Update app.py

app.py

@@ -1,14 +1,15 @@
 # app.py
-# 🎁 GIfty — Smart Gift Recommender (Embeddings + FAISS)
+# 🎁 GIfty — Smart Gift Recommender (Embeddings + FAISS + LLM generator)
 # Dataset: ckandemir/amazon-products (Hugging Face)
 # UI: Gradio (English)
 #
-# Works on common Spaces stacks (no RangeSlider; two sliders for budget)
-# Chosen model: sentence-transformers/all-MiniLM-L6-v2 (fast, strong baseline)
-#
-# Tip: First query builds embeddings+FAISS (cached in-memory).
+# Notes:
+# - Embeddings: sentence-transformers/all-MiniLM-L6-v2 + FAISS IndexFlatIP
+# - LLM generator: google/flan-t5-small (local transformers, no API keys)
+# - Budget uses two sliders (compatible with older Gradio).
+# - Images are rendered as right-side thumbnails per result.
 
-import os, re, random
+import os, re, json, random
 from typing import Dict, List, Tuple
 
 import numpy as np
@@ -18,8 +19,11 @@ from datasets import load_dataset
 from sentence_transformers import SentenceTransformer
 import faiss
 
+# LLM (Flan-T5) for generation
+from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
+
 # ---------------- Config ----------------
-MAX_ROWS = int(os.getenv("MAX_ROWS", "6000"))  # cap to keep build time reasonable on CPU
+MAX_ROWS = int(os.getenv("MAX_ROWS", "6000"))  # keep index build fast on CPU
 TITLE = "# 🎁 GIfty — Smart Gift Recommender\n*Top-3 similar picks + 1 generated idea + personalized message*"
 
 OCCASION_OPTIONS = [
@@ -35,6 +39,8 @@ AGE_OPTIONS = {
     "senior (65+)": "senior",
 }
 
+GENDER_OPTIONS = ["any", "female", "male", "nonbinary"]
+
 INTEREST_OPTIONS = [
     "reading","writing","tech","travel","fitness","cooking","tea","coffee",
     "games","movies","plants","music","design","stationery","home","experience",
@@ -42,7 +48,7 @@ INTEREST_OPTIONS = [
     "photography","outdoors","pets","beauty","jewelry"
 ]
 
-# Query-expansion dictionary (improves semantic match with catalog wording)
+# Query-expansion dictionary (helps matching catalog wording)
 SYNONYMS = {
     "music": ["audio", "headphones", "vinyl", "earbuds", "speaker"],
     "tech": ["electronics", "gadgets", "computer", "smart", "device"],
@@ -140,7 +146,7 @@ def load_catalog() -> pd.DataFrame:
         ds = load_dataset("ckandemir/amazon-products", split="train")
         raw = ds.to_pandas()
     except Exception:
-        # Fallback so the app never crashes if internet is blocked
+        # Fallback (keeps the app alive if internet is blocked)
         raw = pd.DataFrame({
             "Product Name": ["Wireless Earbuds", "Coffee Sampler", "Strategy Board Game"],
             "Description": [
@@ -166,6 +172,19 @@ def _contains_ci(series: pd.Series, needle: str) -> pd.Series:
     pat = re.escape(needle)
     return series.fillna("").str.contains(pat, case=False, regex=True)
 
+def gender_tokens(gender: str) -> List[str]:
+    gender = (gender or "any").lower()
+    if gender == "female": return ["women", "woman", "female", "her"]
+    if gender == "male": return ["men", "man", "male", "him"]
+    if gender == "nonbinary": return ["unisex", "gender neutral", "they"]
+    return ["unisex"]  # "any"
+
+def soft_gender_boost(row: pd.Series, gender: str) -> float:
+    if not gender or gender == "any": return 0.0
+    tokens = gender_tokens(gender)
+    blob = f"{row.get('tags','')} {row.get('short_desc','')}".lower()
+    return 0.08 if any(t in blob for t in tokens) else 0.0
+
 def filter_business(df: pd.DataFrame, budget_min=None, budget_max=None,
                     occasion: str=None, age_range: str="any") -> pd.DataFrame:
     m = pd.Series(True, index=df.index)
@@ -179,7 +198,7 @@ def filter_business(df: pd.DataFrame, budget_min=None, budget_max=None,
         m &= (df["age_range"].fillna("any").isin([age_range, "any"]))
     return df[m]
 
-# ---------------- Embeddings + FAISS ----------------
+# ---------------- Embeddings + FAISS (MiniLM) ----------------
 class EmbeddingIndex:
     def __init__(self, docs: List[str], model_id: str):
         self.model_id = model_id
@@ -194,8 +213,7 @@ class EmbeddingIndex:
         sims, idxs = self.index.search(qv, topn)
         return sims[0], idxs[0]
 
-# Choose the best all-around model for this app:
-EMBED_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2"  # fast & good quality
+EMBED_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2"  # fast & solid on CPU
 EMB_INDEX = EmbeddingIndex(CATALOG["doc"].tolist(), EMBED_MODEL_ID)
 
 # ---------------- Query building ----------------
@@ -209,28 +227,28 @@ def expand_with_synonyms(tokens: List[str]) -> List[str]:
     return out
 
 def profile_to_query(profile: Dict) -> str:
-    """
-    Weighted, doc-aligned query (interests + synonyms) + occasion + age.
-    Repeats interests to give them more weight.
-    """
+    # Weighted, doc-aligned query (interests + synonyms) + occasion + age + gender signal
     interests = [t.strip().lower() for t in profile.get("interests", []) if t.strip()]
     expanded = expand_with_synonyms(interests)
     expanded = expanded + expanded  # weight x2
     occasion = (profile.get("occasion", "") or "").lower()
     age = profile.get("age_range", "any")
+    gender = (profile.get("gender", "any") or "any").lower()
     parts = []
     if expanded: parts.append(", ".join(expanded))
     if occasion: parts.append(occasion)
     if age and age != "any": parts.append(age)
+    if gender and gender != "any":
+        parts.append("women" if gender=="female" else ("men" if gender=="male" else "unisex"))
     return " | ".join(parts).strip()
 
 def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
     query = profile_to_query(profile)
 
-    # Global search on full catalog
-    sims, idxs = EMB_INDEX.search(query, topn=min(max(k*50, k), len(CATALOG)))
+    # Global search
+    sims, idxs = EMB_INDEX.search(query, topn=min(max(k*80, k), len(CATALOG)))
 
-    # Filter down to business subset
+    # Business subset
     df_f = filter_business(
         CATALOG,
         budget_min=profile.get("budget_min"),
@@ -241,107 +259,180 @@ def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
     if df_f.empty:
         df_f = CATALOG
 
-    order = np.argsort(-sims)
+    # Gather candidates within the subset and apply a small gender boost
+    cand = []
+    for i, sim in zip(idxs, sims):
+        i = int(i)
+        if i in df_f.index:
+            boost = soft_gender_boost(CATALOG.loc[i], profile.get("gender","any"))
+            cand.append((i, float(sim) + boost))
+    cand.sort(key=lambda x: -x[1])
+
+    # Pick unique by name
     seen, picks = set(), []
-    for gi in idxs[order]:
-        gi = int(gi)
-        if gi not in df_f.index:
-            continue
+    for gi, score in cand:
         nm = CATALOG.loc[gi, "name"]
-        if nm in seen:
-            continue
+        if nm in seen: continue
         seen.add(nm)
-        picks.append(gi)
-        if len(picks) >= k:
-            break
+        picks.append((gi, score))
+        if len(picks) >= k: break
 
     if not picks:
         res = df_f.head(k).copy()
         res["similarity"] = np.nan
         return res[["name","short_desc","price_usd","occasion_tags","persona_fit","age_range","image_url","similarity"]]
 
-    gi_to_sim = {int(i): float(s) for i, s in zip(idxs, sims)}
-    res = CATALOG.loc[picks].copy()
-    res["similarity"] = [gi_to_sim.get(int(i), np.nan) for i in picks]
+    sel = [gi for gi,_ in picks]
+    res = CATALOG.loc[sel].copy()
+    sim_map = dict(picks)
+    res["similarity"] = [sim_map.get(int(gi), np.nan) for gi in sel]
     return res[["name","short_desc","price_usd","occasion_tags","persona_fit","age_range","image_url","similarity"]]
 
-# ---------------- Generative item + message ----------------
-def generate_item(profile: Dict) -> Dict:
-    random.seed(42)  # stable demo
-    interests = profile.get("interests", [])
-    occasion  = profile.get("occasion","birthday")
-    budget    = profile.get("budget_max", profile.get("budget_usd", 50)) or 50
-    age       = profile.get("age_range","any")
-    core = (interests[0] if interests else "hobby").strip() or "hobby"
-    style = random.choice(["personalized","experience","bundle"])
-    if style == "personalized":
-        base_name = f"Custom {core} accessory with initials"
-        base_desc = f"Thoughtful personalized {core} accessory tailored to their taste."
-    elif style == "experience":
-        base_name = f"{core.title()} workshop voucher"
-        base_desc = f"A guided intro session to explore {core} in a fun, hands-on way."
-    else:
-        base_name = f"{core.title()} starter bundle"
-        base_desc = f"A curated set to kickstart their {core} passion."
-    if age == "kids":
-        base_desc += " Suitable for kids with safe, age-appropriate materials."
-    elif age == "teens":
-        base_desc += " Trendy pick that suits young enthusiasts."
-    elif age == "senior":
-        base_desc += " Comfortable and easy to use."
-    price = float(np.clip(float(budget), 10, 300))
+# ---------------- LLM generator (Flan-T5) ----------------
+LLM_ID = "google/flan-t5-small"
+try:
+    _tok = AutoTokenizer.from_pretrained(LLM_ID)
+    _mdl = AutoModelForSeq2SeqLM.from_pretrained(LLM_ID)
+    LLM = pipeline("text2text-generation", model=_mdl, tokenizer=_tok)
+except Exception as e:
+    LLM = None
+    print("LLM load failed, will fallback to rule-based. Error:", e)
+
+def _run_llm(prompt: str, max_new_tokens=128) -> str:
+    if LLM is None:
+        return ""
+    out = LLM(prompt, max_new_tokens=max_new_tokens, do_sample=False, temperature=0.0)
+    return out[0]["generated_text"]
+
+def _parse_json_maybe(s: str) -> dict:
+    try:
+        return json.loads(s)
+    except Exception:
+        # try to extract {...}
+        m = re.search(r"\{.*\}", s, flags=re.S)
+        if m:
+            try:
+                return json.loads(m.group(0))
+            except Exception:
+                return {}
+        return {}
+
+def llm_generate_item(profile: Dict) -> Dict:
+    # Prompt to produce a single gift idea JSON
+    prompt = f"""
+You are GIfty, a gift recommender. Create ONE gift idea as JSON with keys:
+name, short_desc, price_usd, occasion_tags, persona_fit.
+Constraints:
+- Fit the recipient profile.
+- price_usd must be a number within the budget range.
+- Keep text concise, friendly, and realistic.
+
+Recipient profile:
+interests = {profile.get('interests', [])}
+occasion = {profile.get('occasion','birthday')}
+age_group = {profile.get('age_range','any')}
+gender = {profile.get('gender','any')}
+budget_min = {profile.get('budget_min', 10)}
+budget_max = {profile.get('budget_max', 100)}
+
+Return ONLY JSON.
+"""
+    txt = _run_llm(prompt, max_new_tokens=160)
+    data = _parse_json_maybe(txt)
+    if not data:
+        # fallback rule-based if LLM unavailable or malformed
+        core = (profile.get("interests",[ "hobby" ])[0] or "hobby").strip()
+        name = f"Custom {core} accessory with initials ({profile.get('occasion','birthday')})"
+        return {
+            "name": name,
+            "short_desc": f"Thoughtful personalized {core} accessory tailored to their taste.",
+            "price_usd": float(np.clip(profile.get("budget_max", 50) or 50, 10, 300)),
+            "occasion_tags": profile.get("occasion","birthday"),
+            "persona_fit": ", ".join(profile.get("interests", [])) or "general",
+            "age_range": profile.get("age_range","any"),
+            "image_url": ""
+        }
+    # ensure numeric price and bounds
+    try:
+        p = float(data.get("price_usd", profile.get("budget_max", 50)))
+    except Exception:
+        p = float(profile.get("budget_max", 50) or 50)
+    p = float(np.clip(p, profile.get("budget_min", 10) or 10, profile.get("budget_max", 300) or 300))
     return {
-        "name": f"{base_name} ({occasion})",
-        "short_desc": base_desc,
-        "price_usd": price,
-        "occasion_tags": occasion,
-        "persona_fit": ", ".join(interests) or "general",
-        "age_range": age,
+        "name": data.get("name","Gift idea"),
+        "short_desc": data.get("short_desc","A thoughtful personalized idea."),
+        "price_usd": p,
+        "occasion_tags": data.get("occasion_tags", profile.get("occasion","birthday")),
+        "persona_fit": data.get("persona_fit", ", ".join(profile.get("interests", [])) or "general"),
+        "age_range": profile.get("age_range","any"),
         "image_url": ""
     }
 
-def generate_message(profile: Dict) -> str:
-    name = profile.get("recipient_name","Friend")
-    occasion = profile.get("occasion","birthday")
-    tone = profile.get("tone","warm and friendly")
-    return (f"Dear {name},\n"
-            f"Happy {occasion}! Wishing you health, joy, and wonderful memories. "
-            f"May your goals come true. With {tone}.")
-
-# ---------------- Rendering helpers ----------------
+def llm_generate_message(profile: Dict) -> str:
+    prompt = f"""
+Write a short, warm greeting message (2–3 sentences) in English for a gift card.
+Recipient name: {profile.get('recipient_name','Friend')}
+Occasion: {profile.get('occasion','birthday')}
+Interests: {', '.join(profile.get('interests', []))}
+Age group: {profile.get('age_range','any')}
+Gender: {profile.get('gender','any')}
+Tone: {profile.get('tone','warm and friendly')}
+Avoid emojis. Keep it sincere and concise.
+"""
+    txt = _run_llm(prompt, max_new_tokens=90)
+    if not txt:
+        # fallback
+        return (f"Dear {profile.get('recipient_name','Friend')},\n"
+                f"Happy {profile.get('occasion','birthday')}! Wishing you health, joy, and wonderful memories. "
+                f"With {profile.get('tone','warm and friendly')}.")
+    return txt.strip()
+
+# ---------------- Rendering helpers (HTML with right-side thumbnail) ----------------
 def md_escape(text: str) -> str:
     return str(text).replace("|","\\|").replace("*","\\*").replace("_","\\_")
 
-def render_top3_md(df: pd.DataFrame) -> str:
+def render_top3_html(df: pd.DataFrame) -> str:
     if df is None or df.empty:
-        return "_No results found._"
-    lines = ["**Top-3 recommendations:**\n"]
+        return "<em>No results found.</em>"
+    # Simple cards with image on the right
+    rows = []
     for _, r in df.iterrows():
         name = md_escape(r.get("name",""))
         desc = md_escape(r.get("short_desc",""))
         price = r.get("price_usd")
         sim = r.get("similarity")
         age = r.get("age_range","any")
-        img = r.get("image_url","")
-        if img:
-            lines.append(f"![ ]({img})")
+        img = r.get("image_url","") or ""
         price_str = f"${price:.0f}" if pd.notna(price) else "N/A"
         sim_str = f"{sim:.3f}" if pd.notna(sim) else "—"
-        lines.append(f"**{name}**  \n{desc}  \nPrice: **{price_str}** · Age: `{age}` · Similarity: `{sim_str}`\n")
-    return "\n".join(lines)
+        img_html = f'<img src="{img}" alt="" style="width:84px;height:84px;object-fit:cover;border-radius:10px;margin-left:12px;" />' if img else ""
+        card = f"""
+<div style="display:flex;align-items:flex-start;justify-content:space-between;gap:10px;padding:10px;border:1px solid #eee;border-radius:12px;margin-bottom:8px;">
+  <div style="flex:1;min-width:0;">
+    <div style="font-weight:700;">{name}</div>
+    <div style="font-size:0.95em;margin-top:4px;">{desc}</div>
+    <div style="font-size:0.9em;margin-top:6px;opacity:0.8;">
+      Price: <b>{price_str}</b> · Age: <code>{age}</code> · Similarity: <code>{sim_str}</code>
+    </div>
+  </div>
+  {img_html}
+</div>
+"""
+        rows.append(card)
+    return "\n".join(rows)
 
 # ---------------- Gradio UI ----------------
 EXAMPLES = [
-    [["tech","music"], "birthday", 20, 60, "Noa", "adult (18–64)", "warm and friendly"],
-    [["home","cooking","practical"], "housewarming", 25, 45, "Daniel", "adult (18–64)", "warm"],
-    [["games","photography"], "birthday", 30, 120, "Omer", "teen (13–17)", "fun"],
-    [["reading","design","aesthetic"], "thank_you", 15, 35, "Maya", "any", "friendly"],
+    [["tech","music"], "birthday", 20, 60, "Noa", "adult (18–64)", "any", "warm and friendly"],
+    [["home","cooking","practical"], "housewarming", 25, 45, "Daniel", "adult (18–64)", "male", "warm"],
+    [["games","photography"], "birthday", 30, 120, "Omer", "teen (13–17)", "male", "fun"],
+    [["reading","design","aesthetic"], "thank_you", 15, 35, "Maya", "any", "female", "friendly"],
 ]
 
 def ui_predict(interests_list: List[str], occasion: str, budget_min: float, budget_max: float,
-               recipient_name: str, age_label: str, tone: str):
+               recipient_name: str, age_label: str, gender: str, tone: str):
     try:
-        # budget sanity
+        # sanity
         if budget_min is None: budget_min = 20.0
         if budget_max is None: budget_max = 60.0
         if budget_min > budget_max:
@@ -356,18 +447,19 @@ def ui_predict(interests_list: List[str], occasion: str, budget_min: float, budg
             "budget_max": float(budget_max),
             "budget_usd": float(budget_max),
             "age_range": age_range,
+            "gender": gender or "any",
             "tone": tone or "warm and friendly",
         }
 
         top3 = recommend_topk(profile, k=3)
-        gen  = generate_item(profile)
-        msg  = generate_message(profile)
+        gen  = llm_generate_item(profile)
+        msg  = llm_generate_message(profile)
 
-        top3_md = render_top3_md(top3)
+        top3_html = render_top3_html(top3)
         gen_md  = f"**{md_escape(gen['name'])}**\n\n{md_escape(gen['short_desc'])}\n\n~${gen['price_usd']:.0f}"
-        return top3_md, gen_md, msg
+        return top3_html, gen_md, msg
     except Exception as e:
-        return f":warning: Error: {e}", "", ""
+        return f"<div style='color:#b00;'>⚠️ Error: {e}</div>", "", ""
 
 with gr.Blocks() as demo:
     gr.Markdown(TITLE)
@@ -382,8 +474,9 @@ with gr.Blocks() as demo:
     with gr.Row():
         occasion = gr.Dropdown(label="Occasion", choices=OCCASION_OPTIONS, value="birthday")
         age = gr.Dropdown(label="Age group", choices=list(AGE_OPTIONS.keys()), value="adult (18–64)")
+        gender = gr.Dropdown(label="Recipient gender", choices=GENDER_OPTIONS, value="any")
 
-    # Two sliders (compatible with older Gradio)
+    # Two budget sliders (compatible with older Gradio)
     with gr.Row():
         budget_min = gr.Slider(label="Min budget (USD)", minimum=5, maximum=500, step=1, value=20)
         budget_max = gr.Slider(label="Max budget (USD)", minimum=5, maximum=500, step=1, value=60)
@@ -394,19 +487,19 @@ with gr.Blocks() as demo:
 
     go = gr.Button("Get GIfty 🎯")
 
-    out_top3 = gr.Markdown(label="Top-3 recommendations")
+    out_top3 = gr.HTML(label="Top-3 recommendations")   # HTML to support right-side thumbnails
     out_gen  = gr.Markdown(label="Generated item")
     out_msg  = gr.Markdown(label="Personalized message")
 
     gr.Examples(
         EXAMPLES,
-        [interests, occasion, budget_min, budget_max, recipient_name, age, tone],
+        [interests, occasion, budget_min, budget_max, recipient_name, age, gender, tone],
         label="Quick examples",
     )
 
     go.click(
         ui_predict,
-        [interests, occasion, budget_min, budget_max, recipient_name, age, tone],
+        [interests, occasion, budget_min, budget_max, recipient_name, age, gender, tone],
         [out_top3, out_gen, out_msg]
     )