# storefront_tabs_memory_app.py import os, json import gradio as gr from transformers import pipeline # ---------------- Model ---------------- MODEL_NAME = os.getenv("HF_MODEL_GENERATION", "distilgpt2") _pipe = None def _get_pipe(): global _pipe if _pipe is None: _pipe = pipeline("text-generation", model=MODEL_NAME) return _pipe def model_generate(prompt, max_new_tokens=128, temperature=0.8, top_p=0.95): out = _get_pipe()( prompt, max_new_tokens=int(max_new_tokens), do_sample=True, temperature=float(temperature), top_p=float(top_p), pad_token_id=50256, ) return out[0]["generated_text"] # ---------------- Storefront knowledge (helper module preferred) ---------------- STORE_DATA, USE_HELPERS = None, False try: # Optional helper module under agenticcore/ from agenticcore.storefront_rules import ( load_storefront, answer_faq, get_parking_rules, get_venue_rules, search_products ) STORE_DATA = load_storefront() USE_HELPERS = True except Exception: # Fallback: try JSON next to this file or under agenticcore/ CANDIDATES = [ os.path.join(os.path.dirname(__file__), "storefront_data.json"), os.path.join(os.path.dirname(__file__), "agenticcore", "storefront_data.json"), ] for p in CANDIDATES: if os.path.exists(p): with open(p, "r", encoding="utf-8") as f: STORE_DATA = json.load(f) break # Defaults if JSON/module missing DEFAULT_PRODUCTS = [ {"SKU": "CG-SET", "Name": "Cap & Gown Set", "Price": 59.00, "Notes": "Tassel included; ship until 10 days before event"}, {"SKU": "PK-1", "Name": "Parking Pass", "Price": 10.00, "Notes": "Multiple passes allowed per student"}, ] DEFAULT_PARKING = ["No double parking.", "Vehicles parked in handicap will be towed."] DEFAULT_VENUE = ["Formal attire recommended (not required).", "No muscle shirts.", "No sagging pants."] # Normalize JSON to tables for the UI if STORE_DATA: try: DEFAULT_PRODUCTS = [{ "SKU": p.get("sku",""), "Name": p.get("name",""), "Price": p.get("price_usd",""), "Notes": (p.get("description") or "")[:120], } for p in STORE_DATA.get("products", [])] DEFAULT_PARKING = STORE_DATA.get("policies", {}).get("parking_rules", DEFAULT_PARKING) or DEFAULT_PARKING DEFAULT_VENUE = STORE_DATA.get("policies", {}).get("venue_rules", DEFAULT_VENUE) or DEFAULT_VENUE except Exception: pass # ---------------- Memory seeding ---------------- def seed_storefront_facts() -> str: """Small system-like primer injected ahead of chat history to bias the LLM toward storefront truth.""" lines = ["You are answering questions about a graduation storefront.", "Products:"] for p in DEFAULT_PRODUCTS: price = p.get("Price") if isinstance(price, (int, float)): price = f"${price:.2f}" lines.append(f"- {p.get('Name','Item')} ({p.get('SKU','')}) — {price}: {p.get('Notes','')}") lines.append("Venue rules:") for r in DEFAULT_VENUE: lines.append(f"- {r}") lines.append("Parking rules:") for r in DEFAULT_PARKING: lines.append(f"- {r}") lines.append("Answer concisely using these facts. If unsure, say what’s known from the list above.") return "\n".join(lines) SEED_TEXT = seed_storefront_facts() def build_prompt_from_history(history, user_text, k=4): """history = [[user, bot], ...] — build a short rolling prompt + seed facts.""" lines = [SEED_TEXT, "", "Conversation so far:"] for u, b in (history or [])[-k:]: if u: lines.append(f"User: {u}") if b: lines.append(f"Assistant: {b}") lines.append(f"User: {user_text}") lines.append("Assistant:") return "\n".join(lines) # ---------------- Storefront Q&A router (storefront first, then LLM) ---------------- def storefront_qna(text: str) -> str | None: t = (text or "").lower().strip() if not t: return None # Single-word catches to avoid LLM drift if t in {"parking"}: return "Parking rules:\n- " + "\n- ".join(DEFAULT_PARKING) if t in {"venue", "attire", "dress", "dress code"}: return "Venue rules:\n- " + "\n- ".join(DEFAULT_VENUE) if t in {"passes", "parking pass", "parking passes"}: return "Yes, multiple parking passes are allowed per student." # Prefer helper functions if available if USE_HELPERS and STORE_DATA: a = answer_faq(STORE_DATA, t) if a: return a if "parking" in t and "rule" in t: r = get_parking_rules(STORE_DATA) if r: return "Parking rules:\n- " + "\n- ".join(r) if ("venue" in t and "rule" in t) or "attire" in t or "dress code" in t: r = get_venue_rules(STORE_DATA) if r: return "Venue rules:\n- " + "\n- ".join(r) # Specific timing phrasing to avoid hallucinated dates if "parking" in t and ("hours" in t or "time" in t or "open" in t): return "Parking lots open 2 hours before the ceremony." hits = search_products(STORE_DATA, t) if hits: return "\n".join( f"{p.get('name','Item')} — ${p.get('price_usd',0):.2f}: {p.get('description','')}" for p in hits ) return None # Fallback rules (no helper module) if "parking" in t and ("more than one" in t or "multiple" in t or "extra" in t): return "Yes, multiple parking passes are allowed per student." if "parking" in t and "rule" in t: return "Parking rules:\n- " + "\n- ".join(DEFAULT_PARKING) if "parking" in t and ("hours" in t or "time" in t or "open" in t): return "Parking lots open 2 hours before the ceremony." if "attire" in t or "dress code" in t or ("venue" in t and "rule" in t): return "Venue rules:\n- " + "\n- ".join(DEFAULT_VENUE) if "cap" in t or "gown" in t: return "\n".join( f"{p['Name']} — ${p['Price']:.2f}: {p['Notes']}" for p in DEFAULT_PRODUCTS ) return None def chat_pipeline(history, message, max_new_tokens=128, temperature=0.8, top_p=0.95): # 1) Try storefront knowledge first sf = storefront_qna(message) if sf: return sf # 2) Memory-aware model prompt prompt = build_prompt_from_history(history, message, k=4) return model_generate(prompt, max_new_tokens, temperature, top_p) # ---------------- Gradio UI (Tabs + Accordion) ---------------- CSS = """ :root { --bg:#0b0d12; --panel:#0f172a; --border:#1f2940; --text:#e5e7eb; --muted:#9ca3af; } .gradio-container { background: var(--bg) !important; color: var(--text) !important; } .panel { border:1px solid var(--border); border-radius:16px; background:var(--panel); } .small { font-size:12px; color: var(--muted); } """ with gr.Blocks(title="Storefront Chat", css=CSS) as demo: gr.Markdown("## Storefront Chat") # Keep a single source of truth for chat history (pairs of [user, bot]) history_state = gr.State([]) with gr.Tabs(): # --- TAB 1: Chat (sliders tucked into an accordion) --- with gr.TabItem("Chat"): with gr.Group(elem_classes=["panel"]): chat = gr.Chatbot(height=360, bubble_full_width=False, label="Chat") with gr.Row(): msg = gr.Textbox(placeholder="Ask about parking rules, attire, cap & gown, pickup times…", scale=5) send = gr.Button("Send", scale=1) # Quick chips with gr.Row(): chip1 = gr.Button("Parking rules", variant="secondary") chip2 = gr.Button("Multiple passes", variant="secondary") chip3 = gr.Button("Attire", variant="secondary") chip4 = gr.Button("When do lots open?", variant="secondary") # Advanced options hidden with gr.Accordion("Advanced chat options", open=False): max_new = gr.Slider(32, 512, 128, 1, label="Max new tokens") temp = gr.Slider(0.1, 1.5, 0.8, 0.05, label="Temperature") topp = gr.Slider(0.1, 1.0, 0.95, 0.05, label="Top-p") # Small utilities with gr.Row(): health_btn = gr.Button("Health", variant="secondary") caps_btn = gr.Button("Capabilities", variant="secondary") status_md = gr.Markdown("Status: not checked", elem_classes=["small"]) # --- TAB 2: Products --- with gr.TabItem("Products"): gr.Markdown("### Available Items") cols = list(DEFAULT_PRODUCTS[0].keys()) if DEFAULT_PRODUCTS else ["SKU","Name","Price","Notes"] data = [[p.get(c,"") for c in cols] for p in DEFAULT_PRODUCTS] _products_tbl = gr.Dataframe(headers=cols, value=data, interactive=False, wrap=True, label="Products") # --- TAB 3: Rules --- with gr.TabItem("Rules"): gr.Markdown("### Venue rules") gr.Markdown("- " + "\n- ".join(DEFAULT_VENUE)) gr.Markdown("### Parking rules") gr.Markdown("- " + "\n- ".join(DEFAULT_PARKING)) # --- TAB 4: Logistics --- with gr.TabItem("Logistics"): gr.Markdown( "### Event Logistics\n" "- Shipping available until 10 days before event (typ. 3–5 business days)\n" "- Pickup: Student Center Bookstore during the week prior to event\n" "- Graduates arrive 90 minutes early; guests 60 minutes early\n" "- Lots A & B open 2 hours before; overflow as needed\n" "\n*Try asking the bot:* “What time should I arrive?” • “Where do I pick up the gown?”" ) # ---------- Helpers that keep Chatbot history valid (list of [u,b]) ---------- def _append_bot_md(history, md_text): """Append a bot markdown message without breaking the [user, bot] format.""" history = history or [] return history + [[None, md_text]] # ---------- Callbacks ---------- def on_send(history, message, max_new_tokens, temperature, top_p): t = (message or "").strip() if not t: return history, history, "" # no-op history = (history or []) + [[t, None]] reply = chat_pipeline(history[:-1], t, max_new_tokens, temperature, top_p) history[-1][1] = reply # Return updated state AND what the Chatbot should render return history, history, "" def _health_cb(history): md = (f"### Status: ✅ Healthy\n" f"- Model: `{MODEL_NAME}`\n" f"- Storefront module: {'yes' if USE_HELPERS else 'no'}\n" f"- Storefront JSON: {'loaded' if bool(STORE_DATA) else 'not found'}") new_hist = _append_bot_md(history, md) return new_hist, new_hist, "Status: ✅ Healthy" def _caps_cb(history): caps = [ "Chat (LLM text-generation, memory-aware prompt)", "Storefront Q&A (parking, attire, products, logistics)", "Adjustable: max_new_tokens, temperature, top-p", ] md = "### Capabilities\n- " + "\n- ".join(caps) new_hist = _append_bot_md(history, md) return new_hist, new_hist # Wire up (note: always update both the state and the Chatbot value) send.click(on_send, [history_state, msg, max_new, temp, topp], [history_state, chat, msg]) msg.submit(on_send, [history_state, msg, max_new, temp, topp], [history_state, chat, msg]) # Chips chip1.click(lambda: "What are the parking rules?", outputs=msg) chip2.click(lambda: "Can I buy multiple parking passes?", outputs=msg) chip3.click(lambda: "Is formal attire required?", outputs=msg) chip4.click(lambda: "What time do the parking lots open?", outputs=msg) # Health / capabilities (append to chat, keep valid tuple format) health_btn.click(_health_cb, inputs=[history_state], outputs=[history_state, chat, status_md]) caps_btn.click(_caps_cb, inputs=[history_state], outputs=[history_state, chat]) if __name__ == "__main__": demo.launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", "7860")))