app.py

# app.py
"""
Gradio application entrypoint for Hugging Face Spaces.
"""

import os
import tempfile
import pandas as pd
import gradio as gr
from evaluator import evaluate_dataframe
from synthetic_data import generate_synthetic_dataset

# Helper to save uploaded file to local temp path (gradio File gives a NamedTemporaryFile-like object)
def save_uploaded(file_obj):
    if not file_obj:
        return None
    # file_obj can be a dictionary or a file-like object depending on Gradio version
    try:
        path = file_obj.name
        return path
    except Exception:
        # fallback: write bytes to temp file
        data = file_obj.read()
        suffix = ".csv" if file_obj.name.endswith(".csv") else ".json"
        fd, tmp = tempfile.mkstemp(suffix=suffix)
        with os.fdopen(fd, "wb") as f:
            f.write(data)
        return tmp

def load_file_to_df(path):
    if path is None:
        return None
    # Try CSV
    try:
        if path.endswith(".csv"):
            return pd.read_csv(path)
        # JSONL
        try:
            return pd.read_json(path, lines=True)
        except ValueError:
            return pd.read_json(path)
    except Exception as e:
        # As last resort, raise
        raise e

def run_evaluation(file_obj):
    # If no file provided, use synthetic demo
    if file_obj is None:
        df = generate_synthetic_dataset(num_agents=3, num_samples=12)
    else:
        path = save_uploaded(file_obj)
        df = load_file_to_df(path)

    # Ensure required columns exist; otherwise, attempt to map common alternatives
    if df is None:
        return None, "No data loaded", None

    # Try to normalize column names
    cols = {c.lower(): c for c in df.columns}
    # rename common variants
    rename_map = {}
    for k in ["prompt", "response", "task", "agent", "reference"]:
        if k not in cols:
            # try variants
            if k == "reference":
                for alt in ["answer", "ground_truth", "ref"]:
                    if alt in cols:
                        rename_map[cols[alt]] = k
                        break
            else:
                for alt in [k, k.capitalize(), k.upper()]:
                    if alt.lower() in cols:
                        rename_map[cols[alt.lower()]] = k
    if rename_map:
        df = df.rename(columns=rename_map)

    metrics_df, images, leaderboard = evaluate_dataframe(df)

    # Prepare gallery (list of image file paths). Gradio Gallery accepts list of image paths or PIL images.
    gallery_items = [p for (p, caption) in images]
    captions = [caption for (p, caption) in images]

    # Save a CSV report for download
    out_csv = "/tmp/eval_results.csv"
    metrics_df.to_csv(out_csv, index=False)

    return (gallery_items, captions), metrics_df, leaderboard

# Build Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# Agentic Evaluation Framework")
    gr.Markdown(
        "Upload a CSV/JSON/JSONL with columns: `prompt,response,task,agent,reference` (reference optional). "
        "If no file is uploaded, a small synthetic demo will run."
    )

    with gr.Row():
        file_input = gr.File(label="Upload CSV / JSON / JSONL (optional)", file_types=[".csv", ".json", ".jsonl"])
        run_btn = gr.Button("Run Evaluation")
        download_report = gr.File(label="Download CSV Report")

    # ✅ Fixed Gallery (removed .style, added columns=2)
    gallery = gr.Gallery(
        label="Visualization Outputs",
        columns=2,
        height="auto"
    )
    table = gr.Dataframe(headers=None, label="Per-example Metrics (detailed)")
    leaderboard = gr.Dataframe(headers=None, label="Leaderboard (Avg Final Score per Agent & Task)")

    def on_run(file_in):
        (gallery_items, captions), metrics_df, lb = run_evaluation(file_in)
        # Save gallery captions mapping into a simple list of tuples for Gradio gallery (path, caption)
        gallery_display = []
        for i, p in enumerate(gallery_items):
            caption = captions[i] if i < len(captions) else ""
            gallery_display.append((p, caption))
        return gallery_display, metrics_df, lb

    run_btn.click(fn=on_run, inputs=[file_input], outputs=[gallery, table, leaderboard])

    gr.Markdown("## Usage tips\n- Columns: `prompt,response,task,agent,reference` (case-insensitive). "
                "- `reference` can be empty but accuracy/hallucination will be weaker.\n"
                "- Visualization images are available in the Gallery and a CSV report is downloadable.")

demo.launch()