app.py

"""Epicure Explorer: chef-facing operators over the three sibling embeddings.

Features:
  - Basket pairings (with pairwise cosine heatmap)
  - Supervised SLERP (with "why these results" explainer)
  - Emergent SLERP (with explainer)
  - Arithmetic (Mikolov-style, with explainer)
  - Mode atlas (click row -> highlight on UMAP)
  - Compare siblings (one query, three columns)
  - UMAP visualisation (2D / 3D)
  - Parse my fridge (free-text -> canonical vocab via rapidfuzz)
  - Recipe builder (hybrid retrieval: rapidfuzz + sentence-transformers over mode labels)
  - Saved queries (per-browser persistence via gr.BrowserState)
  - Public developer API (gr.api endpoints for neighbours / slerp / arithmetic / embed)
  - Food-group filter on every ingredient dropdown

Paper: https://arxiv.org/abs/2605.22391
"""

from __future__ import annotations

import os
import re
import sys
import json
import uuid
from datetime import datetime, timezone
from functools import lru_cache

import numpy as np
import gradio as gr
import plotly.graph_objects as go
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

try:
    from epicure import Epicure
except ImportError:
    from huggingface_hub import hf_hub_download
    epicure_py = hf_hub_download("Kaikaku/epicure-cooc", "epicure.py")
    sys.path.insert(0, os.path.dirname(epicure_py))
    from epicure import Epicure

from rapidfuzz import process as fuzz_process, fuzz as fuzz_scorers

# ===== Kaikaku brand =====
KAIKAKU_DARK  = "#0F2D2F"
KAIKAKU_DEEP  = "#0A1F20"
KAIKAKU_MID   = "#1A3D3F"
KAIKAKU_EDGE  = "#2A4D4F"
KAIKAKU_ACCENT = "#288B79"
KAIKAKU_ACCENT_HOVER = "#1E6E5F"
KAIKAKU_ACCENT_LIGHT = "#A8D5CA"
KAIKAKU_TEXT  = "#0F2D2F"
KAIKAKU_MUTED = "#5A7878"

plt.rcParams.update({
    "figure.facecolor": "#ffffff",
    "axes.facecolor": "#ffffff",
    "axes.edgecolor": "#cccccc",
    "axes.labelcolor": "#111111",
    "xtick.color": "#333333",
    "ytick.color": "#333333",
    "text.color": "#111111",
    "savefig.facecolor": "#ffffff",
})

MODELS = {
    "cooc": Epicure.from_pretrained("Kaikaku/epicure-cooc"),
    "core": Epicure.from_pretrained("Kaikaku/epicure-core"),
    "chem": Epicure.from_pretrained("Kaikaku/epicure-chem"),
}
ALL_INGREDIENTS = sorted(MODELS["cooc"].vocab.keys())

_HERE = os.path.dirname(os.path.abspath(__file__))
UMAP_DATA = np.load(os.path.join(_HERE, "umap_2d.npz"))
_lab = json.load(open(os.path.join(_HERE, "ingredient_labels.json")))
NAMES_BY_IDX: list[str] = _lab["names"]
FOOD_GROUPS: list[str]  = _lab["food_groups"]

FG_COLORS = {
    "Vegetable":  "#2ca02c",
    "Fruit":      "#e377c2",
    "Grain":      "#bcbd22",
    "Dairy":      "#17becf",
    "Spice":      "#d62728",
    "Pantry":     "#ff7f0e",
    "Beverage":   "#9467bd",
    "Other":      "#cccccc",
}

print(f"[epicure-explorer] models loaded: {list(MODELS)}", flush=True)
print(f"[epicure-explorer] food group labels: {len(FOOD_GROUPS)} ingredients", flush=True)

# ===== Feature 5: food-group filter helpers =====

_NAME_TO_GROUP: dict[str, str] = {NAMES_BY_IDX[i]: FOOD_GROUPS[i] for i in range(len(NAMES_BY_IDX))}
FOOD_GROUP_CHOICES = ["All", "Vegetable", "Spice", "Fruit", "Dairy", "Grain", "Pantry", "Beverage", "Other"]

def _choices_for_group(group: str) -> list[str]:
    if not group or group == "All":
        return ALL_INGREDIENTS
    return sorted(n for n in ALL_INGREDIENTS if _NAME_TO_GROUP.get(n, "Other") == group)

def _filter_dropdown(group: str, current_value):
    new_choices = _choices_for_group(group)
    allowed = set(new_choices)
    cur = current_value or []
    if isinstance(cur, str):
        kept = cur if cur in allowed else None
    else:
        kept = [v for v in cur if v in allowed]
    return gr.Dropdown(choices=new_choices, value=kept)

# ===== math helpers =====

def _unit(v, eps=1e-9):
    n = np.linalg.norm(v); return v / max(n, eps)

def _basket_centroid(m, names):
    valid = [n for n in (names or []) if n in m.vocab]
    if not valid: return None
    return _unit(m.E[[m.vocab[n] for n in valid]].mean(axis=0))

def _stack_directions(m, keys, use_factor_pole=False):
    poles = []
    for k in keys or []:
        if use_factor_pole:
            for mode in m.modes:
                if mode.mode_id == k:
                    poles.append(_unit(mode.pole)); break
        else:
            if k in m.supervised_poles:
                poles.append(_unit(m.supervised_poles[k]))
    if not poles: return None
    return _unit(np.stack(poles, axis=0).sum(axis=0))

def _topk(m, q, k, exclude):
    sims = m.E @ q
    for n in exclude or []:
        if n in m.vocab: sims[m.vocab[n]] = -np.inf
    order = np.argsort(-sims)
    return [(m.itos[int(i)], float(sims[i])) for i in order[:k]]

def _supervised_choices(sibling):
    return sorted(MODELS[sibling].supervised_poles.keys())

def _factor_mode_choices(sibling):
    return [(f"{m.label} ({m.mode_id})", m.mode_id) for m in MODELS[sibling].modes if m.kind == "factor"]

def _slerp(v, d, theta_deg):
    d_perp = d - (d @ v) * v
    n = np.linalg.norm(d_perp)
    if n < 1e-9: return v
    d_perp = d_perp / n
    th = np.deg2rad(float(theta_deg))
    return _unit(np.cos(th)*v + np.sin(th)*d_perp)

# ===== Feature 4: explainer helpers =====

def _fmt_nb_inline(pairs):
    return ", ".join(f"{n} ({s:+.2f})" for n, s in pairs)

def _slerp_explainer(m, basket, direction_keys, theta, q, v, d, kind):
    if v is None or d is None or q is None:
        return "_(no rotation applied)_"
    cos_theta = float(q @ v)
    travelled = min(max(float(theta) / 90.0, 0.0), 1.0)
    dir_nb = _topk(m, _unit(d), k=5, exclude=basket or [])
    seed_nb = _topk(m, v, k=3, exclude=basket or [])
    dir_names = ", ".join(n for n, _ in dir_nb[:3])
    label = "direction pole" if kind == "supervised" else "factor-mode pole"
    dirs_str = " + ".join(direction_keys) if direction_keys else "(none)"
    return (
        f"**Why these results**  \n"
        f"- Rotated query vs. seed centroid: cos = {cos_theta:.3f} (theta = {float(theta):.0f}°; "
        f"{travelled*100:.0f}% of the way to the {label}).  \n"
        f"- {label.capitalize()} ({dirs_str}) nearest in vocab: {_fmt_nb_inline(dir_nb)}.  \n"
        f"- Seed basket's own top-3 (baseline): {_fmt_nb_inline(seed_nb)}.  \n"
        f"- At {float(theta):.0f}° the query lands near: {dir_names}."
    )

def _arithmetic_explainer(m, positives, negatives, q, pos_v, neg_v):
    if q is None:
        return "_(no result: missing positives)_"
    pos_sims = [(n, float(_unit(m.E[m.vocab[n]]) @ q)) for n in (positives or []) if n in m.vocab]
    neg_sims = [(n, float(_unit(m.E[m.vocab[n]]) @ q)) for n in (negatives or []) if n in m.vocab]
    top = _topk(m, q, k=1, exclude=(positives or []) + (negatives or []))
    top_name, top_sim = top[0] if top else ("(none)", 0.0)
    pos_part = ", ".join(f"{n} ({s:+.2f})" for n, s in pos_sims) or "(none)"
    neg_part = ", ".join(f"{n} ({s:+.2f})" for n, s in neg_sims) or "(none)"
    input_max = max((s for _, s in pos_sims + neg_sims), default=0.0)
    if pos_sims or neg_sims:
        gap = top_sim - input_max
        if gap > 0.05:
            interp = (f"Result sits closer to **{top_name}** ({top_sim:+.2f}) "
                      f"than to any input (max {input_max:+.2f}); the embedding separates these concepts.")
        else:
            interp = (f"Result is dominated by the inputs themselves "
                      f"(top neighbour {top_name} only {gap:+.2f} above max input cosine).")
    else:
        interp = f"Result top neighbour: {top_name} ({top_sim:+.2f})."
    return (
        f"**Why these results**  \n"
        f"- Result vs. positives: {pos_part}.  \n"
        f"- Result vs. negatives: {neg_part}.  \n"
        f"- {interp}"
    )

# ===== heatmap =====

def _basket_heatmap(m, basket):
    valid = [n for n in (basket or []) if n in m.vocab]
    fig, ax = plt.subplots(figsize=(6, 5))
    if len(valid) < 2:
        ax.text(0.5, 0.5, "Add 2+ ingredients to see pairwise cosines",
                ha="center", va="center", fontsize=13, color="#888",
                transform=ax.transAxes)
        ax.axis("off")
        plt.tight_layout()
        return fig
    idxs = [m.vocab[n] for n in valid]
    sub = m.E[idxs]
    sim = sub @ sub.T
    im = ax.imshow(sim, cmap="viridis", vmin=-0.2, vmax=1.0, aspect="auto")
    ax.set_xticks(range(len(valid)))
    ax.set_yticks(range(len(valid)))
    ax.set_xticklabels(valid, rotation=35, ha="right")
    ax.set_yticklabels(valid)
    for i in range(len(valid)):
        for j in range(len(valid)):
            v = float(sim[i, j])
            color = "white" if v < 0.55 else "black"
            ax.text(j, i, f"{v:.2f}", ha="center", va="center", fontsize=10, color=color)
    cb = plt.colorbar(im, ax=ax)
    cb.set_label("cosine")
    ax.set_title("Pairwise cosine within the basket", fontsize=12)
    plt.tight_layout()
    return fig

# ===== UMAP =====

def _umap_coords(sibling, three_d):
    base = UMAP_DATA[sibling]
    if not three_d:
        return base, None
    m = MODELS[sibling]
    E = m.E - m.E.mean(axis=0, keepdims=True)
    _, _, Vt = np.linalg.svd(E, full_matrices=False)
    pc1 = (E @ Vt[0])
    pc1 = (pc1 - pc1.mean()) / (pc1.std() + 1e-9)
    scale = (base.max() - base.min()) * 0.25
    return base, (pc1 * scale).astype(np.float32)

def umap_view(sibling, basket, show_neighbours, k, three_d=False):
    coords2, z = _umap_coords(sibling, three_d)
    m = MODELS[sibling]
    n = len(NAMES_BY_IDX)
    colors = [FG_COLORS.get(fg, "#cccccc") for fg in FOOD_GROUPS]
    hover_text = [f"{NAMES_BY_IDX[i]}<br>group: {FOOD_GROUPS[i]}" for i in range(n)]
    basket_set = set(basket or [])
    basket_idxs = [m.vocab[b] for b in (basket or []) if b in m.vocab]
    neighbour_set: set[str] = set()
    if show_neighbours and basket_idxs:
        centroid = _basket_centroid(m, basket)
        if centroid is not None:
            nb_pairs = _topk(m, centroid, k=int(k), exclude=basket)
            neighbour_set = {nm for nm, _ in nb_pairs}
    bg_keep = lambda i: NAMES_BY_IDX[i] not in basket_set and NAMES_BY_IDX[i] not in neighbour_set
    bg_x = [float(coords2[i, 0]) for i in range(n) if bg_keep(i)]
    bg_y = [float(coords2[i, 1]) for i in range(n) if bg_keep(i)]
    bg_z = [float(z[i]) for i in range(n) if bg_keep(i)] if three_d else None
    bg_c = [colors[i] for i in range(n) if bg_keep(i)]
    bg_h = [hover_text[i] for i in range(n) if bg_keep(i)]
    fig = go.Figure()
    if three_d:
        fig.add_trace(go.Scatter3d(
            x=bg_x, y=bg_y, z=bg_z, mode="markers",
            marker=dict(size=3, color=bg_c, opacity=0.55, line=dict(width=0)),
            text=bg_h, hovertemplate="%{text}<extra></extra>", name="ingredients", showlegend=False,
        ))
    else:
        fig.add_trace(go.Scattergl(
            x=bg_x, y=bg_y, mode="markers",
            marker=dict(size=5, color=bg_c, opacity=0.65, line=dict(width=0)),
            text=bg_h, hovertemplate="%{text}<extra></extra>", name="ingredients", showlegend=False,
        ))
    if neighbour_set:
        ni = [i for i in range(n) if NAMES_BY_IDX[i] in neighbour_set]
        nx = [float(coords2[i, 0]) for i in ni]
        ny = [float(coords2[i, 1]) for i in ni]
        nz = [float(z[i]) for i in ni] if three_d else None
        nlabels = [NAMES_BY_IDX[i] for i in ni]
        marker = dict(size=11 if not three_d else 6, color="#ff8800", opacity=0.95,
                      line=dict(color="#ffffff", width=1.2))
        TR = go.Scatter3d if three_d else go.Scatter
        kwargs = dict(mode="markers+text", marker=marker, text=nlabels, textposition="top center",
                      textfont=dict(size=10),
                      hovertemplate="<b>%{text}</b> (neighbour)<extra></extra>",
                      name=f"top-{k} neighbours")
        fig.add_trace(TR(x=nx, y=ny, z=nz, **kwargs) if three_d else TR(x=nx, y=ny, **kwargs))
    if basket_idxs:
        bx = [float(coords2[i, 0]) for i in basket_idxs]
        by = [float(coords2[i, 1]) for i in basket_idxs]
        bz = [float(z[i]) for i in basket_idxs] if three_d else None
        blabels = [NAMES_BY_IDX[i] for i in basket_idxs]
        marker = dict(size=18 if not three_d else 9, color=KAIKAKU_ACCENT,
                      symbol="star" if not three_d else "diamond",
                      line=dict(color="#111111", width=1.5))
        TR = go.Scatter3d if three_d else go.Scatter
        kwargs = dict(mode="markers+text", marker=marker, text=blabels, textposition="top center",
                      textfont=dict(size=13, color="#111111"),
                      hovertemplate="<b>%{text}</b> (basket)<extra></extra>", name="basket")
        fig.add_trace(TR(x=bx, y=by, z=bz, **kwargs) if three_d else TR(x=bx, y=by, **kwargs))
    title_suffix = " (3D)" if three_d else ""
    fig.update_layout(
        title=dict(text=f"UMAP of Epicure-{sibling.capitalize()}{title_suffix} - {n} ingredients", font=dict(size=15)),
        height=650, margin=dict(l=40, r=40, t=60, b=40),
        paper_bgcolor="#ffffff", plot_bgcolor="#ffffff",
        legend=dict(orientation="v", x=1.02, y=1, font=dict(size=11)),
    )
    if not three_d:
        fig.update_xaxes(showgrid=True, gridcolor="#eeeeee", zeroline=False, title="UMAP 1")
        fig.update_yaxes(showgrid=True, gridcolor="#eeeeee", zeroline=False, title="UMAP 2")
    else:
        fig.update_layout(scene=dict(xaxis=dict(title="UMAP 1"), yaxis=dict(title="UMAP 2"),
                                     zaxis=dict(title="PC1 (z)"), bgcolor="#ffffff"))
    return fig

# ===== tab handlers (with explainers) =====

def basket_pairings(sibling, basket, k):
    m = MODELS[sibling]
    centroid = _basket_centroid(m, basket)
    if centroid is None:
        return [], [], _basket_heatmap(m, [])
    nb = _topk(m, centroid, k, exclude=basket or [])
    scored = [(mode.mode_id, mode.label, mode.kind, float(_unit(mode.pole) @ centroid)) for mode in m.modes]
    scored.sort(key=lambda x: -x[3])
    heatmap = _basket_heatmap(m, basket)
    return (
        [[name, f"{sim:.4f}"] for name, sim in nb],
        [[mid, label, kind, f"{sim:.4f}"] for mid, label, kind, sim in scored[:k]],
        heatmap,
    )

def supervised_slerp_multi(sibling, basket, directions, theta, k):
    m = MODELS[sibling]
    v = _basket_centroid(m, basket)
    if v is None:
        return [], "_(empty basket)_"
    d = _stack_directions(m, directions, use_factor_pole=False)
    if d is None:
        return [[n, f"{s:.4f}"] for n, s in _topk(m, v, k, basket)], "_(no direction selected)_"
    q = _slerp(v, d, theta)
    rows = [[n, f"{s:.4f}"] for n, s in _topk(m, q, k, basket)]
    return rows, _slerp_explainer(m, basket, directions or [], theta, q, v, d, "supervised")

def emergent_slerp_multi(sibling, basket, mode_labels, theta, k):
    m = MODELS[sibling]
    label_to_id = {f"{mode.label} ({mode.mode_id})": mode.mode_id for mode in m.modes if mode.kind == "factor"}
    mode_ids = [label_to_id[lab] for lab in (mode_labels or []) if lab in label_to_id]
    v = _basket_centroid(m, basket)
    if v is None:
        return [], "_(empty basket)_"
    d = _stack_directions(m, mode_ids, use_factor_pole=True)
    if d is None:
        return [[n, f"{s:.4f}"] for n, s in _topk(m, v, k, basket)], "_(no factor mode selected)_"
    q = _slerp(v, d, theta)
    rows = [[n, f"{s:.4f}"] for n, s in _topk(m, q, k, basket)]
    return rows, _slerp_explainer(m, basket, mode_ids, theta, q, v, d, "emergent")

def arithmetic(sibling, positives, negatives, k):
    m = MODELS[sibling]
    pos = _basket_centroid(m, positives)
    if pos is None:
        return [], "_(no positives provided)_"
    neg = _basket_centroid(m, negatives) if negatives else None
    q = _unit(pos - neg) if neg is not None else pos
    rows = [[n, f"{s:.4f}"] for n, s in _topk(m, q, k, (positives or []) + (negatives or []))]
    return rows, _arithmetic_explainer(m, positives or [], negatives or [], q, pos, neg)

def browse_modes(sibling, kind_filter, query):
    m = MODELS[sibling]
    rows, q = [], (query or "").strip().lower()
    for mode in m.modes:
        if kind_filter != "all" and mode.kind != kind_filter:
            continue
        if q and q not in mode.label.lower() and q not in mode.property.lower():
            continue
        rows.append([mode.mode_id, mode.kind, mode.property, mode.label, mode.n_members,
                     ", ".join(mode.members[:12])])
    rows.sort(key=lambda r: (r[1], -r[4]))
    return rows

def compare_siblings(basket, directions, theta, k):
    out = []
    for sib in ["cooc","core","chem"]:
        m = MODELS[sib]
        v = _basket_centroid(m, basket)
        if v is None: out.append([]); continue
        valid_dirs = [d for d in (directions or []) if d in m.supervised_poles]
        if valid_dirs:
            d_vec = _stack_directions(m, valid_dirs)
            q = _slerp(v, d_vec, theta) if d_vec is not None else v
        else:
            q = v
        hits = _topk(m, q, k=k, exclude=basket)
        out.append([[n, f"{s:.4f}"] for n, s in hits])
    return out[0], out[1], out[2]

# ===== Feature 6: recipe builder (lazy-loaded sentence-transformer) =====

_ST_MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
_ST = None
def _get_st():
    global _ST
    if _ST is None:
        print(f"[epicure-explorer] loading {_ST_MODEL_NAME} (first call, ~80MB)", flush=True)
        from sentence_transformers import SentenceTransformer
        _ST = SentenceTransformer(_ST_MODEL_NAME, device="cpu")
    return _ST

@lru_cache(maxsize=4)
def _mode_label_matrix(sibling: str):
    m = MODELS[sibling]
    modes = [md for md in m.modes if md.kind == "factor"]
    if not modes:
        return [], [], np.zeros((0, 384), dtype=np.float32)
    labels = [md.label for md in modes]
    mids = [md.mode_id for md in modes]
    M = _get_st().encode(labels, normalize_embeddings=True, convert_to_numpy=True)
    return mids, labels, M.astype(np.float32)

def _mode_quartile(mode):
    members = list(mode.members or [])
    n = max(4, min(12, (len(members) + 3) // 4))
    return members[:n]

_PROMPT_STOPWORDS = {
    "i","im","i'm","a","an","the","for","of","with","and","or","some","my","me","we",
    "make","making","cook","cooking","prepare","preparing","want","need","to","tonight",
    "people","person","servings","dinner","lunch","dish","recipe","quick","easy",
    "tasty","yummy","good","great","food","meal","style","plate","plates",
}
_TOKEN_RE = re.compile(r"[A-Za-z][A-Za-z\-']{1,}")

def suggest_basket(prompt, sibling, k=10):
    if not prompt or not prompt.strip():
        return [], [], "Type a dish description first."
    vocab = list(MODELS[sibling].vocab.keys())
    vocab_sp = [v.replace("_", " ") for v in vocab]
    raw_tokens = _TOKEN_RE.findall(prompt.lower())
    tokens = [t for t in raw_tokens if t not in _PROMPT_STOPWORDS and len(t) > 2]
    direct = {}
    direct_evidence = []
    for tok in tokens:
        hits = fuzz_process.extract(tok, vocab_sp, scorer=fuzz_scorers.token_set_ratio,
                                    score_cutoff=88, limit=2)
        for _sp, score, idx in hits:
            name = vocab[idx]
            if score > direct.get(name, 0):
                direct[name] = float(score)
                direct_evidence.append((tok, name, float(score)))
    mids, labels, M = _mode_label_matrix(sibling)
    thematic = {}
    thematic_modes = []
    if M.shape[0] > 0:
        q = _get_st().encode([prompt], normalize_embeddings=True, convert_to_numpy=True)[0]
        sims = M @ q
        order = np.argsort(-sims)
        picked = [(mids[i], labels[i], float(sims[i])) for i in order[:3] if sims[i] >= 0.25]
        thematic_modes = picked
        id_to_mode = {md.mode_id: md for md in MODELS[sibling].modes if md.kind == "factor"}
        for mid, lab, sim in picked:
            for name in _mode_quartile(id_to_mode[mid]):
                s_existing, _ = thematic.get(name, (0.0, ""))
                s_new = max(s_existing, sim * 100.0)
                thematic[name] = (s_new, lab)
    combined = {}
    for name, sc in direct.items():
        combined[name] = (sc, "direct")
    for name, (sc, lab) in thematic.items():
        prev = combined.get(name)
        if prev is None or sc > prev[0]:
            tag = "both" if prev else "thematic"
            combined[name] = (sc, tag)
    ranked = sorted(combined.items(),
                    key=lambda kv: (-kv[1][0], 0 if kv[1][1] != "thematic" else 1, kv[0]))[:int(k)]
    rows = [[name, src, round(score, 1)] for name, (score, src) in ranked]
    names = [name for name, _ in ranked]
    lines = []
    if direct_evidence:
        dm = ", ".join(sorted({f"`{n}` (from '{t}')" for t, n, _ in direct_evidence}))
        lines.append(f"**Direct mentions:** {dm}")
    else:
        lines.append("**Direct mentions:** _none cleared score threshold_")
    if thematic_modes:
        bits = []
        id_to_mode = {md.mode_id: md for md in MODELS[sibling].modes if md.kind == "factor"}
        for mid, lab, sim in thematic_modes:
            sample = ", ".join(id_to_mode[mid].members[:4])
            bits.append(f"`{lab}` (cos {sim:.2f}; e.g. {sample})")
        lines.append("**Matched factor modes:** " + "; ".join(bits))
    else:
        lines.append("**Matched factor modes:** _no mode label cleared cosine 0.25_")
    return rows, names, "\n\n".join(lines)

# ===== fridge parser =====

_LINE_SPLIT = re.compile(r"[\n;]")
_BRACKET = re.compile(r"\([^)]*\)")
_QTY = (r"(?:\d+(?:[\.,/]\d+)?|a|an|one|two|three|four|five|six|seven|eight|nine|ten|half|quarter)")
_UNIT = (r"(?:cups?|tbsp\.?|tablespoons?|tsp\.?|teaspoons?|oz\.?|ounces?|lbs?\.?|pounds?|"
         r"grams?|kgs?|kilos?|ml|liters?|litres?|cloves?|bunches?|sprigs?|pinch(?:es)?|"
         r"slices?|pieces?|cans?|packets?|sticks?|leaves?|stalks?|heads?|inch(?:es)?|"
         r"splash(?:es)?|dash(?:es)?|drops?|handfuls?|large|small|medium)")
_LEADING_QTY = re.compile(rf"^\s*{_QTY}\s+(?:{_UNIT}\b\s*)?(?:of\s+)?", re.IGNORECASE)
_LEADING_UNIT_ONLY = re.compile(rf"^\s*{_UNIT}\b\s*(?:of\s+)?", re.IGNORECASE)
_JUICE_OF = re.compile(rf"^\s*(?:juice|zest)\s+(?:of\s+)?(?:{_QTY}\s+)?", re.IGNORECASE)
_LEADING_PREP = re.compile(
    r"^\s*(?:fresh|dried|cooked|frozen|raw|ripe|firm|boneless|skinless|smoked|low[- ]fat)\s+",
    re.IGNORECASE)
_TRAILING_PREP = re.compile(
    r"\s*,\s*(?:chopped|minced|diced|sliced|grated|crushed|whole|ground|peeled|"
    r"to taste|optional|finely|coarsely|cubed|shredded|julienned|halved|quartered|warmed|"
    r"toasted|roasted|bruised|melted|softened|cooked|drained|rinsed|patted dry|trimmed|"
    r"deveined|seeded|stemmed|crumbled).*$", re.IGNORECASE)
_KNOWN_PLURALS = {"tortillas":"tortilla","thighs":"thigh","leaves":"leaf","onions":"onion",
                  "potatoes":"potato","tomatoes":"tomato","cloves":"clove"}

def _clean_line(line):
    s = line.strip().lower()
    s = _BRACKET.sub(" ", s)
    if "juice" in s or "zest" in s:
        s = _JUICE_OF.sub("", s)
    s = _TRAILING_PREP.sub("", s)
    s = _LEADING_QTY.sub("", s)
    s = _LEADING_UNIT_ONLY.sub("", s)
    s = _LEADING_PREP.sub("", s)
    s = _LEADING_PREP.sub("", s)
    tokens = [_KNOWN_PLURALS.get(t, t) for t in s.split()]
    return re.sub(r"\s+", " ", " ".join(tokens)).strip()

def _fuzzy_lookup(cleaned, vocab, vocab_sp, min_score):
    if not cleaned: return None, 0.0
    candidates = []
    for scorer in (fuzz_scorers.token_set_ratio, fuzz_scorers.WRatio, fuzz_scorers.partial_ratio):
        hits = fuzz_process.extract(cleaned, vocab_sp, scorer=scorer, score_cutoff=min_score, limit=10)
        for _name_sp, score, idx in hits:
            candidates.append((vocab[idx], float(score)))
    if not candidates: return None, 0.0
    cleaned_tokens = set(cleaned.split())
    def rank_key(c):
        name, score = c
        nt = set(name.replace("_"," ").split())
        return (-score, 0 if nt.issubset(cleaned_tokens) else 1, -len(name))
    candidates.sort(key=rank_key)
    return candidates[0]

def parse_fridge(raw_text, sibling, min_score=70):
    if not raw_text or not raw_text.strip(): return [], []
    vocab = list(MODELS[sibling].vocab.keys())
    vocab_sp = [v.replace("_"," ") for v in vocab]
    rows, matched = [], []
    for line in _LINE_SPLIT.split(raw_text):
        if not line.strip(): continue
        cleaned = _clean_line(line)
        if not cleaned:
            rows.append([line.strip(), "(empty)", 0.0, ""]); continue
        match, score = _fuzzy_lookup(cleaned, vocab, vocab_sp, int(min_score))
        if match is None:
            tokens = cleaned.split()
            if len(tokens) > 1:
                match, score = _fuzzy_lookup(" ".join(tokens[:-1]), vocab, vocab_sp, int(min_score))
        if match is None:
            rows.append([line.strip(), "(no match)", 0.0, cleaned]); continue
        rows.append([line.strip(), match, round(score, 1), cleaned])
        matched.append(match)
    seen, dedup = set(), []
    for n in matched:
        if n not in seen: seen.add(n); dedup.append(n)
    return rows, dedup

# ===== Feature 8: public API endpoints =====

def _suggest(name: str, sibling: str, n: int = 5) -> list[str]:
    vocab = list(MODELS[sibling].vocab.keys())
    hits = fuzz_process.extract((name or "").lower().replace(" ", "_"),
                                vocab, scorer=fuzz_scorers.WRatio, limit=n)
    return [h[0] for h in hits]

def _validate_sibling(sibling):
    if sibling not in MODELS:
        return {"error": f"sibling '{sibling}' not in {{cooc, core, chem}}",
                "suggestions": ["cooc","core","chem"]}
    return None

def _validate_ingredient(name, sibling, field="ingredient"):
    if not isinstance(name, str) or not name:
        return {"error": f"{field} must be a non-empty string"}
    if name not in MODELS[sibling].vocab:
        return {"error": f"{field} '{name}' not in vocab",
                "suggestions": _suggest(name, sibling)}
    return None

def api_neighbors(ingredient, sibling="chem", k=5):
    err = _validate_sibling(sibling) or _validate_ingredient(ingredient, sibling)
    if err: return err
    m = MODELS[sibling]
    q = _unit(m.E[m.vocab[ingredient]])
    pairs = _topk(m, q, int(k), exclude=[ingredient])
    return [{"name": n, "cosine": round(float(s), 6)} for n, s in pairs]

def api_slerp(seed, direction, theta_deg=30, sibling="chem", k=5):
    err = _validate_sibling(sibling) or _validate_ingredient(seed, sibling, "seed")
    if err: return err
    m = MODELS[sibling]
    if direction not in m.supervised_poles:
        return {"error": f"direction '{direction}' not a supervised pole",
                "suggestions": sorted(m.supervised_poles.keys())[:10]}
    v = _unit(m.E[m.vocab[seed]])
    d = _unit(m.supervised_poles[direction])
    q = _slerp(v, d, float(theta_deg))
    pairs = _topk(m, q, int(k), exclude=[seed])
    return [{"name": n, "cosine": round(float(s), 6)} for n, s in pairs]

def api_arithmetic(positives, negatives, sibling="chem", k=5):
    err = _validate_sibling(sibling)
    if err: return err
    positives = list(positives or [])
    negatives = list(negatives or [])
    if not positives:
        return {"error": "positives must be a non-empty list"}
    m = MODELS[sibling]
    unknown = [x for x in positives + negatives if x not in m.vocab]
    if unknown:
        return {"error": f"unknown ingredients: {unknown}",
                "suggestions": {x: _suggest(x, sibling) for x in unknown}}
    pos = _basket_centroid(m, positives)
    neg = _basket_centroid(m, negatives) if negatives else None
    q = _unit(pos - neg) if neg is not None else pos
    pairs = _topk(m, q, int(k), exclude=positives + negatives)
    return [{"name": n, "cosine": round(float(s), 6)} for n, s in pairs]

def api_embed(ingredient, sibling="chem"):
    err = _validate_sibling(sibling) or _validate_ingredient(ingredient, sibling)
    if err: return err
    m = MODELS[sibling]
    v = _unit(m.E[m.vocab[ingredient]])
    return [float(x) for x in v.tolist()]

def api_list_directions(sibling="chem"):
    err = _validate_sibling(sibling)
    if err: return err
    return sorted(MODELS[sibling].supervised_poles.keys())

def api_list_factor_modes(sibling="chem"):
    err = _validate_sibling(sibling)
    if err: return err
    return [{"mode_id": mode.mode_id, "label": str(mode.label),
             "kind": str(mode.kind), "property": str(mode.property),
             "n_members": int(mode.n_members)}
            for mode in MODELS[sibling].modes if mode.kind == "factor"]

# ===== Feature 9: saved queries helpers =====

TAB_IDS = {
    "basket": "tab_basket",
    "supervised_slerp": "tab_sup",
    "emergent_slerp": "tab_em",
    "arithmetic": "tab_ar",
    "compare": "tab_cmp",
}
TAB_LABELS = {
    "basket": "Basket pairings",
    "supervised_slerp": "Supervised SLERP",
    "emergent_slerp": "Emergent SLERP",
    "arithmetic": "Arithmetic",
    "compare": "Compare siblings",
}

def _summarise(tab, inputs):
    sib = inputs.get("sibling", "")
    if tab == "basket":
        return f"[{sib}] basket: {', '.join(inputs.get('basket', [])[:3])} k={inputs.get('k')}"
    if tab == "supervised_slerp":
        b = ", ".join(inputs.get("basket", [])[:2])
        d = ", ".join(inputs.get("directions", [])[:2])
        return f"[{sib}] {b} +{inputs.get('theta')}° -> {d}"
    if tab == "emergent_slerp":
        b = ", ".join(inputs.get("basket", [])[:2])
        return f"[{sib}] {b} +{inputs.get('theta')}° -> {len(inputs.get('modes', []))} factor modes"
    if tab == "arithmetic":
        p = " + ".join(inputs.get("positives", [])[:2])
        n = " + ".join(inputs.get("negatives", [])[:2])
        return f"[{sib}] {p}" + (f" - {n}" if n else "")
    if tab == "compare":
        return f"[3 siblings] {', '.join(inputs.get('basket', [])[:2])} +{inputs.get('theta')}°"
    return "(unknown)"

def save_query(saved, tab, inputs_dict):
    saved = list(saved or [])
    rec = {
        "id": str(uuid.uuid4()),
        "created_at": datetime.now(timezone.utc).isoformat(timespec="seconds"),
        "tab": tab,
        "inputs": inputs_dict,
        "summary": _summarise(tab, inputs_dict),
    }
    saved.insert(0, rec)
    saved = saved[:200]
    return saved, _render_saved(saved)

def delete_query(saved, qid):
    saved = [q for q in (saved or []) if q.get("id") != qid]
    return saved, _render_saved(saved)

def _render_saved(saved):
    return [[q["created_at"], TAB_LABELS.get(q["tab"], q["tab"]), q["summary"], q["id"]]
            for q in (saved or [])]

# ===== Theme + CSS =====

THEME = gr.themes.Soft(
    primary_hue=gr.themes.Color(
        c50="#E8F4F1", c100="#C8E6DE", c200=KAIKAKU_ACCENT_LIGHT,
        c300="#7BBAA9", c400="#4DA08F", c500=KAIKAKU_ACCENT,
        c600=KAIKAKU_ACCENT_HOVER, c700="#155547", c800="#0F3B33",
        c900=KAIKAKU_DARK, c950=KAIKAKU_DEEP,
    ),
    neutral_hue="slate",
    font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"],
).set(
    block_label_text_color="#1f2937",
    block_label_text_weight="600",
    block_title_text_color="#0f172a",
    block_title_text_weight="700",
    body_text_color="#0f172a",
    body_text_color_subdued="#475569",
    button_primary_background_fill=KAIKAKU_ACCENT,
    button_primary_background_fill_hover=KAIKAKU_ACCENT_HOVER,
    button_primary_text_color="#ffffff",
    button_primary_border_color=KAIKAKU_ACCENT,
    button_secondary_background_fill="#f1f5f9",
    button_secondary_background_fill_hover="#e2e8f0",
    button_secondary_text_color=KAIKAKU_DARK,
    slider_color=KAIKAKU_ACCENT,
    color_accent=KAIKAKU_ACCENT,
)

CUSTOM_CSS = f"""
.gradio-container {{max-width: 1280px !important;}}
footer {{visibility: hidden;}}
.gradio-container label, .gradio-container .label,
.gradio-container [data-testid="block-label"],
.gradio-container .block-label, .gradio-container .gr-block-label {{
  color: #0f172a !important; font-weight: 600 !important; background: transparent !important;
}}
.gradio-container button[role="tab"] {{ color: #334155 !important; font-weight: 500 !important; }}
.gradio-container button[role="tab"][aria-selected="true"] {{
  color: {KAIKAKU_ACCENT} !important; border-bottom-color: {KAIKAKU_ACCENT} !important; font-weight: 700 !important;
}}
.gradio-container button.primary, .gradio-container .primary > button {{
  background: {KAIKAKU_ACCENT} !important; color: #ffffff !important;
  border-color: {KAIKAKU_ACCENT} !important; font-weight: 600 !important;
}}
.gradio-container button.primary:hover {{ background: {KAIKAKU_ACCENT_HOVER} !important; border-color: {KAIKAKU_ACCENT_HOVER} !important; }}
.gradio-container table thead th, .gradio-container .gr-dataframe thead th {{
  color: #0f172a !important; font-weight: 700 !important; background: #f8fafc !important;
}}
.gradio-container table tbody td {{ color: #0f172a !important; }}
.sibling-card {{
  border-left: 3px solid {KAIKAKU_ACCENT}; padding: 10px 14px;
  margin: 6px 0; background: #f8fafc; border-radius: 4px;
}}
.sibling-name {{ color: {KAIKAKU_DARK}; font-weight: 700; font-size: 1.02em; }}
.sibling-desc {{ color: #334155; font-size: 0.95em; line-height: 1.5; }}
"""

_INITIAL_UMAP = umap_view("chem", ["chicken","lemon","garlic"], True, 8, three_d=False)
_INITIAL_HEATMAP = _basket_heatmap(MODELS["chem"], ["chicken","lemon","garlic"])

SIBLING_CARDS = """
<div class="sibling-card">
  <div class="sibling-name">Cooc - recipe-context only</div>
  <div class="sibling-desc">Walks recipe co-occurrence (NPMI graph) only. Neighbours are recipe <em>companions</em>: things that get cooked with the seed. Isotropic geometry (PR=173.6 of 300). Best for "what else do I cook with X".</div>
</div>
<div class="sibling-card">
  <div class="sibling-name">Core - blended (the middle ground)</div>
  <div class="sibling-desc">Typed FlavorDB compound walks blended with injected I-I walks at ii_repeat=10. Concentrated geometry (PR=94.2), tightest emergent modes. Chemistry-aware but keeps recipe context.</div>
</div>
<div class="sibling-card">
  <div class="sibling-name">Chem - chemistry only</div>
  <div class="sibling-desc">Typed FlavorDB compound metapaths only (ii_repeat=0). Neighbours are flavour-profile <em>peers</em>: things that share aroma chemistry with the seed. Best supervised-direction recovery; cuisine Cohen's d = 3.07 across 8 macro-regions.</div>
</div>
"""

# ===== Helper for ingredient picker with food-group filter =====

def _ingredient_picker(label, default_value, multiselect=True, max_choices=10):
    radio = gr.Radio(choices=FOOD_GROUP_CHOICES, value="All",
                     label=f"{label} - food group filter", interactive=True)
    dd = gr.Dropdown(choices=ALL_INGREDIENTS, value=default_value, label=label,
                     multiselect=multiselect, max_choices=max_choices)
    radio.change(_filter_dropdown, inputs=[radio, dd], outputs=dd, show_progress="hidden")
    return radio, dd

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

with gr.Blocks(title="Epicure Explorer", theme=THEME, css=CUSTOM_CSS) as demo:

    saved_state = gr.BrowserState(default_value=[], storage_key="epicure_saved_queries_v1")

    gr.Markdown(
        """# Epicure Explorer
Chef-facing operators over three sibling ingredient embeddings (Cooc / Core / Chem) from
[arXiv:2605.22391](https://arxiv.org/abs/2605.22391). 1,790 canonical ingredients across 7 languages,
300-D Metapath2Vec, controlled chemistry-vs-recipe-context spectrum."""
    )

    gr.HTML(SIBLING_CARDS)

    sibling = gr.Radio(choices=["cooc","core","chem"], value="chem", label="Sibling embedding to query")

    shared_basket = gr.State([])

    with gr.Tabs() as tabs:

        # ---------- Tab 1: Basket pairings ----------
        with gr.Tab("Basket pairings", id="tab_basket"):
            gr.Markdown("Pick one or more ingredients. The tool averages their unit vectors and returns nearest neighbours plus closest modes of that centroid.")
            basket_radio, basket = _ingredient_picker("Ingredient basket (pick 1+)", ["chicken","lemon","garlic"])
            k_pair = gr.Slider(1, 15, value=8, step=1, label="K")
            with gr.Row():
                pair_btn = gr.Button("Find pairings", variant="primary")
                save_basket_btn = gr.Button("Save this query", variant="secondary")
            with gr.Row():
                nb_table = gr.Dataframe(headers=["Neighbour","Cosine"], label="Top-K nearest neighbours", interactive=False)
                mode_table = gr.Dataframe(headers=["Mode id","Label","Kind","Cosine"], label="Closest modes", interactive=False)
            heatmap_plot = gr.Plot(value=_INITIAL_HEATMAP, label="Pairwise cosine (matplotlib)")
            pair_btn.click(basket_pairings, inputs=[sibling, basket, k_pair],
                           outputs=[nb_table, mode_table, heatmap_plot], show_progress="full")
            gr.Examples(
                examples=[
                    ["chem", ["chicken","lemon","garlic"], 8],
                    ["core", ["miso","ginger","sesame_oil"], 8],
                    ["chem", ["tomato","basil","mozzarella_cheese"], 8],
                    ["cooc", ["chocolate","strawberry","cream"], 8],
                    ["chem", ["cumin","coriander","turmeric"], 8],
                    ["core", ["soy_sauce","ginger","scallion"], 8],
                    ["chem", ["red_wine","beef","rosemary"], 8],
                    ["core", ["coconut_milk","lemongrass","fish_sauce"], 8],
                ],
                inputs=[sibling, basket, k_pair],
                label="Try one of these baskets",
            )

        # ---------- Tab 2: Supervised SLERP ----------
        with gr.Tab("Supervised SLERP", id="tab_sup"):
            gr.Markdown("Rotate the seed basket toward one or more supervised pole vectors.")
            sup_radio, sup_basket = _ingredient_picker("Seed basket (pick 1+)", ["rice"])
            sup_dirs = gr.Dropdown(choices=_supervised_choices("chem"), value=["cuisine:South_Asian"],
                                   label="Supervised directions (pick 1+; summed)",
                                   multiselect=True, max_choices=5)
            sup_theta = gr.Slider(0, 90, value=30, step=5, label="Rotation angle (deg)")
            sup_k = gr.Slider(1, 15, value=8, step=1, label="K")
            with gr.Row():
                sup_btn = gr.Button("Rotate", variant="primary")
                save_sup_btn = gr.Button("Save this query", variant="secondary")
            sup_table = gr.Dataframe(headers=["Ingredient","Cosine"], label="Top-K rotated-query neighbours")
            sup_explainer = gr.Markdown()
            sup_btn.click(supervised_slerp_multi,
                          inputs=[sibling, sup_basket, sup_dirs, sup_theta, sup_k],
                          outputs=[sup_table, sup_explainer], show_progress="full")
            sibling.change(lambda s: gr.Dropdown(choices=_supervised_choices(s), value=[]),
                           inputs=sibling, outputs=sup_dirs)
            gr.Examples(
                examples=[
                    ["chem", ["rice"], ["cuisine:South_Asian"], 30, 8],
                    ["chem", ["corn"], ["cuisine:Latin_American"], 30, 8],
                    ["core", ["chicken"], ["cuisine:Mediterranean"], 45, 8],
                    ["core", ["tomato","basil"], ["cuisine:Southeast_Asian"], 45, 8],
                    ["chem", ["beef"], ["cuisine:East_Asian"], 60, 8],
                    ["cooc", ["chocolate"], ["cuisine:Latin_American"], 30, 8],
                ],
                inputs=[sibling, sup_basket, sup_dirs, sup_theta, sup_k],
                label="Try one of these rotations",
            )

        # ---------- Tab 3: Emergent SLERP ----------
        with gr.Tab("Emergent SLERP", id="tab_em"):
            gr.Markdown("Rotate the seed basket toward one or more emergent FastICA factor-mode poles.")
            em_radio, em_basket = _ingredient_picker("Seed basket (pick 1+)", ["chocolate"])
            factor_opts = _factor_mode_choices("chem")
            em_modes = gr.Dropdown(choices=[label for label, _ in factor_opts],
                                   value=[factor_opts[0][0]] if factor_opts else [],
                                   label="Factor modes (pick 1+; summed)", multiselect=True, max_choices=5)
            em_theta = gr.Slider(0, 90, value=30, step=5, label="Rotation angle (deg)")
            em_k = gr.Slider(1, 15, value=8, step=1, label="K")
            with gr.Row():
                em_btn = gr.Button("Rotate", variant="primary")
                save_em_btn = gr.Button("Save this query", variant="secondary")
            em_table = gr.Dataframe(headers=["Ingredient","Cosine"], label="Top-K rotated-query neighbours")
            em_explainer = gr.Markdown()
            em_btn.click(emergent_slerp_multi,
                         inputs=[sibling, em_basket, em_modes, em_theta, em_k],
                         outputs=[em_table, em_explainer], show_progress="full")
            sibling.change(lambda s: gr.Dropdown(choices=[label for label, _ in _factor_mode_choices(s)], value=[]),
                           inputs=sibling, outputs=em_modes)

        # ---------- Tab 4: Arithmetic ----------
        with gr.Tab("Arithmetic", id="tab_ar"):
            gr.Markdown("Mikolov-style vector arithmetic: `centroid(positives) - centroid(negatives)`, then top-K neighbours. Killer demo: `miso - salt` on Core.")
            pos_radio, pos_box = _ingredient_picker("Positives (added)", ["miso"])
            neg_radio, neg_box = _ingredient_picker("Negatives (subtracted)", ["salt"])
            ar_k = gr.Slider(1, 15, value=8, step=1, label="K")
            with gr.Row():
                ar_btn = gr.Button("Compute", variant="primary")
                save_ar_btn = gr.Button("Save this query", variant="secondary")
            ar_table = gr.Dataframe(headers=["Ingredient","Cosine"], label="Top-K nearest to result vector")
            ar_explainer = gr.Markdown()
            ar_btn.click(arithmetic, inputs=[sibling, pos_box, neg_box, ar_k],
                         outputs=[ar_table, ar_explainer], show_progress="full")
            gr.Examples(
                examples=[
                    ["core", ["miso"], ["salt"], 8],
                    ["core", ["chicken","tofu"], ["beef"], 8],
                    ["cooc", ["basil","cumin"], ["parsley"], 8],
                    ["chem", ["chocolate"], ["sugar"], 8],
                    ["chem", ["wine"], ["beer"], 8],
                    ["core", ["bread"], ["flour"], 8],
                    ["core", ["coffee"], ["milk"], 8],
                    ["chem", ["mozzarella_cheese"], ["milk"], 8],
                ],
                inputs=[sibling, pos_box, neg_box, ar_k],
                label="Try one of these arithmetic queries",
            )

        # ---------- Tab 5: Mode atlas (click row -> UMAP) ----------
        with gr.Tab("Mode atlas", id="tab_atlas"):
            gr.Markdown(
                "Browse the GMM mode atlas. Cooc 150 / Core 193 / Chem 200 modes. "
                "**Click any row** to send that mode's members to the UMAP tab as a basket."
            )
            atlas_kind = gr.Radio(choices=["all","factor","continuous","binary"], value="all", label="Mode kind")
            atlas_search = gr.Textbox(label="Search labels / properties", placeholder="e.g. South Asian, baking, fiber", value="")
            atlas_btn = gr.Button("Browse modes", variant="primary")
            atlas_table = gr.Dataframe(
                headers=["mode_id","kind","property","label","n_members","top members"],
                label="Modes (click a row to highlight on UMAP)",
                wrap=True, interactive=False,
            )
            atlas_btn.click(browse_modes, inputs=[sibling, atlas_kind, atlas_search], outputs=atlas_table, show_progress="full")

        # ---------- Tab 6: Compare siblings ----------
        with gr.Tab("Compare siblings", id="tab_cmp"):
            gr.Markdown("Same query, three siblings, side by side.")
            cmp_radio, cmp_basket = _ingredient_picker("Seed basket", ["chicken"])
            cmp_dirs = gr.Dropdown(choices=_supervised_choices("chem"), value=[],
                                   label="Optional directions (empty = pure pairings)",
                                   multiselect=True, max_choices=5)
            cmp_theta = gr.Slider(0, 90, value=30, step=5, label="Rotation angle (deg)")
            cmp_k = gr.Slider(1, 15, value=8, step=1, label="K")
            with gr.Row():
                cmp_btn = gr.Button("Compare across siblings", variant="primary")
                save_cmp_btn = gr.Button("Save this query", variant="secondary")
            with gr.Row():
                cmp_cooc = gr.Dataframe(headers=["Cooc neighbour","Cosine"], label="Cooc (recipe-context)")
                cmp_core = gr.Dataframe(headers=["Core neighbour","Cosine"], label="Core (blended)")
                cmp_chem = gr.Dataframe(headers=["Chem neighbour","Cosine"], label="Chem (chemistry)")
            cmp_btn.click(compare_siblings, inputs=[cmp_basket, cmp_dirs, cmp_theta, cmp_k],
                          outputs=[cmp_cooc, cmp_core, cmp_chem], show_progress="full")

        # ---------- Tab 7: UMAP ----------
        with gr.Tab("UMAP visualisation", id="tab_umap"):
            gr.Markdown(
                "2-D UMAP of the 1,790-ingredient embedding (cosine, n_neighbors=30, min_dist=0.03). "
                "Points coloured by food group. Basket members appear as accent stars; top-K neighbours as amber dots."
            )
            umap_radio, umap_basket = _ingredient_picker("Highlight these ingredients", ["chicken","lemon","garlic"])
            with gr.Row():
                umap_show_nb = gr.Checkbox(value=True, label="Show top-K neighbours of basket centroid")
                umap_3d = gr.Checkbox(value=False, label="3-D perspective (UMAP + PC1)")
                umap_k = gr.Slider(1, 20, value=10, step=1, label="K neighbours")
            umap_btn = gr.Button("Update plot", variant="primary")
            umap_plot = gr.Plot(value=_INITIAL_UMAP, label="UMAP")
            umap_btn.click(umap_view, inputs=[sibling, umap_basket, umap_show_nb, umap_k, umap_3d],
                           outputs=umap_plot, show_progress="full")
            sibling.change(umap_view, inputs=[sibling, umap_basket, umap_show_nb, umap_k, umap_3d],
                           outputs=umap_plot)

        # ---------- Tab 8: Parse my fridge ----------
        with gr.Tab("Parse my fridge", id="tab_fridge"):
            gr.Markdown(
                "Paste a free-text ingredient list. Quantities, units, and prep notes are stripped, "
                "then each line is fuzzy-matched to canonical vocab. Click **Send matched to Basket tab** "
                "to populate the Basket Pairings input."
            )
            fridge_text = gr.Textbox(
                label="Free-text ingredients (one per line or semicolon-separated)",
                lines=8,
                value=("2 boneless chicken thighs\n1 cup coconut milk\n1 tbsp fish sauce (or soy sauce)\n"
                       "fresh lemongrass, bruised\n3 cloves garlic, minced\n1 inch fresh ginger\n"
                       "juice of one lime\nsalt to taste"),
            )
            fridge_min = gr.Slider(40, 100, value=70, step=5, label="Min match score (rapidfuzz)")
            with gr.Row():
                fridge_btn = gr.Button("Parse and match", variant="primary")
                fridge_send = gr.Button("Send matched to Basket tab", variant="secondary")
            fridge_table = gr.Dataframe(
                headers=["Input line", "Canonical match", "Score", "Cleaned"],
                label="Parsed matches", interactive=False,
            )
            fridge_matched = gr.Textbox(label="Matched ingredients", interactive=False)
            def _parse(txt, sib, mn):
                rows, matches = parse_fridge(txt, sib, int(mn))
                return rows, ", ".join(matches), matches
            fridge_btn.click(_parse, inputs=[fridge_text, sibling, fridge_min],
                             outputs=[fridge_table, fridge_matched, shared_basket], show_progress="full")
            fridge_send.click(lambda matches: gr.Dropdown(value=matches[:10] if matches else []),
                              inputs=[shared_basket], outputs=[basket])

        # ---------- Tab 9: Recipe builder ----------
        with gr.Tab("Recipe builder", id="tab_recipe"):
            gr.Markdown(
                "Describe a dish in plain English. Hybrid retrieval: rapidfuzz token matching for direct "
                "ingredient mentions + sentence-transformer cosine against the sibling's factor-mode labels "
                "for thematic matches. First call after Space cold-start downloads ~80MB encoder (one-time)."
            )
            rb_prompt = gr.Textbox(label="Dish description", lines=3,
                                   value="I'm making Thai green curry for 4 people")
            rb_k = gr.Slider(4, 20, value=10, step=1, label="Suggestions (K)")
            rb_btn = gr.Button("Suggest starter basket", variant="primary")
            rb_table = gr.Dataframe(
                headers=["Ingredient", "Source", "Score"],
                label="Suggested basket (source = direct / thematic / both)", interactive=False,
            )
            rb_explainer = gr.Markdown()
            rb_matched = gr.State([])
            rb_send = gr.Button("Send to Basket tab", variant="secondary")
            def _rb(prompt, sib, k):
                rows, names, md = suggest_basket(prompt, sib, int(k))
                return rows, md, names
            rb_btn.click(_rb, inputs=[rb_prompt, sibling, rb_k],
                         outputs=[rb_table, rb_explainer, rb_matched], show_progress="full")
            rb_send.click(lambda names: gr.Dropdown(value=(names or [])[:10]),
                          inputs=[rb_matched], outputs=[basket])
            gr.Examples(
                examples=[
                    ["I'm making Thai green curry for 4 people", 10],
                    ["spicy vegetarian taco filling", 10],
                    ["weeknight pasta with tomatoes and herbs", 10],
                    ["Japanese miso-glazed salmon and greens", 10],
                    ["Moroccan tagine with lamb and dried fruit", 10],
                ],
                inputs=[rb_prompt, rb_k],
                label="Try one of these prompts",
            )

        # ---------- Tab 10: Saved queries ----------
        with gr.Tab("Saved queries", id="tab_saved"):
            gr.Markdown(
                "Stored locally in your browser via `localStorage` (gr.BrowserState). "
                "~5 MB quota; per-browser, not per-account. Clearing browser data wipes them."
            )
            saved_table = gr.Dataframe(
                headers=["created_at", "tab", "summary", "id"],
                label="Your saved queries (newest first)", interactive=False, wrap=True,
            )
            with gr.Row():
                selected_id = gr.State("")
                del_btn = gr.Button("Delete selected", variant="secondary")
            def _on_select(saved, evt: gr.SelectData):
                if evt is None or evt.index is None:
                    return ""
                row = evt.index[0] if isinstance(evt.index, (list, tuple)) else evt.index
                return (saved or [{}])[row].get("id", "") if row < len(saved or []) else ""
            saved_table.select(_on_select, inputs=[saved_state], outputs=selected_id)
            del_btn.click(delete_query, inputs=[saved_state, selected_id],
                          outputs=[saved_state, saved_table])
            demo.load(lambda s: _render_saved(s), inputs=saved_state, outputs=saved_table)

    # ---- Wire Save buttons (after all tabs exist so all components are in scope) ----
    save_basket_btn.click(
        lambda s, sib, b, k: save_query(s, "basket",
            {"sibling": sib, "basket": b or [], "k": int(k)}),
        inputs=[saved_state, sibling, basket, k_pair],
        outputs=[saved_state, saved_table],
    )
    save_sup_btn.click(
        lambda s, sib, b, d, th, k: save_query(s, "supervised_slerp",
            {"sibling": sib, "basket": b or [], "directions": d or [], "theta": float(th), "k": int(k)}),
        inputs=[saved_state, sibling, sup_basket, sup_dirs, sup_theta, sup_k],
        outputs=[saved_state, saved_table],
    )
    save_em_btn.click(
        lambda s, sib, b, m, th, k: save_query(s, "emergent_slerp",
            {"sibling": sib, "basket": b or [], "modes": m or [], "theta": float(th), "k": int(k)}),
        inputs=[saved_state, sibling, em_basket, em_modes, em_theta, em_k],
        outputs=[saved_state, saved_table],
    )
    save_ar_btn.click(
        lambda s, sib, p, n, k: save_query(s, "arithmetic",
            {"sibling": sib, "positives": p or [], "negatives": n or [], "k": int(k)}),
        inputs=[saved_state, sibling, pos_box, neg_box, ar_k],
        outputs=[saved_state, saved_table],
    )
    save_cmp_btn.click(
        lambda s, b, d, th, k: save_query(s, "compare",
            {"basket": b or [], "directions": d or [], "theta": float(th), "k": int(k)}),
        inputs=[saved_state, cmp_basket, cmp_dirs, cmp_theta, cmp_k],
        outputs=[saved_state, saved_table],
    )

    # ---- Mode atlas row click -> UMAP highlight + jump to UMAP tab ----
    def atlas_row_to_umap(sibling_value, table_value, show_nb, k_value, three_d_value, evt: gr.SelectData):
        if evt is None or evt.index is None or table_value is None:
            return gr.update(), gr.update(), gr.update(), gr.update()
        row = evt.index[0] if isinstance(evt.index, (list, tuple)) else evt.index
        try:
            clicked_mode_id = (table_value.iloc[row, 0] if hasattr(table_value, "iloc")
                               else table_value[row][0])
        except Exception:
            return gr.update(), gr.update(), gr.update(), gr.update()
        m = MODELS[sibling_value]
        mode = next((md for md in m.modes if md.mode_id == clicked_mode_id), None)
        if mode is None:
            return gr.update(), gr.update(), gr.update(), gr.update()
        members = [n for n in mode.members if n in m.vocab][:10]
        if not members:
            return gr.update(), gr.update(), gr.update(), gr.update()
        fig = umap_view(sibling_value, members, bool(show_nb), int(k_value), three_d=bool(three_d_value))
        return (
            gr.Dropdown(value=members),
            fig,
            members,
            gr.Tabs(selected="tab_umap"),
        )
    atlas_table.select(
        atlas_row_to_umap,
        inputs=[sibling, atlas_table, umap_show_nb, umap_k, umap_3d],
        outputs=[umap_basket, umap_plot, shared_basket, tabs],
        show_progress="hidden",
    )

    # ---- Public API endpoints ----
    gr.api(api_neighbors,         api_name="neighbors")
    gr.api(api_slerp,             api_name="slerp")
    gr.api(api_arithmetic,        api_name="arithmetic")
    gr.api(api_embed,             api_name="embed")
    gr.api(api_list_directions,   api_name="list_directions")
    gr.api(api_list_factor_modes, api_name="list_factor_modes")

    gr.Markdown(
        """---
### Developer API

These operators are also exposed as JSON endpoints. See `/?view=api` for the auto-generated schema.

```python
from gradio_client import Client
c = Client("Kaikaku/epicure-explorer")
c.predict("garlic", "chem", 5, api_name="/neighbors")
c.predict("rice", "cuisine:South_Asian", 30, "chem", 5, api_name="/slerp")
c.predict(["miso"], ["salt"], "core", 8, api_name="/arithmetic")
c.predict("garlic", "chem", api_name="/embed")          # 300-D L2-normalised vector
c.predict("chem", api_name="/list_directions")
c.predict("chem", api_name="/list_factor_modes")
```

Endpoints validate inputs and return `{"error": "...", "suggestions": [...]}` on bad input. Free-tier limits: ~1-2 req/sec shared, no auth, Space sleeps after ~48h idle (cold start ~30-60s on next request).

---
**Cite:** Radzikowski and Chen, 2026, *Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings*, [arXiv:2605.22391](https://arxiv.org/abs/2605.22391). Artefacts: [epicure-cooc](https://huggingface.co/Kaikaku/epicure-cooc) | [epicure-core](https://huggingface.co/Kaikaku/epicure-core) | [epicure-chem](https://huggingface.co/Kaikaku/epicure-chem) | [corpus dataset](https://huggingface.co/datasets/Kaikaku/epicure-corpus-resources)
"""
    )

if __name__ == "__main__":
    demo.launch()