consistency_eval_check.py

Go to the documentation of this file.

# Copyright (c) 2025-present Aquin Labs Private Limited. All Rights Reserved.
"""`aquin consistency-eval --check` — save JSON + diagram to the current working directory."""
 
from __future__ import annotations
 
import json
from datetime import datetime, timezone
from pathlib import Path
from typing import Any
 
_JSON_NAME = "consistency-eval-check.json"
_PNG_NAME = "consistency-eval-check.png"
 
 
def _normalize_consistency_eval_result(result: dict[str, Any]) -> dict[str, Any]:
    if not isinstance(result, dict):
        return {"error": "Invalid consistency-eval result"}
 
    if result.get("error"):
        return dict(result)
 
    data = result
    if isinstance(result.get("content"), dict):
        data = result["content"]
    if isinstance(data.get("consistency"), dict):
        data = data["consistency"]
    card = result.get("card")
    if isinstance(card, dict) and isinstance(card.get("data"), dict):
        nested = card["data"].get("consistency")
        if isinstance(nested, dict):
            data = nested
 
    return dict(data)
 
 
def has_consistency_eval_payload(result: dict[str, Any]) -> bool:
    data = _normalize_consistency_eval_result(result)
    if data.get("error"):
        return False
    return data.get("consistency_score") is not None or bool(data.get("variants"))
 
 
def write_consistency_eval_check(
    result: dict[str, Any],
    *,
    tool_name: str | None,
    cwd: str | Path,
) -> tuple[Path, Path]:
    cwd = Path(cwd)
    json_path = cwd / _JSON_NAME
    png_path = cwd / _PNG_NAME
 
    flat = _normalize_consistency_eval_result(result)
    payload = {
        "saved_at": datetime.now(timezone.utc).isoformat(),
        "tool": tool_name,
        **flat,
    }
    json_path.write_text(json.dumps(payload, indent=2, default=str), encoding="utf-8")
    _plot_consistency_eval_check(flat, png_path)
    return json_path, png_path
 
 
def _plot_consistency_eval_check(result: dict[str, Any], png_path: Path) -> None:
    import matplotlib
 
    matplotlib.use("Agg")
    import matplotlib.pyplot as plt
 
    if result.get("error"):
        fig, ax = plt.subplots(figsize=(6, 2), facecolor="#0f1117")
        ax.axis("off")
        ax.text(0.5, 0.5, f"Error: {result['error']}", ha="center", va="center", wrap=True, color="#e5e7eb")
        fig.savefig(png_path, dpi=140, bbox_inches="tight", facecolor="#0f1117")
        plt.close(fig)
        return
 
    query = _trunc(str(result.get("query") or ""), 56)
    score = result.get("consistency_score")
    mean_kl = result.get("mean_kl")
    max_kl = result.get("max_kl")
    variants = result.get("variants") or []
 
    title = "Consistency eval"
    if query:
        title += f' — "{query}"'
    subtitle_parts = []
    if mean_kl is not None:
        subtitle_parts.append(f"mean KL {float(mean_kl):.4f}")
    if max_kl is not None:
        subtitle_parts.append(f"max KL {float(max_kl):.4f}")
 
    fig, axes = plt.subplots(1, 2, figsize=(11, 4.5), facecolor="#0f1117")
    fig.suptitle(title, color="#e5e7eb", fontsize=11, y=1.03)
    if subtitle_parts:
        fig.text(0.5, 0.96, "  ·  ".join(subtitle_parts), ha="center", fontsize=8, color="#9ca3af")
 
    _plot_score_panel(axes[0], score)
    _plot_variants_panel(axes[1], variants)
 
    for ax in axes:
        ax.set_facecolor("#0f1117")
        for spine in ax.spines.values():
            spine.set_color("#374151")
 
    fig.tight_layout(rect=(0, 0, 1, 0.94))
    fig.savefig(png_path, dpi=140, bbox_inches="tight", facecolor="#0f1117")
    plt.close(fig)
 
 
def _plot_score_panel(ax: Any, score: Any) -> None:
    ax.set_title("Consistency score", color="#e5e7eb", fontsize=10)
 
    if score is None:
        ax.axis("off")
        ax.text(0.5, 0.5, "No score", ha="center", va="center", color="#9ca3af")
        return
 
    val = float(score)
    color = _score_color(val)
    ax.barh([0], [val], color=color, height=0.45, alpha=0.92)
    ax.set_xlim(0, 1.05)
    ax.set_yticks([0])
    ax.set_yticklabels(["score"], fontsize=9, color="#d1d5db")
    ax.set_xlabel("Score (0–1, higher = more consistent)", color="#9ca3af", fontsize=8)
    ax.text(min(val + 0.02, 0.98), 0, f"{val:.3f}", va="center", fontsize=9, color="#9ca3af")
    ax.tick_params(axis="x", colors="#6b7280", labelsize=8)
    ax.grid(axis="x", alpha=0.2, color="#4b5563")
 
 
def _plot_variants_panel(ax: Any, variants: list[dict[str, Any]]) -> None:
    ax.set_title("Paraphrase KL from anchor", color="#e5e7eb", fontsize=10)
 
    if not variants:
        ax.axis("off")
        ax.text(0.5, 0.5, "No variants", ha="center", va="center", color="#9ca3af")
        return
 
    labels = []
    kls = []
    for i, v in enumerate(variants):
        labels.append("anchor" if i == 0 else f"v{i}")
        kls.append(float(v.get("kl_from_anchor") or 0))
 
    colors = ["#78716c" if i == 0 else "#6366f1" for i in range(len(labels))]
    y_pos = list(range(len(labels)))
    ax.barh(y_pos, kls, color=colors, height=0.65, alpha=0.9)
    ax.set_yticks(y_pos)
    ax.set_yticklabels(labels, fontsize=9, color="#d1d5db")
    ax.invert_yaxis()
    ax.set_xlabel("KL divergence", color="#9ca3af", fontsize=8)
    ax.tick_params(axis="x", colors="#6b7280", labelsize=8)
    ax.grid(axis="x", alpha=0.2, color="#4b5563")
    for i, kl in enumerate(kls):
        if kl > 0:
            ax.text(kl + max(kls) * 0.02 + 0.001, i, f"{kl:.4f}", va="center", fontsize=7, color="#9ca3af")
 
 
def _score_color(val: float) -> str:
    if val >= 0.75:
        return "#34d399"
    if val >= 0.5:
        return "#fbbf24"
    return "#f87171"
 
 
def _trunc(text: str, limit: int) -> str:
    t = text.strip()
    if len(t) <= limit:
        return t or "·"
    return t[: limit - 1] + "…"