app.py

#!/usr/bin/env python3

from __future__ import annotations

# Copyright 2023 Dmitry Ustalov
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

__author__ = "Dmitry Ustalov"
__license__ = "Apache 2.0"

from typing import Protocol, cast

import evalica
import gradio as gr
import networkx as nx
import numpy as np
import pandas as pd
import plotly.express as px
import scipy.stats
from evalica import AlphaBootstrapResult, Winner, alpha_bootstrap
from plotly.graph_objects import Figure

TOLERANCE, LIMIT = 1e-6, 100
MORE_EVALICA = """
**More Evalica:**

- Paper: [2025.coling-demos.6](https://aclanthology.org/2025.coling-demos.6/) ([arXiv](https://arxiv.org/abs/2412.11314))
- GitHub: <https://github.com/dustalov/evalica>
- PyPI: <https://pypi.org/project/evalica/>
- conda-forge: <https://anaconda.org/conda-forge/evalica>
- crates.io: <https://crates.io/crates/evalica>
- LLMFAO: <https://evalovernite.substack.com/p/llmfao-human-ranking>
""".strip()


def visualize(df_pairwise: pd.DataFrame) -> Figure:
    fig = px.imshow(df_pairwise, color_continuous_scale="RdBu", text_auto=".2f")

    fig.update_layout(xaxis_title="Loser", yaxis_title="Winner", xaxis_side="top")

    fig.update_traces(hovertemplate="Winner: %{y}<br>Loser: %{x}<br>Fraction of Wins: %{z}<extra></extra>")

    return fig


def counting(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.counting(xs, ys, ws, index=index)
    return result.scores


def average_win_rate(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.average_win_rate(xs, ys, ws, index=index)
    return result.scores


def bradley_terry(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.bradley_terry(xs, ys, ws, index=index, tolerance=TOLERANCE, limit=LIMIT)
    return result.scores


def elo(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.elo(xs, ys, ws, index=index)
    return result.scores


def eigen(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.eigen(xs, ys, ws, index=index, tolerance=TOLERANCE, limit=LIMIT)
    return result.scores


def pagerank(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.pagerank(xs, ys, ws, index=index, tolerance=TOLERANCE, limit=LIMIT)
    return result.scores


def newman(
    xs: pd.Series[str],
    ys: pd.Series[str],
    ws: pd.Series[Winner],
    index: pd.Index,
) -> pd.Series[float]:
    result = evalica.newman(xs, ys, ws, index=index, tolerance=TOLERANCE, limit=LIMIT)
    return result.scores


class CallableAlgorithm(Protocol):
    def __call__(
        self,
        xs: pd.Series[str],
        ys: pd.Series[str],
        ws: pd.Series[Winner],
        index: pd.Index,
    ) -> pd.Series[float]: ...


ALGORITHMS: dict[str, CallableAlgorithm] = {
    "Counting": counting,
    "Average Win Rate": average_win_rate,
    "Bradley-Terry (1952)": bradley_terry,
    "Elo (1960)": elo,
    "Eigenvector (1987)": eigen,
    "PageRank (1998)": pagerank,
    "Newman (2023)": newman,
}


def largest_strongly_connected_component(df_pairs: pd.DataFrame) -> set[str]:
    G = nx.from_pandas_edgelist(df_pairs, source="left", target="right", create_using=nx.DiGraph)
    H = nx.from_pandas_edgelist(
        df_pairs[df_pairs["winner"] == Winner.Draw],
        source="right",
        target="left",
        create_using=nx.DiGraph,
    )
    F = nx.compose(G, H)
    largest = max(nx.strongly_connected_components(F), key=len)
    return cast("set[str]", largest)


def estimate(
    df_pairs: pd.DataFrame,
    algorithm: CallableAlgorithm,
    index: pd.Index,
) -> pd.DataFrame:
    scores = algorithm(df_pairs["left"], df_pairs["right"], df_pairs["winner"], index)

    df_result = pd.DataFrame(data={"score": scores}, index=index)
    df_result.index.name = "item"

    return df_result


def pairwise_bootstrap(
    df_pairs: pd.DataFrame,
    algorithm: CallableAlgorithm,
    index: pd.Index,
    rounds: int,
) -> pd.DataFrame:
    left = df_pairs["left"].to_numpy()
    right = df_pairs["right"].to_numpy()
    winner = df_pairs["winner"].to_numpy()

    def statistic(xs: np.ndarray, ys: np.ndarray, ws: np.ndarray) -> np.ndarray:
        scores = algorithm(pd.Series(xs), pd.Series(ys), pd.Series(ws), index)
        return scores.to_numpy()

    result = scipy.stats.bootstrap(
        (left, right, winner),
        statistic,
        n_resamples=rounds,
        paired=True,
        method="percentile",
        random_state=0,
    )

    ratings = pd.Series(
        np.median(result.bootstrap_distribution, axis=1),
        index=index,
    )

    ci = pd.Series(
        list(zip(result.confidence_interval.low, result.confidence_interval.high, strict=True)),
        index=index,
    )

    df_result = pd.DataFrame({"score": ratings, "ci": ci})
    df_result.index.name = "item"

    return df_result


def handler(
    file: str | None,
    algorithm: str,
    filtered: bool,
    truncated: bool,
    rounds: int,
) -> tuple[pd.DataFrame, Figure]:
    if file is None:
        raise gr.Error("File must be uploaded")

    if algorithm not in ALGORITHMS:
        raise gr.Error(f"Unknown algorithm: {algorithm}")

    try:
        df_pairs = pd.read_csv(file, dtype=str)
    except ValueError as e:
        raise gr.Error(f"Parsing error: {e}") from e

    if not pd.Series(["left", "right", "winner"]).isin(df_pairs.columns).all():
        raise gr.Error("Columns must exist: left, right, winner")

    if not df_pairs["winner"].str.lower().isin(pd.Series(["left", "right", "tie"])).all():
        raise gr.Error("Allowed winner values: left, right, tie")

    df_pairs = df_pairs[["left", "right", "winner"]]
    df_pairs["winner"] = (
        df_pairs["winner"]
        .str.lower()
        .map(
            {"left": Winner.X, "right": Winner.Y, "tie": Winner.Draw},
        )
    )

    df_pairs = df_pairs.loc[df_pairs.notna().all(axis=1)]

    if filtered:
        largest = largest_strongly_connected_component(df_pairs)
        mask = df_pairs["left"].isin(largest) & df_pairs["right"].isin(largest)
        df_pairs = df_pairs.loc[mask]

    *_, index = evalica.indexing(xs=df_pairs["left"], ys=df_pairs["right"])

    if rounds:
        df_result = pairwise_bootstrap(df_pairs, ALGORITHMS[algorithm], index, rounds)
    else:
        df_result = estimate(df_pairs, ALGORITHMS[algorithm], index)

    df_result["pairs"] = (
        pd.Series(0, dtype=int, index=index)
        .add(
            df_pairs.groupby("left")["left"].count(),
            fill_value=0,
        )
        .add(
            df_pairs.groupby("right")["right"].count(),
            fill_value=0,
        )
        .astype(int)
    )

    df_result["rank"] = df_result["score"].rank(na_option="bottom", ascending=False).astype(int)

    df_result = df_result.fillna(-np.inf)
    df_result = df_result.sort_values(by=["rank", "score"], ascending=[True, False])
    df_result = df_result.reset_index()

    if truncated:
        df_result = pd.concat((df_result.head(5), df_result.tail(5)))
        df_result = df_result[~df_result.index.duplicated(keep="last")]

    pairwise = evalica.pairwise_scores(df_result["score"].to_numpy())

    df_pairwise = pd.DataFrame(data=pairwise, index=df_result["item"], columns=df_result["item"])

    fig = visualize(df_pairwise)

    if "ci" in df_result.columns:
        df_result["ci"] = df_result.apply(
            lambda row: f"({row['score'] - row['ci'][0]:.03f}; {row['ci'][1] - row['score']:.03f})",
            axis=1,
        )

    df_result["score"] = df_result["score"].apply(lambda x: f"{x:.03f}")

    return df_result, fig


def visualize_alpha_ci(bootstrap_result: AlphaBootstrapResult) -> Figure:
    fig = px.histogram(
        bootstrap_result.distribution,
        nbins=50,
        labels={"value": "Alpha", "count": "Frequency"},
    )

    fig.add_vline(
        x=bootstrap_result.alpha,
        line_dash="dash",
        line_color="red",
        annotation_text=f"alpha = {bootstrap_result.alpha:.3f}",
        annotation_position="top right",
    )

    fig.add_vline(
        x=bootstrap_result.low,
        line_dash="dot",
        line_color="blue",
        annotation_text=f"low = {bootstrap_result.low:.3f}",
        annotation_position="top left",
    )

    fig.add_vline(
        x=bootstrap_result.high,
        line_dash="dot",
        line_color="blue",
        annotation_text=f"high = {bootstrap_result.high:.3f}",
        annotation_position="top right",
    )

    fig.update_layout(
        xaxis_title="Alpha",
        yaxis_title="Frequency",
        showlegend=False,
    )

    return fig


def alpha_handler(file: str | None, distance: str, rounds: int) -> tuple[pd.DataFrame, Figure | None]:
    if file is None:
        raise gr.Error("File must be uploaded")

    try:
        df_ratings = pd.read_csv(file, header=None, dtype=str)
    except ValueError as e:
        raise gr.Error(f"Parsing error: {e}") from e

    if df_ratings.empty:
        raise gr.Error("The file is empty")

    try:
        bootstrap_result = alpha_bootstrap(
            df_ratings,
            distance=distance,  # type: ignore[arg-type]
            n_resamples=rounds or 1,
        )
    except evalica.InsufficientRatingsError as e:
        raise gr.Error("Insufficient ratings: no units have at least 2 ratings") from e
    except evalica.UnknownDistanceError as e:
        raise gr.Error(f"Unknown distance: {e}") from e
    except Exception as e:
        raise gr.Error(f"Computation error: {e}") from e

    df_metrics = pd.DataFrame(
        {
            "Metric": ["Alpha", "Observed Disagreement", "Expected Disagreement"],
            "Value": [bootstrap_result.alpha, bootstrap_result.observed, bootstrap_result.expected],
        },
    )

    if not rounds:
        return df_metrics, None

    return df_metrics, visualize_alpha_ci(bootstrap_result)


def alpha_interface() -> gr.Interface:
    return gr.Interface(
        fn=alpha_handler,
        inputs=[
            gr.File(
                file_types=[".csv", ".tsv"],
                label="Ratings",
            ),
            gr.Dropdown(
                choices=["nominal", "ordinal", "interval", "ratio"],
                value="nominal",
                label="Distance",
                info="Nominal for categorical, ordinal for ordered categories, interval/ratio for numeric scales",
            ),
            gr.Number(
                value=0,
                minimum=0,
                maximum=10000,
                label="Bootstrap Rounds",
                info="Number of bootstrap resamples for the confidence interval plot. Set to 0 to skip.",
            ),
        ],
        outputs=[
            gr.Dataframe(
                headers=["Metric", "Value"],
                label="Inter-Rater Reliability",
            ),
            gr.Plot(
                label="Bootstrap Distribution of Alpha",
            ),
        ],
        examples=[
            ["codings.csv", "ordinal", 1000],
            ["gcl.csv", "nominal", 1000],
        ],
        title="Krippendorff's Alpha",
        article=(
            f"""
This tool computes Krippendorff's alpha, an inter-rater reliability coefficient.

As an input, it expects a comma-separated (CSV) file without a header:
rows are raters (observers), columns are units (items), and cell values are ratings.

As the output, this tool provides alpha together with observed and expected disagreement.

{MORE_EVALICA}
            """.strip()
        ),
        analytics_enabled=False,
        flagging_mode="never",
    )


def main() -> None:
    pairwise_iface = gr.Interface(
        fn=handler,
        inputs=[
            gr.File(
                file_types=[".tsv", ".csv"],
                label="Comparisons",
            ),
            gr.Dropdown(
                choices=list(ALGORITHMS),
                value="Bradley-Terry (1952)",
                label="Algorithm",
            ),
            gr.Checkbox(
                value=False,
                label="Largest SCC",
                info="Bradley-Terry, Eigenvector, and Newman algorithms require the comparison graph "
                "to be strongly-connected. "
                "This option keeps only the largest strongly-connected component (SCC) of the input graph. "
                "Some items might be missing as a result of this filtering.",
            ),
            gr.Checkbox(
                value=False,
                label="Truncate Output",
                info="Perform the entire computation but output only five head and five tail items, avoiding overlap.",
            ),
            gr.Number(
                value=0,
                minimum=0,
                maximum=10000,
                label="Bootstrap Rounds",
                info="Number of bootstrap rounds to perform for estimating the confidence interval.",
            ),
        ],
        outputs=[
            gr.Dataframe(
                headers=["item", "score", "ci", "pairs", "rank"],
                label="Ranking",
            ),
            gr.Plot(
                label="Pairwise Chances of Winning the Comparison",
            ),
        ],
        examples=[
            ["food.csv", "Counting", False, False, 0],
            ["food.csv", "Bradley-Terry (1952)", False, False, 1000],
            ["food.csv", "Eigenvector (1987)", False, False, 1000],
            ["food.csv", "PageRank (1998)", False, False, 1000],
            ["food.csv", "Newman (2023)", False, False, 1000],
            ["llmfao.csv", "Average Win Rate", False, True, 100],
            ["llmfao.csv", "Bradley-Terry (1952)", False, True, 100],
            ["llmfao.csv", "Elo (1960)", False, True, 100],
        ],
        title="Pairwise Comparisons",
        article=(
            f"""
This easy-to-use tool transforms pairwise comparisons (*aka* side-by-side) to a meaningful ranking of items.

As an input, it expects a comma-separated (CSV) file with a header containing the following columns:

- `left`: the first compared item
- `right`: the second compared item
- `winner`: the label indicating the winning item

Possible values for `winner` are `left`, `right`, or `tie`. The provided examples might be a good starting point.

As the output, this tool provides a table with items, their estimated scores, and ranks.

{MORE_EVALICA}
            """.strip()
        ),
        flagging_mode="never",
        analytics_enabled=False,
    )

    iface = gr.TabbedInterface(
        [pairwise_iface, alpha_interface()],
        ["Pairwise Ranking", "Krippendorff's Alpha"],
        title="Evalica",
        analytics_enabled=False,
    )

    iface.launch()


if __name__ == "__main__":
    main()