Update app.py

app.py

@@ -1,15 +1,26 @@
 """
-Governance-GPT Quiz – AI-Maturity Self-Assessment (bucket version)
+Governance-GPT Quiz – bucket version (LLM-powered summary)
 Author: Rishabh Sharma – 2025-09-14
-Single-file Gradio Space: 15-question Likert survey ➜ bucket scores ➜ PDF+summary
+Single-file Gradio Space: 15-question Likert survey → bucket scores → LLM summary + PDF
 """
 
-import os, tempfile, datetime
+import os, tempfile, datetime, warnings
 import gradio as gr
 import pandas as pd
-from fpdf import FPDF  # pure-python PDF – no external binaries
-
-# ------------------- Question bank ------------------- #
+from fpdf import FPDF               # pure-python PDF
+from transformers import pipeline    # 🔹 NEW: LLM summariser
+
+# ───────────── Initialize LLM once ───────────── #
+# Small, instruction-tuned model that runs comfortably on free CPU Spaces
+warnings.filterwarnings("ignore", category=UserWarning)  # tokenizer warnings
+summariser = pipeline(
+    task="text2text-generation",
+    model="google/flan-t5-base",
+    tokenizer="google/flan-t5-base",
+    max_new_tokens=220,
+)
+
+# ───────────── Question bank ───────────── #
 QUESTIONS = [
     "Governance framework is documented and communicated across the organisation.",
     "Roles & responsibilities for AI oversight are clearly assigned.",
@@ -32,7 +43,6 @@ QUESTIONS = [
     "Model cards or equivalent documentation exist for each deployed model.",
 ]
 
-# Five buckets, 3 questions each (index positions are 0-based)
 BUCKETS = {
     "Governance & Strategy":      [0, 1, 2],
     "Data & Privacy":             [3, 4, 5],
@@ -49,25 +59,7 @@ TIERS = {
     "Optimized":  (4.51, 5.00),
 }
 
-# Generic tier actions (overall)
-OVERALL_ACTION = {
-    "Initial":    "kick-off a cross-functional task-force and establish a basic AI policy.",
-    "Repeatable": "formalise core processes and create mandatory model documentation.",
-    "Defined":    "scale governance with automated monitoring and periodic audits.",
-    "Managed":    "integrate governance KPIs into OKRs and adopt continuous compliance tooling.",
-    "Optimized":  "benchmark externally, publish transparency reports, and champion open governance.",
-}
-
-# Bucket-specific remediation snippets
-BUCKET_ACTION = {
-    "Governance & Strategy":     "formalise decision rights and publish an organisation-wide AI charter.",
-    "Data & Privacy":            "implement end-to-end data lineage and privacy impact assessments.",
-    "Risk & Compliance":         "create incident playbooks and third-party risk registers.",
-    "Security & Infrastructure": "harden model artefact storage and secure inference endpoints.",
-    "Lifecycle & Oversight":     "define human-override mechanisms and retirement criteria for models.",
-}
-
-# ------------------- Helpers ------------------- #
+# ───────────── Helpers ───────────── #
 def score_to_tier(avg: float) -> str:
     for tier, (low, high) in TIERS.items():
         if low <= avg <= high:
@@ -75,120 +67,111 @@ def score_to_tier(avg: float) -> str:
     return "Unclassified"
 
 def latin1(txt: str) -> str:
-    """Convert to Latin-1 safe string for FPDF."""
     return txt.replace("…", "...").encode("latin-1", "replace").decode("latin-1")
 
+def llm_remediation(product: str, bucket_avgs: dict, overall_tier: str) -> str:
+    """Call Flan-T5 to get a markdown summary with recommendations."""
+    # Craft a concise prompt to stay within max_new_tokens
+    bucket_lines = "\n".join(
+        f"- {b}: {v:.2f}" for b, v in bucket_avgs.items()
+    )
+    prompt = (
+        "You are an AI governance consultant.\n"
+        f"Product: {product}\n"
+        f"Overall tier: {overall_tier}\n"
+        "Bucket scores (1-5):\n"
+        f"{bucket_lines}\n\n"
+        "Write a short markdown summary (≤120 words) with:\n"
+        "• A one-sentence overall assessment\n"
+        "• 3-5 bullet remediation actions, prioritised, referencing bucket names.\n"
+        "Do not add any other sections."
+    )
+    try:
+        gen = summariser(prompt)[0]["generated_text"]
+        return gen.strip()
+    except Exception as e:
+        # Fallback in rare failure cases
+        return f"LLM summary unavailable – {e}"
+
 def build_pdf(product: str, bucket_df: pd.DataFrame, overall_avg: float,
               overall_tier: str, file_path: str, summary_text: str):
     pdf = FPDF()
     pdf.set_auto_page_break(auto=True, margin=15)
     pdf.add_page()
 
-    # Title
     pdf.set_font("Helvetica", "B", 16)
     pdf.cell(0, 10, latin1(f"AI Governance Maturity Report – {product}"), ln=1, align="C")
     pdf.set_font("Helvetica", "", 12)
     pdf.cell(0, 8, f"Generated on {datetime.date.today().isoformat()}", ln=1, align="C")
     pdf.ln(4)
 
-    # Summary
     pdf.set_font("Helvetica", "B", 12)
     pdf.cell(0, 8, latin1(f"Overall Score: {overall_avg:.2f}  |  Tier: {overall_tier}"), ln=1)
     pdf.set_font("Helvetica", "", 11)
     pdf.multi_cell(0, 6, latin1(summary_text))
     pdf.ln(4)
 
-    # Bucket table header
     pdf.set_font("Helvetica", "B", 11)
     pdf.cell(80, 8, "Bucket", 1)
     pdf.cell(35, 8, "Avg Score", 1)
     pdf.cell(35, 8, "Tier", 1, ln=1)
     pdf.set_font("Helvetica", "", 10)
 
-    # Rows
     for _, row in bucket_df.iterrows():
         pdf.cell(80, 8, latin1(row["Bucket"][:40]), 1)
         pdf.cell(35, 8, f'{row["Avg Score"]:.2f}', 1)
         pdf.cell(35, 8, row["Tier"], 1, ln=1)
 
-    # Overall row
     pdf.cell(80, 8, "Overall", 1)
     pdf.cell(35, 8, f"{overall_avg:.2f}", 1)
     pdf.cell(35, 8, overall_tier, 1, ln=1)
 
     pdf.output(file_path)
 
-def generate_summary(product: str, bucket_avgs: dict, overall_tier: str) -> str:
-    """Return markdown remediation summary."""
-    lines = [f"### 🏷️ AI Governance Summary for **{product}**",
-             f"- **Overall tier:** {overall_tier}"]
-
-    # Bucket-specific feedback
-    for bucket, avg in bucket_avgs.items():
-        tier = score_to_tier(avg)
-        # Only add remediation if below Managed
-        if tier in ("Initial", "Repeatable", "Defined"):
-            lines.append(f"- **{bucket}** score is {avg:.2f} → *{tier}*. "
-                         f"{product} should {BUCKET_ACTION[bucket]}")
-    # If all buckets high, generic kudos
-    if len(lines) == 2:
-        lines.append("Great job! All buckets are Managed or Optimized.")
-
-    # Overall action line
-    lines.append(f"- Next step: {product} should {OVERALL_ACTION[overall_tier]}")
-    return "  \n".join(lines)
-
-# ------------------- Gradio callback ------------------- #
+# ───────────── Gradio callback ───────────── #
 def generate_report(product_name, *scores):
     product = product_name.strip() or "your product"
     scores = list(scores)
 
-    # Compute bucket averages
-    bucket_avgs = {}
-    for bucket, idxs in BUCKETS.items():
-        bucket_avgs[bucket] = sum(scores[i] for i in idxs) / len(idxs)
-
+    bucket_avgs = {
+        bucket: sum(scores[i] for i in idxs) / len(idxs)
+        for bucket, idxs in BUCKETS.items()
+    }
     overall_avg = sum(scores) / len(scores)
     overall_tier = score_to_tier(overall_avg)
 
-    # Prepare DF for PDF table
     bucket_df = pd.DataFrame({
         "Bucket": list(bucket_avgs.keys()),
         "Avg Score": list(bucket_avgs.values()),
         "Tier": [score_to_tier(v) for v in bucket_avgs.values()],
     })
 
-    # Build remediation summary
-    summary_md = generate_summary(product, bucket_avgs, overall_tier)
+    summary_md = llm_remediation(product, bucket_avgs, overall_tier)
 
-    # Generate PDF
     tmp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=".pdf")
     build_pdf(product, bucket_df, overall_avg, overall_tier, tmp_pdf.name, summary_md)
 
     return summary_md, tmp_pdf.name
 
-# ------------------- Gradio UI ------------------- #
+# ───────────── UI ───────────── #
 with gr.Blocks(title="Governance-GPT Quiz") as demo:
     gr.Markdown(
         """
 # Governance-GPT Quiz  
 Enter your **product / system name**, rate each statement from **1 (Strongly Disagree)** to **5 (Strongly Agree)**,  
-and get a bucket-wise governance score with a personalised remediation plan.
+and receive an LLM-generated remediation plan plus PDF bucket report.
         """
     )
 
     product_inp = gr.Textbox(label="Product / System Name", placeholder="e.g. AcmeAI Recommendation Engine")
-
     sliders = [gr.Slider(1, 5, value=3, step=1, label=q) for q in QUESTIONS]
 
     btn = gr.Button("Generate PDF Report")
     summary_out = gr.Markdown()
     pdf_out = gr.File(label="⬇️ Download your PDF")
 
-    btn.click(
-        fn=generate_report,
-        inputs=[product_inp] + sliders,
-        outputs=[summary_out, pdf_out],
-    )
+    btn.click(fn=generate_report,
+              inputs=[product_inp] + sliders,
+              outputs=[summary_out, pdf_out])
 
 demo.launch()