 
@@ -48,10 +50,47 @@ Rules:
 - Keep the draft reply professional and appropriate for business use
 """
 
+ADVANCED_TOOL_CALLING_PROMPT = """You are an enterprise email triage agent.
+
+Your job is to analyze incoming emails and decide how they should be handled.
+
+You have access to one tool:
+
+lookup_status(reference_id)
+
+Use this tool only when the email refers to a specific order, shipment, incident, or ticket and the current status is needed to answer well.
+
+If you decide to use the tool, respond with exactly this format and nothing else:
+
+CALL_TOOL: lookup_status <reference_id>
+
+Examples:
+CALL_TOOL: lookup_status SG-18427
+CALL_TOOL: lookup_status INC-2041
+
+If no tool is needed, then provide your final answer directly.
+
+Final output format:
+
+Intent:
+Priority:
+Owner:
+Action:
+Reasoning:
+Draft Reply:
+
+Rules:
+- Be concise and practical
+- If the email is ambiguous, say so clearly
+- If there is urgency, reflect that in both priority and action
+- Keep the draft reply professional and appropriate for business use
+- Do not call the tool unless it is genuinely needed
+"""
+
 SAMPLE_EMAILS = {
-    "Customer Complaint": """Subject: Delayed shipment and no updates
+    "Customer Complaint": """Subject: Delayed shipment SG-18427
 
-We placed our order nearly two weeks ago and still have not received the shipment. We have also not had any meaningful update from your team despite following up twice.
+We placed our order nearly two weeks ago and still have not received shipment SG-18427. We have also not had any meaningful update from your team despite following up twice.
 
 This delay is now affecting our own delivery commitments to customers. Please let us know what is going on and when we can expect resolution.
 
@@ -88,11 +127,11 @@ Please keep this confidential.
 
 Thank you.""",
 
-    "Urgent Escalation": """Subject: Production outage affecting client operations
+    "Urgent Escalation": """Subject: Update on incident INC-2041
 
 Team,
 
-A major system outage started 30 minutes ago and several enterprise clients are unable to access critical workflows. We are already receiving escalation calls.
+A major outage is still affecting several enterprise clients. Can someone update us on incident INC-2041 and confirm the current status?
 
 We need immediate support and clear ownership on incident response.
 
@@ -139,16 +178,91 @@ def format_output(raw_text: str) -> str:
 
     return "\n".join(formatted_lines).strip()
 
+STATUS_DB = {
+    "SG-18427": {
+        "record_type": "shipment",
+        "status": "Delayed in customs",
+        "last_update": "2026-04-03 14:20 SGT",
+        "eta": "2026-04-09",
+        "owner_team": "Operations",
+    },
+    "INC-2041": {
+        "record_type": "incident",
+        "status": "Engineering investigating root cause",
+        "last_update": "2026-04-06 10:05 SGT",
+        "eta": "Next update due in 30 minutes",
+        "owner_team": "Operations",
+    },
+    "TKT-48392": {
+        "record_type": "ticket",
+        "status": "Assigned to support tier 2",
+        "last_update": "2026-04-04 17:40 SGT",
+        "eta": "Expected resolution within 1 business day",
+        "owner_team": "Support",
+    },
+}
+
+
+def lookup_status(reference_id: str) -> dict:
+    ref = reference_id.strip().upper()
+    result = STATUS_DB.get(ref)
+
+    if result:
+        return {
+            "found": True,
+            "reference_id": ref,
+            **result
+        }
+
+    return {
+        "found": False,
+        "reference_id": ref,
+        "message": "No matching record found in the status system."
+    }
+
+
+def detect_tool_call(text: str):
+    pattern = r"^\s*CALL_TOOL:\s*(\w+)\s+([A-Z0-9\-]+)\s*$"
+    match = re.match(pattern, text.strip())
+    if match:
+        return match.group(1), match.group(2)
+    return None, None
+
+
+def render_tool_trace(tool_used, tool_name="", tool_input=None, tool_output=None):
+    if not tool_used:
+        return """## Tool Trace
+
+**Tool used:** No
+"""
+
+    input_block = json.dumps(tool_input, indent=2)
+    output_block = json.dumps(tool_output, indent=2)
+
+    return f"""## Tool Trace
+
+**Tool used:** Yes  
+**Tool name:** `{tool_name}`
+
+### Tool input
+```json
+{input_block}
+```
+### Tool output
+```json
+{output_block}
+```
+"""
 
 def run_agent(system_prompt, email, show_raw_output):
     if not system_prompt or not system_prompt.strip():
-        return "Please enter a system prompt.", ""
+        return "Please enter a system prompt.", "", "## Tool Trace\n\nNo run yet."
 
     if not email or not email.strip():
-        return "Please enter an email to analyze.", ""
+        return "Please enter an email to analyze.", "", "## Tool Trace\n\nNo run yet."
 
     try:
-        response = client.chat.completions.create(
+        first_response = client.chat.completions.create(
             model="gpt-4o-mini",
             messages=[
                 {"role": "system", "content": system_prompt},
@@ -157,16 +271,51 @@ def run_agent(system_prompt, email, show_raw_output):
             temperature=0.4,
         )
 
-        raw_output = response.choices[0].message.content or ""
+        first_output = first_response.choices[0].message.content or ""
+        tool_name, reference_id = detect_tool_call(first_output)
+
+        if tool_name == "lookup_status" and reference_id:
+            tool_result = lookup_status(reference_id)
+
+            tool_trace = render_tool_trace(
+                tool_used=True,
+                tool_name=tool_name,
+                tool_input={"reference_id": reference_id},
+                tool_output=tool_result,
+            )
+
+            second_response = client.chat.completions.create(
+                model="gpt-4o-mini",
+                messages=[
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": email},
+                    {"role": "assistant", "content": first_output},
+                    {
+                        "role": "user",
+                        "content": (
+                            "Tool result:\n"
+                            f"{json.dumps(tool_result)}\n\n"
+                            "Now provide the final answer for the email."
+                        ),
+                    },
+                ],
+                temperature=0.4,
+            )
+
+            raw_output = second_response.choices[0].message.content or ""
+        else:
+            tool_trace = render_tool_trace(tool_used=False)
+            raw_output = first_output
+
         formatted_output = format_output(raw_output)
 
         if show_raw_output:
-            return formatted_output, raw_output
-        return formatted_output, ""
+            return formatted_output, raw_output, tool_trace
+        return formatted_output, "", tool_trace
 
     except Exception as e:
         error_message = f"Error: {str(e)}"
-        return error_message, error_message
+        return error_message, error_message, "## Tool Trace\n\nAn error occurred."
 
 
 with gr.Blocks(title="Executive Agent Lab") as demo:
@@ -199,6 +348,7 @@ with gr.Blocks(title="Executive Agent Lab") as demo:
                 load_basic_btn = gr.Button("Load Basic")
                 load_structured_btn = gr.Button("Load Structured")
                 load_advanced_btn = gr.Button("Load Advanced")
+                load_advanced_tool_btn = gr.Button("Load Advanced (Tool)")
 
             with gr.Accordion("What makes a good agent prompt?", open=False):
                 gr.Markdown(
@@ -208,7 +358,8 @@ with gr.Blocks(title="Executive Agent Lab") as demo:
 - routing rules  
 - output format  
 - tone  
-- handling ambiguity"""
+- handling ambiguity
+- when to use the tool and when not to use it"""
                 )
 
         with gr.Column(scale=1):
@@ -243,6 +394,10 @@ with gr.Blocks(title="Executive Agent Lab") as demo:
             formatted_output = gr.Markdown(
                 value="Run the agent to see results here."
             )
+            tool_trace = gr.Markdown(
+                value="## Tool Trace\n\nNo run yet."
+            )
+
 
             show_raw_output = gr.Checkbox(
                 label="Show raw output",
@@ -264,12 +419,16 @@ with gr.Blocks(title="Executive Agent Lab") as demo:
     def load_advanced():
         return ADVANCED_PROMPT
 
+    def load_advanced_tool():
+        return ADVANCED_TOOL_CALLING_PROMPT
+    
     def load_email(sample_name):
         return SAMPLE_EMAILS[sample_name]
 
     load_basic_btn.click(fn=load_basic, outputs=system_prompt)
     load_structured_btn.click(fn=load_structured, outputs=system_prompt)
     load_advanced_btn.click(fn=load_advanced, outputs=system_prompt)
+    load_advanced_tool_btn.click(fn=load_advanced_tool, outputs=system_prompt)
 
     btn_customer.click(fn=lambda: load_email("Customer Complaint"), outputs=email)
     btn_sales.click(fn=lambda: load_email("Sales Inquiry"), outputs=email)
@@ -280,7 +439,7 @@ with gr.Blocks(title="Executive Agent Lab") as demo:
     run_btn.click(
         fn=run_agent,
         inputs=[system_prompt, email, show_raw_output],
-        outputs=[formatted_output, raw_output],
+        outputs=[formatted_output, raw_output, tool_trace],
     )
 
 demo.launch(server_name="0.0.0.0", server_port=7860)