Update app.py

app.py

@@ -1,70 +1,250 @@
 import gradio as gr
+import pandas as pd
+import os
+import json
 from huggingface_hub import InferenceClient
 
+# ===============================
+# LLM CLIENT
+# ===============================
+HF_TOKEN = os.getenv("HF")
 
-def respond(
-    message,
-    history: list[dict[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-    hf_token: gr.OAuthToken,
-):
-    """
-    For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-    """
-    client = InferenceClient(token=hf_token.token, model="openai/gpt-oss-20b")
+# client = InferenceClient(
+#     model="Qwen/Qwen2.5-7B-Instruct",
+#     token=HF_TOKEN
+# )
+client = InferenceClient(model="meta-llama/Meta-Llama-3-8B-Instruct", token=HF_TOKEN)
 
-    messages = [{"role": "system", "content": system_message}]
+# ===============================
+# SAFETY
+# ===============================
+BLOCKED_KEYWORDS = [
+    "import os", "import sys", "subprocess",
+    "open(", "eval(", "exec(", "__",
+    "socket", "requests"
+]
 
-    messages.extend(history)
+# ===============================
+# AGENT: DECIDE MODE
+# ===============================
+def decide_mode(user_question, df):
+    prompt = f"""
+You are an expert data analysis agent.
 
-    messages.append({"role": "user", "content": message})
+Dataset columns:
+{list(df.columns)}
 
-    response = ""
+Decide how to answer the user's question.
+
+Choose ONLY ONE mode:
+- code_and_insight → requires exact computation
+- insight_only → qualitative reasoning only
+
+Respond ONLY in valid JSON:
+{{
+  "mode": "code_and_insight | insight_only",
+  "needs_code_visible": true | false
+}}
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
+Set needs_code_visible = true ONLY if the user explicitly asks for code.
+
+User question:
+{user_question}
+"""
+    response = ""
+    for chunk in client.chat_completion(
+        messages=[{"role": "user", "content": prompt}],
+        max_tokens=150,
+        temperature=0,
         stream=True,
-        temperature=temperature,
-        top_p=top_p,
     ):
-        choices = message.choices
-        token = ""
-        if len(choices) and choices[0].delta.content:
-            token = choices[0].delta.content
+        if chunk.choices and chunk.choices[0].delta.content:
+            response += chunk.choices[0].delta.content
+    return response
+
+
+# ===============================
+# AGENT: CODE GEN + EXEC
+# ===============================
+def generate_and_run_code(user_question, df, retries=1):
+    column_info = {col: str(dtype) for col, dtype in df.dtypes.items()}
+    last_error = None
+
+    for attempt in range(retries + 1):
+        planner_prompt = f"""
+You are a Python data analyst.
+
+Dataset columns and types:
+{column_info}
+
+Rules:
+- Use pandas only
+- Dataframe name: df
+- Store final output in variable named: result
+- No explanations
+- No markdown
+- No imports
+
+User question:
+{user_question}
+"""
+
+        if attempt > 0:
+            planner_prompt += f"\nPrevious error:\n{last_error}\nFix the code."
+
+        code = ""
+        for chunk in client.chat_completion(
+            messages=[{"role": "user", "content": planner_prompt}],
+            max_tokens=400,
+            temperature=0.2,
+            stream=True,
+        ):
+            if chunk.choices and chunk.choices[0].delta.content:
+                code += chunk.choices[0].delta.content
+
+        if any(bad in code for bad in BLOCKED_KEYWORDS):
+            return None, None, "Unsafe code detected"
+
+        local_env = {"df": df, "result": None}
+        try:
+            exec(code, {}, local_env)
+            return code, local_env["result"], None
+        except Exception as e:
+            last_error = str(e)
+
+    return None, None, last_error
+
+
+# ===============================
+# CORE CHATBOT
+# ===============================
+def analyze_excel(message, history, file):
+
+    if file is None:
+        yield "⚠️ Please upload an Excel file first."
+        return
 
-        response += token
-        yield response
+    user_question = message["content"] if isinstance(message, dict) else message
 
+    try:
+        df = pd.read_excel(file.name, engine="openpyxl")
 
+        # 🧠 Decide mode
+        decision_raw = decide_mode(user_question, df)
+
+        try:
+            decision = json.loads(decision_raw)
+        except Exception:
+            yield "❌ Unable to interpret the request. Please rephrase."
+            return
+
+        mode = decision["mode"]
+        show_code = decision["needs_code_visible"]
+
+        # ===============================
+        # CODE + INSIGHT MODE
+        # ===============================
+        if mode == "code_and_insight":
+            yield "🧠 Running analysis…"
+
+            code, result, error = generate_and_run_code(
+                user_question=user_question,
+                df=df,
+                retries=1
+            )
+
+            if error:
+                yield f"❌ Computation failed: {error}"
+                return
+
+            # Build insight prompt
+            insight_prompt = f"""
+You are a senior data analyst.
+
+User question:
+{user_question}
+
+Computed result:
+{result}
+
+Explain the insight clearly in natural language.
+Focus on meaning and implications.
 """
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
+
+            response = ""
+            if show_code:
+                response += f"🧾 Generated Python code:\n\n```python\n{code}\n```\n\n"
+
+            for chunk in client.chat_completion(
+                messages=[{"role": "user", "content": insight_prompt}],
+                max_tokens=350,
+                temperature=0.4,
+                stream=True,
+            ):
+                if chunk.choices and chunk.choices[0].delta.content:
+                    response += chunk.choices[0].delta.content
+                    yield response
+            return
+
+        # ===============================
+        # INSIGHT ONLY MODE
+        # ===============================
+        summary = f"""
+Rows: {len(df)}
+Columns: {list(df.columns)}
+Missing values:
+{df.isnull().sum().to_string()}
+"""
+
+        insight_prompt = f"""
+Dataset summary:
+{summary}
+
+User question:
+{user_question}
+
+Provide high-level analytical insights.
+Do not compute exact numbers.
+Do not generate code.
 """
-chatbot = gr.ChatInterface(
-    respond,
-    type="messages",
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
-
-with gr.Blocks() as demo:
-    with gr.Sidebar():
-        gr.LoginButton()
-    chatbot.render()
 
+        response = ""
+        for chunk in client.chat_completion(
+            messages=[{"role": "user", "content": insight_prompt}],
+            max_tokens=400,
+            temperature=0.4,
+            stream=True,
+        ):
+            if chunk.choices and chunk.choices[0].delta.content:
+                response += chunk.choices[0].delta.content
+                yield response
+
+    except Exception as e:
+        yield f"❌ Error: {str(e)}"
+
+
+# ===============================
+# UI
+# ===============================
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 📊 Excel Analysis Agent (Code + Insight)")
+
+    excel_file = gr.File(
+        label="Upload Excel File",
+        file_types=[".xlsx"]
+    )
+
+    gr.ChatInterface(
+        fn=analyze_excel,
+        additional_inputs=[excel_file],
+        type="messages",
+        examples=[
+            ["Which provider has the highest average claim amount?", None],
+            ["How many unique members are there?", None],
+            ["Explain trends in processing time across regions.", None],
+            ["Show python code to compute correlation between ClaimAmount and ProcessingCost.", None],
+        ],
+    )
 
 if __name__ == "__main__":
     demo.launch()