Update app.py

app.py

@@ -1,11 +1,13 @@
-# app.py
-# 🎁 GIfty — Smart Gift Recommender (Embeddings + FAISS + LLM + Image Gen)
+# app.py — Gifty (revised)
+# 🎁 GIfty — Smart Gift Recommender
 # Data: ckandemir/amazon-products
-# Retrieval: MiniLM embeddings + FAISS (cosine)
-# Generation: Flan-T5-small (text), SD-Turbo (image)
-# UI: Gradio; Quick Examples on top; Budget range: RangeSlider if present, else two sliders
+# Retrieval: MiniLM-L12-v2 embeddings + FAISS (cosine), with simple on-disk cache
+# DIY Generation: small instruct LMs via HF pipeline (default: flan-t5-small) with JSON validate+repair (no padding)
+# Greeting: short LLM completion
+# Image: SD-Turbo (optional)
+# UI: Gradio; Quick Examples; Budget RangeSlider; DIY JSON + readable card
 
-import os, re, json, random
+import os, re, json, random, hashlib, pathlib
 from typing import Dict, List, Tuple
 
 import numpy as np
@@ -16,23 +18,37 @@ from datasets import load_dataset
 from sentence_transformers import SentenceTransformer
 import faiss
 
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForCausalLM, pipeline
 
 import torch
 from diffusers import AutoPipelineForText2Image
 
 # --------------------- Config ---------------------
 MAX_ROWS = int(os.getenv("MAX_ROWS", "8000"))
-TITLE = "# 🎁 GIfty — Smart Gift Recommender\n*Top-3 similar picks + 1 invented gift (with image) + personalized message*"
+TITLE = "# 🎁 GIfty — Smart Gift Recommender\n*Top-3 catalog picks + 1 DIY gift (JSON) + personalized message*"
 
-# ===== Updated Interests (exact) =====
+# Retrieval model (embedding)
+EMBED_MODEL_ID = os.getenv("EMBED_MODEL_ID", "sentence-transformers/all-MiniLM-L12-v2")
+EMBED_CACHE_DIR = os.getenv("EMBED_CACHE_DIR", "./.gifty_cache")
+pathlib.Path(EMBED_CACHE_DIR).mkdir(parents=True, exist_ok=True)
+
+# DIY generation model (text)
+GEN_MODEL_ID = os.getenv("GEN_MODEL_ID", "google/flan-t5-small")
+OUTPUT_LANG = os.getenv("OUTPUT_LANG", "en")  # "en" or "he"
+MAX_INPUT_TOKENS = int(os.getenv("MAX_INPUT_TOKENS", "360"))
+MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "260"))
+DIY_MAX_ATTEMPTS = int(os.getenv("DIY_MAX_ATTEMPTS", "4"))
+
+# Image gen toggle
+ENABLE_IMAGE = os.getenv("ENABLE_IMAGE", "1") == "1"
+
+# ===== UI options =====
 INTEREST_OPTIONS = [
     "Sports","Travel","Cooking","Technology","Music","Art","Reading","Gardening","Fashion",
     "Gaming","Photography","Hiking","Movies","Crafts","Pets","Wellness","Collecting","Food",
     "Home decor","Science"
 ]
 
-# ===== Updated Occasions (exact) =====
 OCCASION_UI = [
     "Birthday","Wedding / Engagement","Anniversary","Graduation","New baby","Housewarming",
     "Retirement","Holidays","Valentine’s Day","Promotion / New job","Get well soon"
@@ -52,7 +68,6 @@ OCCASION_CANON = {
     "Get well soon":"get_well"
 }
 
-# ===== Updated Relationship & Tone =====
 RECIPIENT_RELATIONSHIPS = [
     "Family - Parent",
     "Family - Sibling",
@@ -68,15 +83,7 @@ RECIPIENT_RELATIONSHIPS = [
 ]
 
 MESSAGE_TONES = [
-    "Formal",
-    "Casual",
-    "Funny",
-    "Heartfelt",
-    "Inspirational",
-    "Playful",
-    "Romantic",
-    "Appreciative",
-    "Encouraging",
+    "Formal","Casual","Funny","Heartfelt","Inspirational","Playful","Romantic","Appreciative","Encouraging",
 ]
 
 AGE_OPTIONS = {
@@ -211,9 +218,13 @@ def load_catalog() -> pd.DataFrame:
             ],
             "Category": ["Electronics | Audio","Grocery | Coffee","Toys & Games | Board Games"],
             "Selling Price": ["$59.00","$34.00","$39.00"],
-            "Image": ["","",""],
+            "Image": ["","",""]
         })
     df = map_amazon_to_schema(raw).drop_duplicates(subset=["name","short_desc"])
+    # EDA cleanups: drop missing price, cap to <= 500
+    df = df[pd.notna(df["price_usd"])].copy()
+    df = df[df["price_usd"] <= 500].reset_index(drop=True)
+    # limit rows
     if len(df) > MAX_ROWS:
         df = df.sample(n=MAX_ROWS, random_state=42).reset_index(drop=True)
     df["doc"] = df.apply(build_doc, axis=1)
@@ -221,38 +232,43 @@ def load_catalog() -> pd.DataFrame:
 
 CATALOG = load_catalog()
 
-# --------------------- Business filters ---------------------
-def _contains_ci(series: pd.Series, needle: str) -> pd.Series:
-    if not needle: return pd.Series(True, index=series.index)
-    return series.fillna("").str.contains(re.escape(needle), case=False, regex=True)
-
-def filter_business(df: pd.DataFrame, budget_min=None, budget_max=None,
-                    occasion_canon: str=None, age_range: str="any") -> pd.DataFrame:
-    m = pd.Series(True, index=df.index)
-    if budget_min is not None:
-        m &= df["price_usd"].fillna(0) >= float(budget_min)
-    if budget_max is not None:
-        m &= df["price_usd"].fillna(1e9) <= float(budget_max)
-    if occasion_canon:
-        m &= _contains_ci(df["occasion_tags"], occasion_canon)
-    if age_range and age_range != "any":
-        m &= (df["age_range"].fillna("any").isin([age_range, "any"]))
-    return df[m]
-
-# --------------------- Embeddings + FAISS ---------------------
+# --------------------- Embeddings + FAISS (with simple cache) ---------------------
 class EmbeddingIndex:
     def __init__(self, docs: List[str], model_id: str):
+        self.model_id = model_id
         self.model = SentenceTransformer(model_id)
-        embs = self.model.encode(docs, convert_to_numpy=True, normalize_embeddings=True)
-        self.index = faiss.IndexFlatIP(embs.shape[1])  # cosine via normalized vectors
-        self.index.add(embs)
+        self.embs = self._load_or_build(docs)
+        self.index = faiss.IndexFlatIP(self.embs.shape[1])  # cosine via normalized vectors
+        self.index.add(self.embs)
+
+    def _cache_paths(self, n_docs: int) -> Tuple[str, str]:
+        h = hashlib.md5((self.model_id + f"|{n_docs}").encode()).hexdigest()[:10]
+        npy = os.path.join(EMBED_CACHE_DIR, f"emb_{h}.npy")
+        idx = os.path.join(EMBED_CACHE_DIR, f"faiss_{h}.index")
+        return npy, idx
+
+    def _load_or_build(self, docs: List[str]) -> np.ndarray:
+        npy_path, _ = self._cache_paths(len(docs))
+        if os.path.exists(npy_path):
+            try:
+                embs = np.load(npy_path)
+                if embs.shape[0] == len(docs):
+                    return embs
+            except Exception:
+                pass
+        # build
+        embs = self.model.encode(docs, convert_to_numpy=True, normalize_embeddings=True, show_progress_bar=True)
+        try:
+            np.save(npy_path, embs)
+        except Exception:
+            pass
+        return embs
 
     def search(self, query: str, topn: int):
         qv = self.model.encode([query], convert_to_numpy=True, normalize_embeddings=True)
         sims, idxs = self.index.search(qv, topn)
         return sims[0], idxs[0]
 
-EMBED_MODEL_ID = "sentence-transformers/all-MiniLM-L6-v2"  # fast & solid on CPU
 EMB_INDEX = EmbeddingIndex(CATALOG["doc"].tolist(), EMBED_MODEL_ID)
 
 # --------------------- Query building ---------------------
@@ -281,9 +297,26 @@ def profile_to_query(profile: Dict) -> str:
     if g != "any": parts.append("women" if g=="female" else ("men" if g=="male" else "unisex"))
     return " | ".join(parts)
 
+def _contains_ci(series: pd.Series, needle: str) -> pd.Series:
+    if not needle: return pd.Series(True, index=series.index)
+    return series.fillna("").str.contains(re.escape(needle), case=False, regex=True)
+
+def filter_business(df: pd.DataFrame, budget_min=None, budget_max=None,
+                    occasion_canon: str=None, age_range: str="any") -> pd.DataFrame:
+    m = pd.Series(True, index=df.index)
+    if budget_min is not None:
+        m &= df["price_usd"].fillna(0) >= float(budget_min)
+    if budget_max is not None:
+        m &= df["price_usd"].fillna(1e9) <= float(budget_max)
+    if occasion_canon:
+        m &= _contains_ci(df["occasion_tags"], occasion_canon)
+    if age_range and age_range != "any":
+        m &= (df["age_range"].fillna("any").isin([age_range, "any"]))
+    return df[m]
+
 def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
     query = profile_to_query(profile)
-    sims, idxs = EMBB_INDEX.search(query, topn=min(max(k*80, k), len(CATALOG))) if False else EMB_INDEX.search(query, topn=min(max(k*80, k), len(CATALOG)))
+    sims, idxs = EMB_INDEX.search(query, topn=min(max(k*80, k), len(CATALOG)))
     df_f = filter_business(
         CATALOG,
         budget_min=profile.get("budget_min"),
@@ -292,6 +325,7 @@ def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
         age_range=profile.get("age_range","any"),
     )
     if df_f.empty: df_f = CATALOG
+    df_f_idx = set(df_f.index.tolist())
 
     # soft gender boost
     def gender_tokens(g: str) -> List[str]:
@@ -305,7 +339,7 @@ def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
     cand = []
     for i, sim in zip(idxs, sims):
         i = int(i)
-        if i in df_f.index:
+        if i in df_f_idx:
             blob = f"{CATALOG.loc[i,'tags']} {CATALOG.loc[i,'short_desc']}".lower()
             boost = 0.08 if any(t in blob for t in gts) else 0.0
             cand.append((i, float(sim) + boost))
@@ -329,95 +363,192 @@ def recommend_topk(profile: Dict, k: int=3) -> pd.DataFrame:
     res["similarity"] = [dict(picks).get(int(i), np.nan) for i in sel]
     return res[["name","short_desc","price_usd","occasion_tags","persona_fit","age_range","image_url","similarity"]]
 
-# --------------------- LLM (text) ---------------------
-LLM_ID = "google/flan-t5-small"
+# --------------------- LLM plumbing (DIY + Greeting) ---------------------
+
+def load_text_pipeline(model_id: str):
+    trust=True
+    if "flan" in model_id or "t5" in model_id:
+        tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=trust)
+        mdl = AutoModelForSeq2SeqLM.from_pretrained(model_id, trust_remote_code=trust)
+        return pipeline("text2text-generation", model=mdl, tokenizer=tok, device_map="auto", trust_remote_code=trust)
+    else:
+        tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=trust)
+        mdl = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=trust)
+        return pipeline("text-generation", model=mdl, tokenizer=tok, device_map="auto", trust_remote_code=trust)
+
 try:
-    _tok = AutoTokenizer.from_pretrained(LLM_ID)
-    _mdl = AutoModelForSeq2SeqLM.from_pretrained(LLM_ID)
-    LLM = pipeline("text2text-generation", model=_mdl, tokenizer=_tok)
+    DIY_PIPE = load_text_pipeline(GEN_MODEL_ID)
 except Exception as e:
-    LLM = None
-    print("LLM load failed, fallback to rule-based. Error:", e)
+    DIY_PIPE = None
+    print("DIY LLM load failed:", e)
 
-def _run_llm(prompt: str, max_new_tokens=160) -> 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"]
+# Small greeting model (can reuse DIY_PIPE)
+GREETING_PIPE = DIY_PIPE
 
-def _parse_json_maybe(s: str) -> dict:
-    try:
-        return json.loads(s)
-    except Exception:
-        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 = f"""
-You are GIfty. Invent ONE gift that matches the catalog style with keys:
-name, short_desc, price_usd, occasion_tags, persona_fit. Use JSON only.
-Constraints:
-- Fit the recipient profile and relationship.
-- price_usd must be numeric within the budget range.
-Profile:
-name={profile.get('recipient_name','Friend')}
-relationship={profile.get('relationship','Friend')}
-gender={profile.get('gender','any')}
-age_group={profile.get('age_range','any')}
-interests={profile.get('interests',[])}
-occasion={profile.get('occ_ui','Birthday')}
-budget_min={profile.get('budget_min',10)}
-budget_max={profile.get('budget_max',100)}
-"""
-    txt = _run_llm(prompt, max_new_tokens=180)
-    data = _parse_json_maybe(txt)
-    if not data:
-        core = (profile.get("interests",["hobby"])[0] or "hobby").lower()
-        return {
-            "name": f"{core.title()} starter bundle ({profile.get('occ_ui','Birthday')})",
-            "short_desc": f"A curated set to kickstart their {core} passion.",
-            "price_usd": float(np.clip(profile.get("budget_max", 50) or 50, 10, 300)),
-            "occasion_tags": OCCASION_CANON.get(profile.get("occ_ui","Birthday"), "birthday"),
-            "persona_fit": ", ".join(profile.get("interests", [])) or "general",
-            "age_range": profile.get("age_range","any"),
-            "image_url": ""
-        }
+# ---- JSON helpers ----
+GENERIC_NAMES = {"diy gift","gift","personalized gift","handmade gift","custom gift","מתנה","מתנה אישית","עשה זאת בעצמך"}
+
+def _f(x, fb=0.0):
+    try: return float(x)
+    except: return float(fb)
+
+def try_parse_json(text: str):
+    if not text: return None
+    m = re.search(r"(\{[\s\S]*\})", text.strip())
+    if not m: return None
+    blob = m.group(1)
     try:
-        p = float(data.get("price_usd", profile.get("budget_max", 50)))
+        return json.loads(blob)
     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": data.get("name","Gift Idea"),
-        "short_desc": data.get("short_desc","A thoughtful idea."),
-        "price_usd": p,
-        "occasion_tags": data.get("occasion_tags", OCCASION_CANON.get(profile.get("occ_ui","Birthday"), "birthday")),
-        "persona_fit": data.get("persona_fit", ", ".join(profile.get("interests", [])) or "general"),
-        "age_range": profile.get("age_range","any"),
-        "image_url": ""
-    }
+        blob = re.sub(r",\s*}\s*$", "}", blob)
+        blob = re.sub(r",\s*\]", "]", blob)
+        try: return json.loads(blob)
+        except: return None
+
+def truncate_prompt(pipe, text: str, max_tokens: int) -> str:
+    tok = pipe.tokenizer
+    ids = tok(text, truncation=True, max_length=max_tokens, return_tensors=None).get("input_ids", [])
+    return tok.decode(ids, skip_special_tokens=True) if ids else text
+
+# ---- DIY prompt, validate & repair (no padding) ----
+
+def diy_prompt(profile: Dict) -> str:
+    lang = "English" if OUTPUT_LANG == "en" else "Hebrew"
+    name = profile.get("recipient_name","Recipient")
+    rel  = profile.get("relationship","Friend")
+    age  = profile.get("age_range","any")
+    gen  = profile.get("gender","any")
+    ints = ", ".join(profile.get("interests",[])) or "general"
+    occ  = profile.get("occ_ui","Birthday")
+    lo, hi = int(profile.get("budget_min",10)), int(profile.get("budget_max",100))
+
+    return "\n".join([
+        f"Invent ONE original DIY gift idea from scratch for this recipient. Write all VALUES in {lang}.",
+        "Return JSON ONLY with exactly these keys (and nothing else):",
+        "gift_name, overview, materials_needed, step_by_step_instructions, estimated_cost_usd, estimated_time_minutes",
+        "",
+        "Hard requirements:",
+        "- Strongly reflect the recipient's interests and the occasion.",
+        "- overview MUST mention the recipient by NAME and include relationship, age_group, gender, and the occasion.",
+        "- gift_name must be SPECIFIC (not generic), 4–10 words, include at least one interest keyword.",
+        f"- estimated_cost_usd between ${lo}-${hi}; estimated_time_minutes 20–240.",
+        "- materials_needed: at least 5 concise items with quantities.",
+        "- step_by_step_instructions: at least 6 practical, ordered steps.",
+        "Forbidden gift_name terms: DIY Gift, Gift, Personalized Gift, Handmade Gift, Custom Gift.",
+        "",
+        f"Recipient: name={name}; relationship={rel}; age_group={age}; gender={gen}.",
+        f"Interests: {ints}. Occasion: {occ}.",
+        "JSON:"
+    ])
 
-def llm_generate_message(profile: Dict) -> str:
-    prompt = f"""
+def diy_validate(g: dict, profile: Dict) -> Tuple[bool, List[str]]:
+    errs=[]
+    # keys
+    req=["gift_name","overview","materials_needed","step_by_step_instructions","estimated_cost_usd","estimated_time_minutes"]
+    for k in req:
+        if k not in g: errs.append(f"missing key: {k}")
+    # name
+    n=str(g.get("gift_name",""))
+    if not n.strip(): errs.append("gift_name empty")
+    if any(b in n.strip().lower() for b in GENERIC_NAMES): errs.append("gift_name generic")
+    if len(n.split())<3: errs.append("gift_name too short")
+    # overview mentions
+    ov=str(g.get("overview",""))
+    if profile.get("recipient_name","") and profile.get("recipient_name") not in ov: errs.append("overview missing recipient name")
+    for field,label in [("relationship","relationship"),("age_range","age_group"),("gender","gender"),("occ_ui","occasion")]:
+        val=str(profile.get(field,""))
+        if val and (val.split()[0] not in ov): errs.append(f"overview missing {label}")
+    # lists
+    mats=g.get("materials_needed", [])
+    steps=g.get("step_by_step_instructions", [])
+    if not isinstance(mats, list) or len(mats)<5: errs.append("materials_needed len < 5")
+    if not isinstance(steps, list) or len(steps)<6: errs.append("steps len < 6")
+    # numbers
+    lo, hi = _f(profile.get("budget_min",10),10), _f(profile.get("budget_max",100),100)
+    cost=_f(g.get("estimated_cost_usd"), -1)
+    if not (lo <= cost <= hi): errs.append(f"cost not in budget [{lo},{hi}]")
+    mins=int(_f(g.get("estimated_time_minutes"), -1))
+    if not (20 <= mins <= 240): errs.append("time not in 20..240")
+    return (len(errs)==0), errs
+
+def diy_repair_prompt(profile: Dict, last: dict, errors: List[str]) -> str:
+    lang = "English" if OUTPUT_LANG == "en" else "Hebrew"
+    return "\n".join([
+        f"Fix ONLY the following problems in this JSON. Keep the same idea and style. Return JSON ONLY. Write all VALUES in {lang}.",
+        "Errors:",
+        *[f"- {e}" for e in errors],
+        "JSON to fix:",
+        json.dumps(last, ensure_ascii=False)
+    ])
+
+def diy_generate(profile: Dict) -> Tuple[dict, str]:
+    if DIY_PIPE is None:
+        return {}, "DIY model not loaded"
+    # attempt 1: creative
+    prompt = diy_prompt(profile)
+    pr = truncate_prompt(DIY_PIPE, prompt, MAX_INPUT_TOKENS)
+    out = DIY_PIPE(pr, do_sample=True, temperature=0.9, top_p=0.95, max_new_tokens=MAX_NEW_TOKENS, truncation=True)
+    if not isinstance(out, list): out=[out]
+    texts = [o.get("generated_text","") for o in out]
+    candidates = [try_parse_json(t) or {} for t in texts]
+
+    # pick first valid
+    for cand in candidates:
+        ok, errs = diy_validate(cand, profile)
+        if ok:
+            return cand, "ok"
+        last = cand
+
+    # repair loop (deterministic)
+    attempts = 1
+    while attempts < DIY_MAX_ATTEMPTS:
+        ok, errs = diy_validate(last, profile)
+        if ok:
+            return last, "ok"
+        fix_pr = diy_repair_prompt(profile, last, errs)
+        fix_pr = truncate_prompt(DIY_PIPE, fix_pr, MAX_INPUT_TOKENS)
+        fixed = DIY_PIPE(fix_pr, do_sample=False, max_new_tokens=MAX_NEW_TOKENS, truncation=True)
+        fixed = (fixed if isinstance(fixed, list) else [fixed])[0].get("generated_text","")
+        fixed = try_parse_json(fixed) or last
+        last = fixed
+        attempts += 1
+    return last, "partial"
+
+# ---- Greeting generation ----
+
+def greeting_prompt(profile: Dict) -> str:
+    tone = profile.get('tone','Heartfelt')
+    name = profile.get('recipient_name','Friend')
+    rel  = profile.get('relationship','Friend')
+    occ  = profile.get('occ_ui','Birthday')
+    ints = ", ".join(profile.get('interests', []))
+    age  = profile.get('age_range','any')
+    gen  = profile.get('gender','any')
+    return f"""
 Write a short greeting (2–3 sentences) in English for a gift card.
-Tone: {profile.get('tone','Heartfelt')}
-Use the relationship to set warmth/formality.
-Recipient: {profile.get('recipient_name','Friend')} ({profile.get('relationship','Friend')})
-Occasion: {profile.get('occ_ui','Birthday')}
-Interests: {', '.join(profile.get('interests', []))}
-Age group: {profile.get('age_range','any')}; Gender: {profile.get('gender','any')}
+Tone: {tone}
+Recipient: {name} ({rel})
+Occasion: {occ}
+Interests: {ints}
+Age group: {age}; Gender: {gen}
 Avoid emojis.
 """
-    txt = _run_llm(prompt, max_new_tokens=90)
-    if not txt:
-        return (f"Dear {profile.get('recipient_name','Friend')}, "
-                f"happy {profile.get('occ_ui','Birthday').lower()}! Wishing you joy and wonderful memories.")
-    return txt.strip()
+
+def llm_generate_message(profile: Dict) -> str:
+    if GREETING_PIPE is None:
+        return (f"Dear {profile.get('recipient_name','Friend')}, happy {profile.get('occ_ui','Birthday').lower()}! "
+                f"Wishing you joy and wonderful memories.")
+    pr = truncate_prompt(GREETING_PIPE, greeting_prompt(profile), MAX_INPUT_TOKENS)
+    out = GREETING_PIPE(pr, do_sample=False, max_new_tokens=90, truncation=True)
+    out = out if isinstance(out, list) else [out]
+    txt = out[0].get("generated_text","")
+    return txt.strip() or (f"Dear {profile.get('recipient_name','Friend')}, happy {profile.get('occ_ui','Birthday').lower()}!")
 
 # --------------------- Image generation (SD-Turbo) ---------------------
+
 def load_image_pipeline():
+    if not ENABLE_IMAGE:
+        return None
     try:
         device = "cuda" if torch.cuda.is_available() else "cpu"
         dtype = torch.float16 if torch.cuda.is_available() else torch.float32
@@ -430,12 +561,14 @@ def load_image_pipeline():
 
 IMG_PIPE = load_image_pipeline()
 
-def generate_gift_image(gift: Dict):
-    if IMG_PIPE is None:
+
+def generate_gift_image_from_diy(diy: Dict):
+    if IMG_PIPE is None or not diy:
         return None
+    name = diy.get('gift_name','gift')
+    ov = diy.get('overview','product photo of handmade gift')
     prompt = (
-        f"{gift.get('name','gift')}, {gift.get('short_desc','')}. "
-        f"Style: product photo, soft studio lighting, minimal background, realistic, high detail."
+        f"{name}: {ov}. Style: product photo, soft studio lighting, minimal background, realistic, high detail."
     )
     try:
         img = IMG_PIPE(
@@ -450,6 +583,7 @@ def generate_gift_image(gift: Dict):
         return None
 
 # --------------------- Rendering ---------------------
+
 def md_escape(text: str) -> str:
     return str(text).replace("|","\\|").replace("*","\\*").replace("_","\\_")
 
@@ -482,6 +616,32 @@ def render_top3_html(df: pd.DataFrame) -> str:
         rows.append(card)
     return "\n".join(rows)
 
+
+def render_diy_md(d: Dict) -> str:
+    if not d:
+        return "<em>DIY generation failed.</em>"
+    name = md_escape(d.get("gift_name",""))
+    ov   = md_escape(d.get("overview",""))
+    cost = d.get("estimated_cost_usd", "—")
+    mins = d.get("estimated_time_minutes", "—")
+    mats = d.get("materials_needed", [])
+    steps= d.get("step_by_step_instructions", [])
+    mats_md = "\n".join([f"- {md_escape(str(m))}" for m in mats]) if isinstance(mats, list) else "- —"
+    steps_md= "\n".join([f"{i+1}. {md_escape(str(s))}" for i,s in enumerate(steps)]) if isinstance(steps, list) else "1. —"
+    return f"""
+### DIY Gift — {name}
+
+{ov}
+
+**Estimated cost:** ${cost} · **Estimated time:** {mins} min
+
+**Materials needed:**
+{mats_md}
+
+**Step-by-step:**
+{steps_md}
+"""
+
 # --------------------- Gradio UI ---------------------
 CSS = """
 #examples { order: 1; }
@@ -491,62 +651,61 @@ CSS = """
 with gr.Blocks(css=CSS) as demo:
     gr.Markdown(TITLE)
 
-    # top section (examples placeholder)
     with gr.Column(elem_id="examples"):
         gr.Markdown("### Quick examples")
 
     with gr.Column(elem_id="form"):
         with gr.Row():
-            recipient_name = gr.Textbox(label="Recipient name", value="Noa")
-            relationship = gr.Dropdown(label="Relationship", choices=RECIPIENT_RELATIONSHIPS, value="Friend")
+            recipient_name = gr.Textbox(label="Recipient name", value="Rotem")
+            relationship = gr.Dropdown(label="Relationship", choices=RECIPIENT_RELATIONSHIPS, value="Romantic partner")
 
         with gr.Row():
             interests = gr.CheckboxGroup(
                 label="Interests (select a few)", choices=INTEREST_OPTIONS,
-                value=["Technology","Music"], interactive=True
+                value=["Reading","Fashion","Home decor"], interactive=True
             )
 
         with gr.Row():
-            occasion = gr.Dropdown(label="Occasion", choices=OCCASION_UI, value="Birthday")
+            occasion = gr.Dropdown(label="Occasion", choices=OCCASION_UI, value="Valentine’s Day")
             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")
+            gender = gr.Dropdown(label="Recipient gender", choices=GENDER_OPTIONS, value="female")
 
-        # Budget: try RangeSlider else two sliders
         RangeSlider = getattr(gr, "RangeSlider", None)
         if RangeSlider is not None:
-            budget_range = RangeSlider(label="Budget range (USD)", minimum=5, maximum=500, step=1, value=[20, 60])
+            budget_range = RangeSlider(label="Budget range (USD)", minimum=5, maximum=500, step=1, value=[30, 60])
             budget_min, budget_max = None, None
         else:
             with gr.Row():
-                budget_min = gr.Slider(label="Min budget (USD)", minimum=5, maximum=500, step=1, value=20)
+                budget_min = gr.Slider(label="Min budget (USD)", minimum=5, maximum=500, step=1, value=30)
                 budget_max = gr.Slider(label="Max budget (USD)", minimum=5, maximum=500, step=1, value=60)
             budget_range = gr.State(value=None)
 
-        tone = gr.Dropdown(label="Message tone", choices=MESSAGE_TONES, value="Heartfelt")
+        tone = gr.Dropdown(label="Message tone", choices=MESSAGE_TONES, value="Romantic")
 
         go = gr.Button("Get GIfty 🎯")
 
         out_top3 = gr.HTML(label="Top-3 recommendations")
-        out_gen_text  = gr.Markdown(label="Invented gift")
-        out_gen_img   = gr.Image(label="Invented gift image", type="pil")
-        out_msg  = gr.Markdown(label="Personalized message")
+        out_diy_json = gr.JSON(label="DIY Gift (JSON)")
+        out_diy_md   = gr.Markdown(label="DIY Gift (readable)")
+        out_gen_img  = gr.Image(label="DIY Gift image", type="pil")
+        out_msg      = gr.Markdown(label="Personalized message")
 
         # examples (render on top via CSS)
         if RangeSlider:
             example_inputs = [interests, occasion, budget_range, recipient_name, relationship, age, gender, tone]
             EXAMPLES = [
-                [["Technology","Music"], "Birthday", [20,60], "Noa", "Friend", "adult (18–64)", "any", "Heartfelt"],
-                [["Home decor","Cooking"], "Housewarming", [25,45], "Daniel", "Neighbor", "adult (18–64)", "male", "Appreciative"],
+                [["Reading","Fashion","Home decor"], "Valentine’s Day", [30,60], "Rotem", "Romantic partner", "adult (18–64)", "female", "Romantic"],
+                [["Technology","Movies"], "Birthday", [25,45], "Daniel", "Friend", "adult (18–64)", "male", "Funny"],
                 [["Gaming","Photography"], "Birthday", [30,120], "Omer", "Family - Sibling", "teen (13–17)", "male", "Playful"],
-                [["Reading","Art"], "Graduation", [15,35], "Maya", "Romantic partner", "any", "female", "Romantic"],
+                [["Home decor","Cooking"], "Housewarming", [25,45], "Noa", "Neighbor", "adult (18–64)", "any", "Appreciative"],
             ]
         else:
             example_inputs = [interests, occasion, budget_min, budget_max, recipient_name, relationship, age, gender, tone]
             EXAMPLES = [
-                [["Technology","Music"], "Birthday", 20, 60, "Noa", "Friend", "adult (18–64)", "any", "Heartfelt"],
-                [["Home decor","Cooking"], "Housewarming", 25, 45, "Daniel", "Neighbor", "adult (18–64)", "male", "Appreciative"],
+                [["Reading","Fashion","Home decor"], "Valentine’s Day", 30, 60, "Rotem", "Romantic partner", "adult (18–64)", "female", "Romantic"],
+                [["Technology","Movies"], "Birthday", 25, 45, "Daniel", "Friend", "adult (18–64)", "male", "Funny"],
                 [["Gaming","Photography"], "Birthday", 30, 120, "Omer", "Family - Sibling", "teen (13–17)", "male", "Playful"],
-                [["Reading","Art"], "Graduation", 15, 35, "Maya", "Romantic partner", "any", "female", "Romantic"],
+                [["Home decor","Cooking"], "Housewarming", 25, 45, "Noa", "Neighbor", "adult (18–64)", "any", "Appreciative"],
             ]
 
     with gr.Column(elem_id="examples"):
@@ -601,27 +760,29 @@ with gr.Blocks(css=CSS) as demo:
         top3 = recommend_topk(profile, k=3)
         top3_html = render_top3_html(top3)
 
-        # invented gift + image
-        gen = llm_generate_item(profile)
-        gen_md = f"**{md_escape(gen['name'])}**\n\n{md_escape(gen['short_desc'])}\n\n~${gen['price_usd']:.0f}"
-        gen_img = generate_gift_image(gen)
+        # DIY gift (generate-from-scratch, JSON)
+        diy_json, diy_status = diy_generate(profile)
+        diy_md = render_diy_md(diy_json)
+
+        # DIY image (optional)
+        diy_img = generate_gift_image_from_diy(diy_json)
 
         # greeting
         msg = llm_generate_message(profile)
 
-        return top3_html, gen_md, gen_img, msg
+        return top3_html, diy_json, diy_md, diy_img, msg
 
     if RangeSlider:
         go.click(
             ui_predict,
             [interests, occasion, budget_range, recipient_name, relationship, age, gender, tone],
-            [out_top3, out_gen_text, out_gen_img, out_msg]
+            [out_top3, out_diy_json, out_diy_md, out_gen_img, out_msg]
         )
     else:
         go.click(
             ui_predict,
             [interests, occasion, budget_min, budget_max, recipient_name, relationship, age, gender, tone],
-            [out_top3, out_gen_text, out_gen_img, out_msg]
+            [out_top3, out_diy_json, out_diy_md, out_gen_img, out_msg]
         )
 
 if __name__ == "__main__":