app.py

import os
import csv
import pandas as pd
import gradio as gr
import plotly.express as px

DEFAULT_CSV = os.environ.get("RESULTS_CSV_PATH", "results.csv")

EXPECTED_COLS = [
    "timestamp_iso","run_id","model","prompt_id","category",
    "quality_score","latency_s","energy_wh","tokens","notes"
]

def _load_df(file: gr.File | None):
    path = DEFAULT_CSV
    if file is not None:
        path = file.name
    if not os.path.exists(path):
        return pd.DataFrame(columns=EXPECTED_COLS)
    df = pd.read_csv(path)
    # ensure expected cols exist
    for c in EXPECTED_COLS:
        if c not in df.columns:
            df[c] = None
    # numeric coercion
    for c in ["quality_score","latency_s","energy_wh","tokens"]:
        df[c] = pd.to_numeric(df[c], errors="coerce")
    return df

def _summaries(df: pd.DataFrame):
    if df.empty:
        return df, pd.DataFrame(), pd.DataFrame(), None, None, None, None

    def q_per_wh(row):
        if pd.notna(row["mean_energy"]) and row["mean_energy"] > 0 and pd.notna(row["mean_quality"]):
            return row["mean_quality"] / row["mean_energy"]
        return None

    per_model = df.groupby("model", dropna=False).agg(
        n_runs=("run_id","count"),
        mean_quality=("quality_score","mean"),
        median_latency=("latency_s","median"),
        p95_latency=("latency_s", lambda x: x.dropna().quantile(0.95) if len(x.dropna()) else None),
        mean_latency=("latency_s","mean"),
        mean_energy=("energy_wh","mean"),
        mean_tokens=("tokens","mean")
    ).reset_index()
    per_model["quality_per_wh"] = per_model.apply(q_per_wh, axis=1)

    per_model_cat = df.groupby(["model","category"], dropna=False).agg(
        n_runs=("run_id","count"),
        mean_quality=("quality_score","mean"),
        mean_latency=("latency_s","mean"),
        p95_latency=("latency_s", lambda x: x.dropna().quantile(0.95) if len(x.dropna()) else None),
        mean_energy=("energy_wh","mean")
    ).reset_index()

    c1 = px.bar(per_model.sort_values("mean_quality", ascending=False),
                x="model", y="mean_quality", title="Mean Quality by Model")
    c2 = px.bar(per_model.sort_values("mean_latency"),
                x="model", y="mean_latency", title="Mean Latency (s) by Model")
    c3 = px.bar(per_model.sort_values("p95_latency"),
                x="model", y="p95_latency", title="P95 Latency (s) by Model")
    c4 = px.bar(per_model.sort_values("quality_per_wh", ascending=False),
                x="model", y="quality_per_wh", title="Quality per Watt-hour (↑ better)")

    return df, per_model, per_model_cat, c1, c2, c3, c4

def _filter(df, model_sel, cat_sel, prompt_sel):
    if df.empty:
        return pd.DataFrame()
    out = df.copy()
    if model_sel and model_sel != "ALL":
        out = out[out["model"] == model_sel]
    if cat_sel and cat_sel != "ALL":
        out = out[out["category"] == cat_sel]
    if prompt_sel and prompt_sel != "ALL":
        out = out[out["prompt_id"] == prompt_sel]
    return out

def _choices(df):
    models = ["ALL"] + sorted([m for m in df["model"].dropna().unique().tolist()])
    cats = ["ALL"] + sorted([c for c in df["category"].dropna().unique().tolist()])
    prompts = ["ALL"] + sorted([p for p in df["prompt_id"].dropna().unique().tolist()])
    return models, cats, prompts

with gr.Blocks(title="Compare’IA — Benchmark Dashboard") as demo:
    gr.Markdown("## Compare’IA — Benchmark Dashboard\nUpload your CSV or use the default `results.csv` in the Space repo.")

    with gr.Row():
        csv_file = gr.File(label="Upload results CSV", file_types=[".csv"])
        refresh_btn = gr.Button("Refresh data")

    raw_df = gr.Dataframe(label="Raw data", interactive=False, wrap=True, height=300)

    with gr.Row():
        model_dd = gr.Dropdown(choices=["ALL"], value="ALL", label="Model")
        cat_dd = gr.Dropdown(choices=["ALL"], value="ALL", label="Category")
        prompt_dd = gr.Dropdown(choices=["ALL"], value="ALL", label="Prompt ID")
        apply_filter = gr.Button("Apply filter")

    filtered_df = gr.Dataframe(label="Filtered rows", interactive=False, height=250)

    with gr.Accordion("Aggregates & Charts", open=True):
        per_model_df = gr.Dataframe(label="Per-model summary", interactive=False)
        per_model_cat_df = gr.Dataframe(label="Per-model-per-category", interactive=False)
        chart_quality = gr.Plot(label="Mean Quality by Model")
        chart_mean_lat = gr.Plot(label="Mean Latency by Model")
        chart_p95_lat = gr.Plot(label="P95 Latency by Model")
        chart_q_per_wh = gr.Plot(label="Quality per Wh")

    def _refresh(file):
        df = _load_df(file)
        models, cats, prompts = _choices(df)
        full_df, pm, pmc, c1, c2, c3, c4 = _summaries(df)
        return (full_df, gr.update(choices=models, value="ALL"),
                gr.update(choices=cats, value="ALL"),
                gr.update(choices=prompts, value="ALL"),
                pm, pmc, c1, c2, c3, c4)

    refresh_btn.click(_refresh, inputs=csv_file,
                      outputs=[raw_df, model_dd, cat_dd, prompt_dd, per_model_df, per_model_cat_df,
                               chart_quality, chart_mean_lat, chart_p95_lat, chart_q_per_wh])

    def _apply(file, model_sel, cat_sel, prompt_sel):
        df = _load_df(file)
        out = _filter(df, model_sel, cat_sel, prompt_sel)
        return out

    apply_filter.click(_apply, inputs=[csv_file, model_dd, cat_dd, prompt_dd], outputs=[filtered_df])

demo.launch()