app.py

import gradio as gr
import difflib
import pandas as pd

from samples import BASE_INSTRUCTION, DRIFT_SCENARIOS
from drift_env import evaluate_instruction_drift
from drift_agent import generate_response as flan_generate

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# =====================================================
# MODELS
# =====================================================

ALT_MODEL = "google/flan-t5-small"
alt_tokenizer = AutoTokenizer.from_pretrained(ALT_MODEL)
alt_model = AutoModelForSeq2SeqLM.from_pretrained(ALT_MODEL)
alt_model.eval()

def alt_generate(prompt: str) -> str:
    inputs = alt_tokenizer(prompt, return_tensors="pt", truncation=True)
    with torch.no_grad():
        outputs = alt_model.generate(**inputs, max_new_tokens=160)
    return alt_tokenizer.decode(outputs[0], skip_special_tokens=True)

MODELS = {
    "FLAN-T5-Base": flan_generate,
    "FLAN-T5-Small": alt_generate
}

SAMPLE_QUESTION = "Explain how the system should respond when API latency increases."

# =====================================================
# HELPERS
# =====================================================

def build_prompt(extra):
    return BASE_INSTRUCTION + "\n\n" + extra + "\n\nQuestion: " + SAMPLE_QUESTION

def diff_highlight(a, b):
    diff = difflib.ndiff(a.split(), b.split())
    html = ""
    for t in diff:
        if t.startswith("+"):
            html += f"<span style='color:#22c55e'> {t}</span>"
        elif t.startswith("-"):
            html += f"<span style='color:#ef4444'> {t}</span>"
        else:
            html += f"{t} "
    return html

# =====================================================
# CORE
# =====================================================

def run_eval(model_name):
    gen = MODELS[model_name]

    steps = []
    scores = []

    baseline = gen(build_prompt(""))

    for i, s in enumerate(DRIFT_SCENARIOS):
        resp = gen(build_prompt(s["extra_context"]))
        _, score = evaluate_instruction_drift(resp)

        steps.append(i)
        scores.append(score)

    df = pd.DataFrame({"Step": steps, "Drift": scores})

    final_resp = gen(build_prompt(DRIFT_SCENARIOS[-1]["extra_context"]))

    verdict = "🟢 Stable" if scores[-1] == 0 else "🔴 Drift Detected"

    return df, baseline, final_resp, diff_highlight(baseline, final_resp), verdict

# =====================================================
# COMPARISON MODE
# =====================================================

def compare_models(m1, m2):
    df1, base1, drift1, diff1, v1 = run_eval(m1)
    df2, base2, drift2, diff2, v2 = run_eval(m2)

    return df1, df2, diff1, diff2, v1, v2

# =====================================================
# UI
# =====================================================

with gr.Blocks(css="""
body {background:#020617;color:#e5e7eb;font-family:Inter}
button {background:linear-gradient(135deg,#0ea5e9,#22c55e);color:white}
a {color:#38bdf8}
""") as app:

    gr.Markdown("# 🚀 Instruction Drift Simulator")
    gr.Markdown("### Compare how models lose instruction fidelity under pressure")

    with gr.Tabs():

        # ---------------- SINGLE MODEL ----------------
        with gr.Tab("Single Model"):
            model = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Base")
            run = gr.Button("Run Evaluation")

            chart = gr.LinePlot(label="Drift Over Time")
            baseline = gr.Textbox(label="Baseline", lines=4)
            drifted = gr.Textbox(label="Drifted", lines=4)
            diff = gr.HTML(label="Diff")
            verdict = gr.Textbox(label="Verdict")

            run.click(
                run_eval,
                model,
                [chart, baseline, drifted, diff, verdict]
            )

        # ---------------- COMPARISON ----------------
        with gr.Tab("Model Comparison"):
            m1 = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Base", label="Model A")
            m2 = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Small", label="Model B")
            compare = gr.Button("Compare")

            chart1 = gr.LinePlot(label="Model A Drift")
            chart2 = gr.LinePlot(label="Model B Drift")

            diff1 = gr.HTML(label="Diff A")
            diff2 = gr.HTML(label="Diff B")

            v1 = gr.Textbox(label="Verdict A")
            v2 = gr.Textbox(label="Verdict B")

            compare.click(
                compare_models,
                [m1, m2],
                [chart1, chart2, diff1, diff2, v1, v2]
            )

    # FOOTER
    gr.Markdown("""
---
### Built by <b>Aditi Khare</b>  
🌐 <a href="https://aditikhare.com" target="_blank">
AditiKhare.com — Enterprise AI Product Ecosystem | Decision Intelligence
</a>
""")

app.launch()