app.py

import html
from datetime import datetime, timezone

import gradio as gr
import numpy as np
import pandas as pd
from sentence_transformers import SentenceTransformer


DATA_PATH = "users_ai_ml_interests_only.parquet"
EMBEDDINGS_PATH = "embeddings_ai_ml_interests.npy"
MODEL_NAME = "BAAI/bge-small-en-v1.5"


profile_df = pd.read_parquet(DATA_PATH)
profile_embeddings = np.load(EMBEDDINGS_PATH).astype(np.float32)

print(f"✅ Loaded {len(profile_df)} HF Atlas profiles from parquet")
print(f"✅ Loaded embeddings: {profile_embeddings.shape}")

if len(profile_df) != profile_embeddings.shape[0]:
    raise ValueError(
        f"Parquet / embeddings mismatch: {len(profile_df)} rows vs {profile_embeddings.shape[0]} embeddings"
    )


def detect_username_col(df):
    for col in ["user", "username", "namespace"]:
        if col in df.columns:
            return col
    return None


USERNAME_COL = detect_username_col(profile_df)

if USERNAME_COL is None:
    raise ValueError("No username column found. Expected one of: user, username, namespace")


def normalize_embeddings_if_needed(embeddings):
    norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
    norms = np.where(norms == 0, 1.0, norms)
    return embeddings / norms


profile_embeddings = normalize_embeddings_if_needed(profile_embeddings)

embedder = SentenceTransformer(MODEL_NAME)


def safe_text(value, default=""):
    if value is None:
        return default

    try:
        if pd.isna(value):
            return default
    except Exception:
        pass

    text = str(value).strip()

    if text.lower() in {"nan", "none", "null"}:
        return default

    return text


def parse_last_seen(value):
    text = safe_text(value)

    if not text:
        return pd.NaT

    return pd.to_datetime(text, errors="coerce", utc=True)


def prepare_dates():
    if "last_seen_all_repo" not in profile_df.columns:
        profile_df["_last_seen_dt"] = pd.NaT
        print("⚠️ Column last_seen_all_repo not found. Date filter will only work as no-filter.")
        return

    profile_df["_last_seen_dt"] = profile_df["last_seen_all_repo"].map(parse_last_seen)

    known = int(profile_df["_last_seen_dt"].notna().sum())
    unknown = int(profile_df["_last_seen_dt"].isna().sum())

    print(f"🕒 Known last_seen_all_repo dates: {known}")
    print(f"🕳️ Unknown last_seen_all_repo dates: {unknown}")

    if known > 0:
        print(f"🕒 Min last_seen: {profile_df['_last_seen_dt'].min()}")
        print(f"🕒 Max last_seen: {profile_df['_last_seen_dt'].max()}")


prepare_dates()


def filter_by_activity(df, activity_filter, custom_days):
    if "_last_seen_dt" not in df.columns:
        return df

    custom_days = int(custom_days) if custom_days else 0

    if activity_filter == "No filter":
        return df

    if activity_filter == "Has known activity date":
        return df[df["_last_seen_dt"].notna()]

    if activity_filter == "No known activity date":
        return df[df["_last_seen_dt"].isna()]

    if activity_filter == "Max age in days":
        if custom_days <= 0:
            return df

        now = pd.Timestamp(datetime.now(timezone.utc))
        cutoff = now - pd.Timedelta(days=custom_days)

        return df[df["_last_seen_dt"].notna() & (df["_last_seen_dt"] >= cutoff)]

    return df


def format_number(value):
    text = safe_text(value, "0")

    try:
        number = int(float(text))
        return f"{number:,}".replace(",", " ")
    except Exception:
        return html.escape(text)


def format_date(value):
    text = safe_text(value)

    if not text:
        return "unknown"

    try:
        dt = pd.to_datetime(text, errors="coerce", utc=True)
        if pd.isna(dt):
            return "unknown"
        return dt.strftime("%Y-%m-%d")
    except Exception:
        return html.escape(text)


def truncate(text, max_len=900):
    text = safe_text(text)

    if len(text) <= max_len:
        return text

    return text[:max_len].rsplit(" ", 1)[0] + "..."


def get_profile_url(row):
    if "atlas_request_url" in row and safe_text(row["atlas_request_url"]):
        return (
            safe_text(row["atlas_request_url"])
            .replace("/api/users/", "/")
            .replace("/overview", "")
        )

    username = safe_text(row[USERNAME_COL])
    return f"https://huggingface.co/{username}"


def render_profile_card(row, score, rank):
    username = safe_text(row[USERNAME_COL], "unknown")
    fullname = safe_text(row.get("fullname", ""), "")
    details = safe_text(row.get("details", ""), "")
    ai_ml_interests = truncate(row.get("ai_ml_interests", ""), 800)
    last_seen = format_date(row.get("last_seen_all_repo", ""))

    num_models = format_number(row.get("numModels", row.get("n_models", 0)))
    num_datasets = format_number(row.get("numDatasets", row.get("n_datasets", 0)))
    num_spaces = format_number(row.get("numSpaces", row.get("n_spaces", 0)))
    followers = format_number(row.get("numFollowers", 0))
    likes = format_number(row.get("numLikes", row.get("numUpvotes", 0)))

    url = get_profile_url(row)

    title = html.escape(username)
    fullname_html = html.escape(fullname) if fullname else "—"
    details_html = html.escape(truncate(details, 300)) if details else ""
    interests_html = html.escape(ai_ml_interests).replace("\n", "<br>")

    extra_details = ""
    if details_html:
        extra_details = f"""
        <div class="details">{details_html}</div>
        """

    return f"""
    <div class="result-card">
        <div class="result-topline">
            <div>
                <div class="rank">#{rank}</div>
                <a class="username" href="{url}" target="_blank">{title}</a>
                <div class="fullname">{fullname_html}</div>
            </div>
            <div class="score">{score * 100:.2f}%</div>
        </div>

        {extra_details}

        <div class="interests">
            <div class="label">AI/ML interests</div>
            <div>{interests_html}</div>
        </div>

        <div class="stats">
            <span>🧠 Models: <b>{num_models}</b></span>
            <span>📚 Datasets: <b>{num_datasets}</b></span>
            <span>🚀 Spaces: <b>{num_spaces}</b></span>
            <span>❤️ Likes: <b>{likes}</b></span>
            <span>👥 Followers: <b>{followers}</b></span>
            <span>🕒 Last seen: <b>{last_seen}</b></span>
        </div>
    </div>
    """


def build_search_results(query, activity_filter, custom_days, display_count):
    query = safe_text(query)

    if not query:
        return """
        <div class="empty-state">
            Describe an AI/ML topic, research area, tool, model family, or technical interest.
        </div>
        """, 0, False

    custom_days = int(custom_days) if custom_days else 0
    display_count = int(display_count)

    eligible = filter_by_activity(profile_df, activity_filter, custom_days)

    print("FILTER:", activity_filter, "DAYS:", custom_days, "ELIGIBLE:", len(eligible))

    if "_last_seen_dt" in eligible.columns and len(eligible) > 0:
        known_eligible = eligible[eligible["_last_seen_dt"].notna()]
        if len(known_eligible) > 0:
            print("ELIGIBLE MIN LAST_SEEN:", known_eligible["_last_seen_dt"].min())
            print("ELIGIBLE MAX LAST_SEEN:", known_eligible["_last_seen_dt"].max())

    if len(eligible) == 0:
        return """
        <div class="empty-state">
            No profile found for this activity filter.
        </div>
        """, display_count, False

    eligible_indices = eligible.index.to_numpy()
    eligible_embeddings = profile_embeddings[eligible_indices]

    query_emb = embedder.encode(
        [query],
        convert_to_numpy=True,
        normalize_embeddings=True,
    ).astype(np.float32)

    similarities = np.dot(query_emb, eligible_embeddings.T)[0]

    display_count = max(1, min(display_count, len(eligible_indices)))
    best_local_indices = np.argsort(-similarities)[:display_count]

    cards = []

    if activity_filter == "Max age in days" and custom_days > 0:
        filter_label = f"active in last {custom_days} days"
    elif activity_filter == "Has known activity date":
        filter_label = "with known public activity date"
    elif activity_filter == "No known activity date":
        filter_label = "with no known public activity date"
    else:
        filter_label = "no date filter"

    header = f"""
    <div class="search-summary">
        <b>{len(eligible):,}</b> eligible profiles · showing top <b>{display_count}</b> · <b>{filter_label}</b>
    </div>
    """.replace(",", " ")

    cards.append(header)

    for rank, local_idx in enumerate(best_local_indices, start=1):
        global_idx = eligible_indices[local_idx]
        row = profile_df.iloc[global_idx]
        score = float(similarities[local_idx])
        cards.append(render_profile_card(row, score, rank))

    has_more = display_count < len(eligible_indices)

    if not has_more:
        cards.append("""
        <div class="empty-state">
            All eligible profiles are already displayed.
        </div>
        """)

    return "\n".join(cards), display_count, has_more


def search_hf_atlas(query, activity_filter, custom_days):
    results_html, display_count, has_more = build_search_results(
        query=query,
        activity_filter=activity_filter,
        custom_days=custom_days,
        display_count=10,
    )

    more_button_update = gr.update(visible=has_more)

    return results_html, display_count, more_button_update


def search_more_hf_atlas(query, activity_filter, custom_days, display_count):
    if display_count is None:
        display_count = 10

    display_count = int(display_count)

    if display_count <= 0:
        display_count = 10

    new_display_count = display_count + 10

    results_html, final_display_count, has_more = build_search_results(
        query=query,
        activity_filter=activity_filter,
        custom_days=custom_days,
        display_count=new_display_count,
    )

    more_button_update = gr.update(visible=has_more)

    return results_html, final_display_count, more_button_update


css = """
body {
    background: radial-gradient(circle at top left, #172554 0%, #020617 35%, #020617 100%);
    font-family: Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, sans-serif;
}

.gradio-container {
    max-width: 980px !important;
}

#title {
    font-size: 3.1em;
    font-weight: 900;
    text-align: center;
    color: #e0f2fe;
    text-shadow: 0 0 18px rgba(56, 189, 248, 0.45);
    margin-bottom: 0;
}

#subtitle {
    color: #bae6fd;
    text-align: center;
    margin-top: 0.6em;
    margin-bottom: 2.2em;
    font-size: 1.15em;
}

textarea {
    background: rgba(15, 23, 42, 0.92) !important;
    border: 1px solid rgba(125, 211, 252, 0.55) !important;
    color: #e0f2fe !important;
    border-radius: 18px !important;
}

input, select {
    background: rgba(15, 23, 42, 0.90) !important;
    border: 1px solid rgba(56, 189, 248, 0.35) !important;
    color: #e0f2fe !important;
}

button {
    background: linear-gradient(135deg, #38bdf8, #818cf8) !important;
    border: none !important;
    color: #020617 !important;
    font-weight: 900 !important;
    font-size: 1.08em !important;
    border-radius: 18px !important;
    box-shadow: 0 0 24px rgba(56, 189, 248, 0.35);
}

button:hover {
    transform: scale(1.015);
    box-shadow: 0 0 34px rgba(129, 140, 248, 0.55);
}

.result-card {
    background: linear-gradient(135deg, rgba(15, 23, 42, 0.96), rgba(30, 41, 59, 0.88));
    border: 1px solid rgba(125, 211, 252, 0.35);
    border-radius: 24px;
    padding: 22px;
    margin: 18px 0;
    box-shadow: 0 16px 44px rgba(0, 0, 0, 0.34);
}

.result-topline {
    display: flex;
    justify-content: space-between;
    gap: 16px;
    align-items: flex-start;
}

.rank {
    color: #7dd3fc;
    font-size: 0.92em;
    font-weight: 800;
    letter-spacing: 0.08em;
}

.username {
    color: #e0f2fe !important;
    font-size: 1.55em;
    font-weight: 900;
    text-decoration: none !important;
}

.username:hover {
    color: #38bdf8 !important;
    text-decoration: underline !important;
}

.fullname {
    color: #cbd5e1;
    margin-top: 4px;
    font-size: 0.98em;
}

.score {
    color: #020617;
    background: linear-gradient(135deg, #67e8f9, #a5b4fc);
    padding: 9px 13px;
    border-radius: 999px;
    font-weight: 900;
    min-width: 90px;
    text-align: center;
}

.details {
    color: #cbd5e1;
    background: rgba(2, 6, 23, 0.38);
    border-left: 3px solid rgba(56, 189, 248, 0.65);
    padding: 12px 14px;
    margin-top: 16px;
    border-radius: 14px;
}

.interests {
    margin-top: 16px;
    color: #e0f2fe;
    line-height: 1.55;
}

.label {
    color: #7dd3fc;
    font-weight: 900;
    margin-bottom: 6px;
    text-transform: uppercase;
    letter-spacing: 0.08em;
    font-size: 0.78em;
}

.stats {
    display: flex;
    flex-wrap: wrap;
    gap: 10px;
    margin-top: 18px;
}

.stats span {
    background: rgba(14, 165, 233, 0.10);
    color: #bae6fd;
    border: 1px solid rgba(125, 211, 252, 0.22);
    border-radius: 999px;
    padding: 7px 11px;
    font-size: 0.92em;
}

.search-summary {
    color: #bae6fd;
    text-align: center;
    background: rgba(15, 23, 42, 0.7);
    border: 1px solid rgba(125, 211, 252, 0.25);
    border-radius: 18px;
    padding: 12px;
    margin-bottom: 18px;
}

.empty-state {
    text-align: center;
    color: #bae6fd;
    background: rgba(15, 23, 42, 0.75);
    border: 1px solid rgba(125, 211, 252, 0.25);
    border-radius: 18px;
    padding: 24px;
    margin-top: 16px;
}

.more-button-wrap {
    margin-top: 10px;
    margin-bottom: 30px;
}

.nebula {
    position: fixed;
    inset: 0;
    pointer-events: none;
    z-index: 0;
    opacity: 0.40;
    background:
        radial-gradient(circle at 20% 20%, rgba(56,189,248,0.24), transparent 28%),
        radial-gradient(circle at 80% 30%, rgba(129,140,248,0.22), transparent 30%),
        radial-gradient(circle at 50% 80%, rgba(14,165,233,0.16), transparent 26%);
}
"""


with gr.Blocks(title="HF Atlas Explorer") as demo:
    gr.HTML('<div class="nebula"></div>')
    gr.HTML('<h1 id="title"> 🌐 HF Collab Center </h1>')
    gr.HTML(
        '<p id="subtitle">Search Hugging Face profiles by AI/ML interests and filter by public activity.</p>'
    )

    query = gr.Textbox(
        label="Search query",
        placeholder="e.g. diffusion models, biomedical NLP, reinforcement learning, graph neural networks, robotics...",
        lines=4,
    )

    with gr.Row():
        activity_filter = gr.Dropdown(
            choices=[
                "No filter",
                "Max age in days",
                "Has known activity date",
                "No known activity date",
            ],
            value="Max age in days",
            label="Last public activity filter",
        )

        custom_days = gr.Number(
            label="Max last_seen age in days, 0 = no limit",
            value=365,
            precision=0,
        )

    display_count_state = gr.State(value=0)

    submit_btn = gr.Button("🔎 Search HF Atlas")

    output = gr.HTML()

    with gr.Row(elem_classes=["more-button-wrap"]):
        more_btn = gr.Button("➕ Show more", visible=False)

    submit_btn.click(
        fn=search_hf_atlas,
        inputs=[query, activity_filter, custom_days],
        outputs=[output, display_count_state, more_btn],
    )

    more_btn.click(
        fn=search_more_hf_atlas,
        inputs=[query, activity_filter, custom_days, display_count_state],
        outputs=[output, display_count_state, more_btn],
    )


demo.launch(css=css, theme=gr.themes.Base())