import spaces # MUST BE THE VERY FIRST IMPORT FOR ZEROGPU import os import io import re import json import uuid import zipfile import threading import xml.etree.ElementTree as ET from html.parser import HTMLParser from datetime import datetime, timezone from pathlib import Path import requests import torch import gradio as gr import sympy as sp from sympy.parsing.sympy_parser import ( parse_expr, standard_transformations, convert_xor, ) from transformers import ( AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer, ) from peft import PeftModel # ----------------------------- CONFIG ----------------------------- GITHUB_REPO = "rickeysharma942-stack/english-math-tutor" RELEASE_TAG = "chatbot" # local file name -> asset name on the GitHub release ASSETS = { "english_adapter": { "adapter_config.json": "english_adapter_config.json", "adapter_model.safetensors": "english_adapter_model.safetensors", }, "math_adapter": { "adapter_config.json": "math_adapter_config.json", "adapter_model.safetensors": "math_adapter_model.safetensors", }, "general_adapter": { "adapter_config.json": "general_adapter_config.json", "adapter_model.safetensors": "general_adapter_model.safetensors", }, } ADAPTER_DIR = Path("adapters") / RELEASE_TAG MAX_FILE_CHARS = 12000 # per attached file MAX_HISTORY_CHARS = 40000 # prompt budget (oldest turns are dropped first) MAX_CHATS = 40 # chats kept in the browser STORAGE_KEY = "tutor_studio_chats_v1" # Your adapters were trained on Qwen3-8B, so the base model MUST be Qwen3-8B. BASE_MODEL = os.environ.get("BASE_MODEL") or "Qwen/Qwen3-8B" # Set USE_4BIT=1 only on a normal GPU Space (T4/L4). Leave off for ZeroGPU. USE_4BIT = os.environ.get("USE_4BIT", "0") == "1" MODES = { "General": {"adapter": "general_adapter"}, "English": {"adapter": "english_adapter"}, "Math": {"adapter": "math_adapter"}, } # ----------------------------- SUPPORTED FILE TYPES (30) ----------------------------- DOC_TYPES = [".pdf", ".docx", ".xlsx", ".pptx", ".odt", ".rtf", ".epub", ".html", ".ipynb"] DATA_TYPES = [".txt", ".md", ".csv", ".tsv", ".json", ".xml", ".yaml", ".log", ".ini"] CODE_TYPES = [".py", ".js", ".ts", ".java", ".c", ".cpp", ".cs", ".go", ".php", ".sql", ".sh", ".css"] ALIAS_TYPES = [".htm", ".yml", ".toml", ".cfg", ".markdown", ".jsx", ".tsx", ".h", ".hpp", ".rb", ".rs"] SUPPORTED_30 = DOC_TYPES + DATA_TYPES + CODE_TYPES ACCEPTED_TYPES = SUPPORTED_30 + ALIAS_TYPES SPECIAL_TYPES = set(DOC_TYPES) | {".htm"} PLAIN_TEXT_TYPES = set(ACCEPTED_TYPES) - SPECIAL_TYPES # ----------------------------- MODEL LOADING ----------------------------- def download_file(url: str, dest: Path): dest.parent.mkdir(parents=True, exist_ok=True) tmp = dest.with_suffix(dest.suffix + ".part") with requests.get(url, stream=True, timeout=60, allow_redirects=True) as r: r.raise_for_status() with open(tmp, "wb") as f: for chunk in r.iter_content(chunk_size=1024 * 1024): if chunk: f.write(chunk) tmp.replace(dest) def ensure_adapters(): base_url = f"https://github.com/{GITHUB_REPO}/releases/download/{RELEASE_TAG}" for adapter_name, files in ASSETS.items(): for local_name, asset_name in files.items(): dest = ADAPTER_DIR / adapter_name / local_name if not dest.exists(): print(f"Downloading {asset_name} ...") download_file(f"{base_url}/{asset_name}", dest) def load_model(): ensure_adapters() tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL) if USE_4BIT: bnb = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, ) base = AutoModelForCausalLM.from_pretrained( BASE_MODEL, quantization_config=bnb, device_map={"": 0}, torch_dtype=torch.float16, ) else: # bf16 (~16 GB) fits on a ZeroGPU slice; no bitsandbytes needed. base = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype=torch.bfloat16, ) # stays on CPU for now # ZeroGPU has no real GPU at startup, so read adapter files onto the CPU # (PEFT would otherwise try to load them straight onto "cuda" and crash). dev = None if USE_4BIT else "cpu" model = PeftModel.from_pretrained( base, str(ADAPTER_DIR / "english_adapter"), adapter_name="english_adapter", torch_device=dev, ) model.load_adapter(str(ADAPTER_DIR / "math_adapter"), adapter_name="math_adapter", torch_device=dev) model.load_adapter(str(ADAPTER_DIR / "general_adapter"), adapter_name="general_adapter", torch_device=dev) model.eval() if not USE_4BIT: model.to("cuda") # ZeroGPU defers the real transfer until a @spaces.GPU call return tokenizer, model print(f"Loading {BASE_MODEL} and adapters...") tokenizer, model = load_model() generation_lock = threading.Lock() print("Model ready.") # ----------------------------- FILE PARSING ----------------------------- class _HTMLText(HTMLParser): def __init__(self): super().__init__() self.parts, self._skip = [], 0 def handle_starttag(self, tag, attrs): if tag in ("script", "style"): self._skip += 1 def handle_endtag(self, tag): if tag in ("script", "style") and self._skip: self._skip -= 1 def handle_data(self, data): if not self._skip and data.strip(): self.parts.append(data.strip()) def html_to_text(raw: str) -> str: p = _HTMLText() p.feed(raw) return "\n".join(p.parts) def _read_pdf(data: bytes) -> str: from pypdf import PdfReader reader = PdfReader(io.BytesIO(data)) pages = [] for i, page in enumerate(reader.pages, 1): t = (page.extract_text() or "").strip() if t: pages.append(f"[Page {i}]\n{t}") return "\n\n".join(pages) or "[No extractable text found - this PDF may be a scan.]" def _read_docx(data: bytes) -> str: from docx import Document doc = Document(io.BytesIO(data)) parts = [p.text for p in doc.paragraphs if p.text.strip()] for table in doc.tables: for row in table.rows: cells = [c.text.strip() for c in row.cells] if any(cells): parts.append(" | ".join(cells)) return "\n".join(parts) def _read_xlsx(data: bytes) -> str: from openpyxl import load_workbook wb = load_workbook(io.BytesIO(data), read_only=True, data_only=True) parts, size = [], 0 for sheet in wb.worksheets: parts.append(f"## Sheet: {sheet.title}") for row in sheet.iter_rows(values_only=True): row_str = " | ".join(str(v) for v in row if v is not None) if row_str.strip(): parts.append(row_str) size += len(row_str) if size > MAX_FILE_CHARS: return "\n".join(parts) return "\n".join(parts) def _read_pptx(data: bytes) -> str: from pptx import Presentation prs = Presentation(io.BytesIO(data)) parts = [] for i, slide in enumerate(prs.slides, 1): parts.append(f"## Slide {i}") for shape in slide.shapes: if getattr(shape, "has_text_frame", False) and shape.text_frame.text.strip(): parts.append(shape.text_frame.text.strip()) return "\n".join(parts) def _read_odt(data: bytes) -> str: with zipfile.ZipFile(io.BytesIO(data)) as z: root = ET.fromstring(z.read("content.xml")) parts = [] for el in root.iter(): if el.tag.endswith("}p") or el.tag.endswith("}h"): text = "".join(el.itertext()).strip() if text: parts.append(text) return "\n".join(parts) def _read_epub(data: bytes) -> str: parts = [] with zipfile.ZipFile(io.BytesIO(data)) as z: for name in sorted(z.namelist()): if name.lower().endswith((".xhtml", ".html", ".htm")): text = html_to_text(z.read(name).decode("utf-8", errors="replace")) if text.strip(): parts.append(text) return "\n\n".join(parts) def _read_rtf(data: bytes) -> str: raw = data.decode("utf-8", errors="replace") try: from striprtf.striprtf import rtf_to_text return rtf_to_text(raw) except Exception: text = re.sub(r"\\[a-z]+-?\d* ?|[{}]", "", raw) return text.strip() def _read_ipynb(data: bytes) -> str: nb = json.loads(data.decode("utf-8", errors="replace")) parts = [] for cell in nb.get("cells", []): src = cell.get("source", "") src = "".join(src) if isinstance(src, list) else str(src) if not src.strip(): continue if cell.get("cell_type") == "code": parts.append(f"```python\n{src}\n```") else: parts.append(src) return "\n\n".join(parts) def file_path_of(file_obj) -> Path: """Gradio may hand us a str path, a dict with 'path', or an object with .name.""" if isinstance(file_obj, str): return Path(file_obj) if isinstance(file_obj, dict): return Path(file_obj.get("path") or file_obj.get("name")) return Path(file_obj.name) def extract_text_from_upload(file_obj) -> str: path = file_path_of(file_obj) ext = path.suffix.lower() if ext not in ACCEPTED_TYPES: return f"[Unsupported file type: {ext or 'no extension'}]" try: data = path.read_bytes() if ext in PLAIN_TEXT_TYPES: return data.decode("utf-8", errors="replace") if ext == ".pdf": return _read_pdf(data) if ext == ".docx": return _read_docx(data) if ext == ".xlsx": return _read_xlsx(data) if ext == ".pptx": return _read_pptx(data) if ext == ".odt": return _read_odt(data) if ext == ".epub": return _read_epub(data) if ext == ".rtf": return _read_rtf(data) if ext == ".ipynb": return _read_ipynb(data) if ext in (".html", ".htm"): return html_to_text(data.decode("utf-8", errors="replace")) except Exception as e: return f"[Could not read {path.name}: {e}]" return "[Unsupported file type]" def build_file_context(files): """Returns (context_text, file_names, skipped_names).""" chunks, names, skipped = [], [], [] for f in files or []: path = file_path_of(f) names.append(path.name) if path.suffix.lower() not in ACCEPTED_TYPES: skipped.append(path.name) continue text = extract_text_from_upload(f) total = len(text) text = text[:MAX_FILE_CHARS] if total > MAX_FILE_CHARS: text += f"\n...[truncated: showing {MAX_FILE_CHARS} of {total} characters]" chunks.append(f"--- File: {path.name} ---\n{text}") return "\n\n".join(chunks), names, skipped # ----------------------------- MATH CHECKER ----------------------------- SYMPY_NAMES = [ "diff", "integrate", "limit", "solve", "simplify", "expand", "factor", "sin", "cos", "tan", "exp", "log", "sqrt", "pi", "E", ] SAFE_SYMPY = {n: getattr(sp, n) for n in SYMPY_NAMES if hasattr(sp, n)} SAFE_SYMPY.update({letter: sp.Symbol(letter) for letter in "xyztnabc"}) # parse_expr's auto_number / auto_symbol transforms emit these names, and # they must be resolvable because __builtins__ and globals are locked down. SAFE_SYMPY.update({ "Integer": sp.Integer, "Float": sp.Float, "Rational": sp.Rational, "Symbol": sp.Symbol, "Function": sp.Function, }) SYMPY_TRANSFORMS = standard_transformations + (convert_xor,) def run_sympy(expression: str) -> str: expression = (expression or "").strip() if not expression: return "Enter an expression first." try: result = sp.simplify( parse_expr( expression, local_dict=dict(SAFE_SYMPY), global_dict={"__builtins__": {}}, transformations=SYMPY_TRANSFORMS, ) ) return str(result) except Exception as e: return f"SymPy Error: {e}" # ----------------------------- GENERATION ----------------------------- def trim_messages(messages): """Drop the oldest turns until the prompt fits MAX_HISTORY_CHARS (always keep the last message).""" def total(ms): return sum(len(m["content"]) for m in ms) while len(messages) > 1 and total(messages) > MAX_HISTORY_CHARS: messages.pop(0) while len(messages) > 1 and messages[0]["role"] == "assistant": messages.pop(0) return messages class StopOnFlag(StoppingCriteria): """Lets us stop generation (and free the GPU) when the user presses Stop.""" def __init__(self, flag: threading.Event): self.flag = flag def __call__(self, input_ids, scores, **kwargs): return torch.full((input_ids.shape[0],), self.flag.is_set(), dtype=torch.bool, device=input_ids.device) @spaces.GPU(duration=120) def generate_response(messages, mode, temperature, max_new_tokens): adapter_name = MODES.get(mode, MODES["General"])["adapter"] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False, ) inputs = tokenizer(prompt, return_tensors="pt").to(model.device) streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) stop_flag = threading.Event() errors = [] temperature = float(temperature) gen_kwargs = dict( max_new_tokens=int(max_new_tokens), repetition_penalty=1.05, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.convert_tokens_to_ids("<|im_end|>"), stopping_criteria=StoppingCriteriaList([StopOnFlag(stop_flag)]), ) if temperature > 0.01: gen_kwargs.update(do_sample=True, temperature=temperature, top_p=0.8) else: gen_kwargs.update(do_sample=False) def worker(): try: with generation_lock: model.set_adapter(adapter_name) with torch.inference_mode(): model.generate(**inputs, streamer=streamer, **gen_kwargs) except Exception as e: # make sure the streaming loop never hangs errors.append(e) streamer.on_finalized_text("", stream_end=True) t = threading.Thread(target=worker, daemon=True) t.start() partial = "" try: for new_text in streamer: partial += new_text yield partial t.join() if errors: yield partial + f"\n\nโš ๏ธ Generation error: {errors[0]}" finally: stop_flag.set() # also fires when the user presses Stop t.join(timeout=10) # ----------------------------- CHAT STORE (kept in the visitor's own browser) ----------------------------- def now_iso(): return datetime.now(timezone.utc).isoformat() def new_chat_obj(): return {"id": uuid.uuid4().hex[:12], "title": "New chat", "updated": now_iso(), "turns": []} def ensure_store(store): if not isinstance(store, dict) or not isinstance(store.get("chats"), dict): store = {"chats": {}, "current": None} if store.get("current") not in store["chats"]: if store["chats"]: store["current"] = max(store["chats"].values(), key=lambda c: c["updated"])["id"] else: chat = new_chat_obj() store["chats"][chat["id"]] = chat store["current"] = chat["id"] return store def current_chat(store): return store["chats"][store["current"]] def normalize_store(store): """Turns that never got an answer (stopped / interrupted) get a note instead of staying blank.""" for cid, chat in store["chats"].items(): if cid == store.get("busy"): continue for turn in chat["turns"]: if not turn.get("a"): turn["a"] = "*(no response)*" return store def make_title(text, names): base = (text or "").strip().splitlines()[0] if (text or "").strip() else "" if not base and names: base = f"๐Ÿ“Ž {names[0]}" base = base or "New chat" return base if len(base) <= 38 else base[:37] + "โ€ฆ" def history_update(store): chats = sorted(store["chats"].values(), key=lambda c: c["updated"], reverse=True) return gr.update(choices=[(c["title"], c["id"]) for c in chats], value=store["current"]) def user_display(turn): parts = [] if turn.get("u"): parts.append(turn["u"]) if turn.get("files"): parts.append("\n".join(f"๐Ÿ“Ž {n}" for n in turn["files"])) return "\n\n".join(parts) def chat_display(chat): return [[user_display(t), (t.get("a") or "โ€ฆ")] for t in chat["turns"]] def store_to_json(store): store = ensure_store(store) data = {"chats": store["chats"], "current": store["current"]} s = json.dumps(data) # localStorage is ~5 MB: drop the oldest chats if we are getting close while len(s) > 3_500_000 and len(data["chats"]) > 1: oldest = min( (c for cid, c in data["chats"].items() if cid != data["current"]), key=lambda c: c["updated"], ) del data["chats"][oldest["id"]] s = json.dumps(data) return s def prune_chats(store): while len(store["chats"]) > MAX_CHATS: oldest = min( (c for cid, c in store["chats"].items() if cid != store["current"]), key=lambda c: c["updated"], ) del store["chats"][oldest["id"]] def build_messages(chat): msgs = [] for t in chat["turns"]: content = t.get("u") or "" if t.get("ctx"): content += "\n\nAttached files:\n" + t["ctx"] msgs.append({"role": "user", "content": content}) if t.get("a"): msgs.append({"role": "assistant", "content": t["a"]}) return trim_messages(msgs) # ----------------------------- UI HANDLERS ----------------------------- def on_load(raw): try: parsed = json.loads(raw) if raw else None except Exception: parsed = None store = normalize_store(ensure_store(parsed)) return store, history_update(store), chat_display(current_chat(store)) def new_chat(store): store = ensure_store(store) if store.get("busy"): return store, history_update(store), chat_display(current_chat(store)) if not current_chat(store)["turns"]: # already on an empty chat return store, history_update(store), chat_display(current_chat(store)) chat = new_chat_obj() store["chats"][chat["id"]] = chat store["current"] = chat["id"] prune_chats(store) return store, history_update(store), [] def select_chat(chat_id, store): store = ensure_store(store) if not store.get("busy") and chat_id in store["chats"]: store["current"] = chat_id normalize_store(store) return store, history_update(store), chat_display(current_chat(store)) def delete_chat(store): store = ensure_store(store) if store.get("busy"): return store, history_update(store), chat_display(current_chat(store)) del store["chats"][store["current"]] store["current"] = None store = normalize_store(ensure_store(store)) return store, history_update(store), chat_display(current_chat(store)) def add_user(msg, store): store = ensure_store(store) chat = current_chat(store) keep = gr.skip() if store.get("busy"): gr.Warning("Please wait for the current answer to finish (or press Stop).") return store, chat_display(chat), keep, history_update(store) msg = msg or {} text = (msg.get("text") or "").strip() files = msg.get("files") or [] if not text and not files: return store, chat_display(chat), keep, history_update(store) ctx, names, skipped = build_file_context(files) if skipped: gr.Warning("Unsupported file type skipped: " + ", ".join(skipped)) if not text: text = "Please summarize the attached file(s)." chat["turns"].append({"u": text, "files": names, "ctx": ctx, "a": ""}) if len(chat["turns"]) == 1: chat["title"] = make_title(text, names) chat["updated"] = now_iso() store["pending"] = True return store, chat_display(chat), {"text": "", "files": []}, history_update(store) def respond(store, mode, temperature, max_new_tokens): store = ensure_store(store) chat = current_chat(store) if not store.get("pending") or not chat["turns"]: yield store, chat_display(chat), gr.skip() return store["pending"] = False store["busy"] = chat["id"] turn = chat["turns"][-1] messages = build_messages(chat) yield store, chat_display(chat), gr.skip() try: for partial in generate_response(messages, mode, temperature, max_new_tokens): turn["a"] = partial if store.get("current") == chat["id"]: yield store, chat_display(chat), gr.skip() except Exception as e: turn["a"] = (turn.get("a") or "") + f"\n\nโš ๏ธ {e}" finally: store["busy"] = None if not turn.get("a"): turn["a"] = "*(no response)*" chat["updated"] = now_iso() yield store, chat_display(chat), history_update(store) def stop_save(store): store = ensure_store(store) store["busy"] = None store["pending"] = False normalize_store(store) return store, chat_display(current_chat(store)) # ----------------------------- GRADIO UI ----------------------------- CSS = """ :root { --bg:#faf9f5; --side:#f0eee6; --line:#e5e2d8; --ink:#1f1e1d; --muted:#6b6a65; --accent:#c96442; --accent-d:#b5573a; } body, .gradio-container { background: var(--bg) !important; color: var(--ink) !important; } .gradio-container { max-width: 100% !important; padding: 0 !important; font-family: 'Inter', system-ui, sans-serif !important; } footer { display: none !important; } /* ---- sidebar ---- */ #sidebar { background: var(--side); border-right: 1px solid var(--line); padding: 18px 14px !important; min-height: 100vh; position: sticky; top: 0; align-self: flex-start; max-height: 100vh; overflow-y: auto; gap: 10px !important; } #brand { font-family: Georgia, 'Times New Roman', serif; font-size: 1.35rem; color: var(--ink); margin: 2px 4px 6px; } #brand p { margin: 0 !important; } #new-chat { background: var(--accent) !important; color: #fff !important; border: none !important; border-radius: 10px !important; font-weight: 600; } #new-chat:hover { background: var(--accent-d) !important; } #recents-label { color: var(--muted); font-size: .75rem; text-transform: uppercase; letter-spacing: .06em; margin: 8px 6px 0; } #recents-label p { margin: 0 !important; } #history, #history .wrap { background: transparent !important; border: none !important; padding: 0 !important; box-shadow: none !important; } #history .wrap { display: flex; flex-direction: column; gap: 2px; } #history label { background: transparent !important; border: none !important; border-radius: 8px !important; padding: 8px 10px !important; width: 100%; cursor: pointer; color: var(--ink) !important; box-shadow: none !important; } #history label:hover { background: #e8e5da !important; } #history label.selected { background: #e3dfd2 !important; font-weight: 600; } #history label span { overflow: hidden; text-overflow: ellipsis; white-space: nowrap; display: block; max-width: 210px; } #history input { display: none !important; } #delete-chat { background: transparent !important; border: 1px solid var(--line) !important; color: var(--muted) !important; border-radius: 10px !important; } #delete-chat:hover { color: #b3261e !important; border-color: #b3261e !important; } #sidebar .block, #sidebar .form { background: transparent !important; border: none !important; box-shadow: none !important; } #sidebar .accordion, #sidebar .label-wrap { background: transparent !important; color: var(--ink) !important; } /* ---- main column ---- */ #main { max-width: 860px; margin: 0 auto; padding: 18px 20px 10px !important; } #chatbot { background: transparent !important; border: none !important; box-shadow: none !important; } #chatbot .message-row .message { border: none !important; box-shadow: none !important; font-size: 1rem; line-height: 1.65; } #chatbot .message.bot, #chatbot .bot.message { background: transparent !important; color: var(--ink) !important; padding-left: 4px !important; } #chatbot .message.user, #chatbot .user.message { background: var(--side) !important; color: var(--ink) !important; border-radius: 16px !important; } #chatbot .placeholder-content, #chatbot .placeholder { text-align: center; font-family: Georgia, serif; color: var(--ink); } #chatbot .placeholder h2 { font-weight: 400; font-size: 2rem; } #chatbot .placeholder p { font-family: 'Inter', sans-serif; color: var(--muted); } #mode-pills, #mode-pills .wrap { background: transparent !important; border: none !important; box-shadow: none !important; padding: 0 !important; justify-content: center; } #mode-pills label { border-radius: 999px !important; border: 1px solid var(--line) !important; background: #fff !important; padding: 6px 16px !important; color: var(--ink) !important; cursor: pointer; box-shadow: none !important; } #mode-pills label.selected { background: var(--accent) !important; border-color: var(--accent) !important; color: #fff !important; } #mode-pills input { display: none !important; } #composer { background: #fff !important; border: 1px solid var(--line) !important; border-radius: 22px !important; box-shadow: 0 2px 10px rgba(0,0,0,.05) !important; } #composer textarea { background: #fff !important; color: var(--ink) !important; } #composer button.submit-button, #composer .submit-button { background: var(--accent) !important; color: #fff !important; border-radius: 10px !important; } #stop-btn { background: transparent !important; border: 1px solid var(--line) !important; color: var(--muted) !important; border-radius: 10px !important; max-width: 160px; margin: 0 auto; } #hint { color: var(--muted); font-size: .8rem; text-align: center; } #hint p { margin: 0 !important; } @media (max-width: 800px) { #sidebar { min-height: auto; position: static; max-height: none; } } """ FORCE_LIGHT_JS = "() => { document.body.classList.remove('dark'); document.documentElement.classList.remove('dark'); }" LOAD_JS = f"""(v) => {{ try {{ return [localStorage.getItem('{STORAGE_KEY}') || '']; }} catch (e) {{ return ['']; }} }}""" SAVE_JS = f"""(v) => {{ try {{ localStorage.setItem('{STORAGE_KEY}', v); }} catch (e) {{ console.warn('Could not save chats', e); }} }}""" LATEX = [ {"left": "$$", "right": "$$", "display": True}, {"left": "\\[", "right": "\\]", "display": True}, {"left": "\\(", "right": "\\)", "display": False}, {"left": "$", "right": "$", "display": False}, ] WELCOME = "## โœจ How can I help you today?\n\nPick a mode, ask a question, or attach a file with the ๐Ÿ“Ž button." SUPPORTED_MD = ( "**Documents:** " + ", ".join(DOC_TYPES) + " \n" "**Data & text:** " + ", ".join(DATA_TYPES) + " \n" "**Code:** " + ", ".join(CODE_TYPES) + " \n" "**Also accepted:** " + ", ".join(ALIAS_TYPES) + " \n\n" f"Long files are cut to {MAX_FILE_CHARS:,} characters each. Images, audio and legacy " ".doc/.xls/.ppt files are not supported (save them as .docx/.xlsx/.pptx first)." ) theme = gr.themes.Base( primary_hue="orange", neutral_hue="stone", font=[gr.themes.GoogleFont("Inter"), "system-ui", "sans-serif"], ).set( body_background_fill="#faf9f5", body_text_color="#1f1e1d", block_background_fill="transparent", block_border_width="0px", input_background_fill="#ffffff", button_primary_background_fill="#c96442", button_primary_background_fill_hover="#b5573a", button_primary_text_color="#ffffff", ) with gr.Blocks(theme=theme, css=CSS, title="Tutor Studio", js=FORCE_LIGHT_JS) as demo: session = gr.State(None) # the chat store for this visitor (server-side, per session) persist_in = gr.Textbox(visible=False) # receives localStorage content on page load persist_out = gr.Textbox(visible=False) # carries JSON to the browser for saving with gr.Row(equal_height=False): # ---------------- sidebar ---------------- with gr.Column(scale=1, min_width=270, elem_id="sidebar"): gr.Markdown("โœจ Tutor Studio", elem_id="brand") new_btn = gr.Button("๏ผ‹ New chat", elem_id="new-chat") gr.Markdown("Recents", elem_id="recents-label") history_list = gr.Radio(choices=[], value=None, show_label=False, container=False, elem_id="history") delete_btn = gr.Button("๐Ÿ—‘ Delete this chat", elem_id="delete-chat", size="sm") with gr.Accordion("Generation settings", open=False): temp_slider = gr.Slider(0.0, 1.0, value=0.35, step=0.05, label="Creativity (temperature)") tokens_slider = gr.Slider(256, 4096, value=1600, step=128, label="Max response tokens") with gr.Accordion("๐Ÿงฎ SymPy math checker", open=False): sympy_input = gr.Textbox(label="Expression", value="diff(x**3 + 5*x, x)") sympy_output = gr.Textbox(label="Result", interactive=False) sympy_btn = gr.Button("Calculate") sympy_btn.click(run_sympy, inputs=sympy_input, outputs=sympy_output) with gr.Accordion("๐Ÿ“Ž Supported file types (30)", open=False): gr.Markdown(SUPPORTED_MD) # ---------------- main chat ---------------- with gr.Column(scale=4, elem_id="main"): chatbot = gr.Chatbot( elem_id="chatbot", show_label=False, height="62vh", bubble_full_width=False, show_copy_button=True, latex_delimiters=LATEX, placeholder=WELCOME, ) mode_radio = gr.Radio( choices=list(MODES.keys()), value="General", show_label=False, container=False, elem_id="mode-pills", ) composer = gr.MultimodalTextbox( elem_id="composer", show_label=False, container=False, file_types=ACCEPTED_TYPES, file_count="multiple", placeholder="Message Tutor Studioโ€ฆ (use ๐Ÿ“Ž to attach files)", lines=1, max_lines=8, ) stop_btn = gr.Button("โ–  Stop generating", elem_id="stop-btn", size="sm") gr.Markdown("Chats are saved only in this browser. Tutor Studio can make mistakes - check important answers.", elem_id="hint") # ---------------- events ---------------- def persist(event): """After any change: serialize the store and save it to the browser's localStorage.""" return event.then(store_to_json, [session], [persist_out], show_progress="hidden").then( None, [persist_out], None, js=SAVE_JS ) sidebar_outputs = [session, history_list, chatbot] demo.load(on_load, [persist_in], sidebar_outputs, js=LOAD_JS) persist(new_btn.click(new_chat, [session], sidebar_outputs, show_progress="hidden")) persist(history_list.input(select_chat, [history_list, session], sidebar_outputs, show_progress="hidden")) persist(delete_btn.click(delete_chat, [session], sidebar_outputs, show_progress="hidden")) send = composer.submit( add_user, [composer, session], [session, chatbot, composer, history_list], show_progress="hidden" ) generate = send.then( respond, [session, mode_radio, temp_slider, tokens_slider], [session, chatbot, history_list], show_progress="hidden", ) persist(generate) stop_event = stop_btn.click(stop_save, [session], [session, chatbot], cancels=[generate], show_progress="hidden") persist(stop_event) if __name__ == "__main__": demo.queue(max_size=20).launch()