Upload 2 files

app.py

@@ -0,0 +1,197 @@
+# -*- coding: utf-8 -*-
+"""
+Jev-style 判別式決策互動 POC（繁體中文）
+後端：jevos（feder-cr/jev，OpenVINO int8，CPU 可跑）
+- 是非題 (noul)：輸入狀態＋問題，回傳 P(yes)
+- 多選題 (choice)：輸入狀態＋問題＋候選選項，回傳每選項機率
+- 每次回應標示延遲與 output_tokens=0（零解碼，判別 vs 生成的核心差異）
+
+免費部署：Hugging Face Spaces（CPU Basic 免費方案即可）
+  1. 在 huggingface.co/new-space 建一個 Gradio Space（公開）
+  2. 上傳 app.py＋requirements.txt
+  3. 首次啟動會自動下載 jevos（約 660MB）並快取於 /data，之後秒啟
+本地運行：見 README.md
+"""
+import json
+import os
+import subprocess
+import sys
+import threading
+import time
+import urllib.request
+import zipfile
+import tarfile
+
+import gradio as gr
+import requests
+
+BASE = os.path.dirname(os.path.abspath(__file__))
+DATA = "/data" if os.path.isdir("/data") and os.access("/data", os.W_OK) else os.path.join(BASE, ".poc-data")
+os.makedirs(DATA, exist_ok=True)
+
+BIN_URL = "https://github.com/feder-cr/jev/releases/download/jevos-v3/jev-linux-x64.tar.gz"
+MODEL_URL = "https://github.com/feder-cr/jev/releases/download/jevos-v3/jevos-v3-openvino-int8.zip"
+API = "http://127.0.0.1:8017/v1/systemone"
+_server = None
+_lock = threading.Lock()
+
+
+def _dl(url, dst):
+    if os.path.exists(dst):
+        return dst
+    print(f"[poc] 下載 {url}", flush=True)
+    tmp = dst + ".part"
+    urllib.request.urlretrieve(url, tmp)
+    os.rename(tmp, dst)
+    return dst
+
+
+def ensure_backend():
+    """下載、解壓、啟動 jevos（冪等，可重複呼叫）。"""
+    global _server
+    with _lock:
+        if _server is not None and _server.poll() is None:
+            return
+        _ensure_backend_locked()
+
+
+def _ensure_backend_locked():
+    global _server
+    # 先看是不是已經有 server 在跑（例如本機手動啟動的）→ 有就直接用，不下載
+    try:
+        requests.post(API, json={"model": "jev-latest", "state": "ping",
+                                 "questions": {"q": {"type": "noul", "instructions": "ok?"}}},
+                      timeout=5)
+        print("[poc] jevos 已在運行", flush=True)
+        return
+    except Exception:
+        pass
+    jev_dir = os.path.join(DATA, "jev")
+    binary = os.path.join(jev_dir, "jev")
+    model_dir = os.path.join(jev_dir, "model")
+    if not (os.path.exists(binary) and os.path.isdir(model_dir)):
+        os.makedirs(DATA, exist_ok=True)
+        tgz = _dl(BIN_URL, os.path.join(DATA, "jev-linux-x64.tar.gz"))
+        with tarfile.open(tgz) as tf:
+            tf.extractall(DATA)
+        zf = _dl(MODEL_URL, os.path.join(DATA, "jevos-v3-openvino-int8.zip"))
+        with zipfile.ZipFile(zf) as z:
+            z.extractall(jev_dir)  # 內含 model/
+        os.chmod(binary, 0o755)
+    print("[poc] 啟動 jevos…", flush=True)
+    _server = subprocess.Popen([binary, "serve"], cwd=jev_dir,
+                               stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    for _ in range(60):
+        time.sleep(1)
+        try:
+            requests.post(API, json={"model": "jev-latest", "state": "ping",
+                                     "questions": {"q": {"type": "noul", "instructions": "ok?"}}},
+                          timeout=5)
+            print("[poc] jevos 就緒", flush=True)
+            return
+        except Exception:
+            continue
+    raise RuntimeError("jevos 啟動逾時")
+
+
+def decide(state, mode, question, options_text):
+    if not (state or "").strip() or not (question or "").strip():
+        return "請先輸入狀態文字與問題", "", ""
+    ensure_backend()
+    t0 = time.perf_counter()
+    if mode == "是非題":
+        payload = {"model": "jev-latest", "state": state,
+                   "questions": {"q1": {"type": "noul", "instructions": question}}}
+    else:
+        criteria = {}
+        for line in options_text.strip().split("\n"):
+            line = line.strip()
+            if not line:
+                continue
+            if ":" in line:
+                name, desc = line.split(":", 1)
+            elif "：" in line:
+                name, desc = line.split("：", 1)
+            else:
+                name, desc = line, ""
+            name = name.strip()
+            desc = desc.strip() or name  # 無描述時用名稱代替（choice 需要 criteria 描述）
+            criteria[name] = desc
+        if not criteria:
+            return "請至少輸入一個候選選項", "", ""
+        payload = {"model": "jev-latest", "state": state,
+                   "questions": {"q1": {"type": "choice", "instructions": question,
+                                       "criteria": criteria}}}
+    try:
+        r = requests.post(API, json=payload, timeout=120)
+        r.raise_for_status()
+        data = r.json()
+    except Exception as e:
+        return f"後端呼叫失敗：{type(e).__name__}（{e}）", "", ""
+    ms = (time.perf_counter() - t0) * 1000
+    try:
+        ans = data["answers"]["q1"]
+    except (KeyError, TypeError):
+        return f"後端回傳格式異常：{str(data)[:300]}", "", ""
+    usage = data.get("usage", {})
+    if ans["type"] == "noul":
+        p = ans["noul"]
+        bars = f"P(yes) = {p:.3f}\n" + "█" * int(p * 40) + f" {p*100:.1f}%"
+        summary = f"答案：{'是' if p >= 0.5 else '否'}（信心 {p if p>=0.5 else 1-p:.1%}）"
+    else:
+        probs = ans["probabilities"]
+        bars = "\n".join(f"{k}: {'█' * int(v*40)} {v*100:.1f}%"
+                         for k, v in sorted(probs.items(), key=lambda x: -x[1]))
+        summary = (f"答案：{ans['choice']}（信心 {ans.get('confidence', 0):.1%}）\n"
+                   f"各選項機率：{json.dumps(probs, ensure_ascii=False)}")
+    meta = (f"延遲 {ms:.0f} ms ｜ input_tokens={usage.get('input_tokens','?')} ｜ "
+            f"output_tokens={usage.get('output_tokens','?')}（零解碼）")
+    return summary, bars, meta
+
+
+PRESETS = {
+    "客服訊息分類": ("客人說：「這個月帳單怎麼多了一筆我沒買的東西？」",
+                    "多選題", "這封訊息應該派給哪個團隊？",
+                    "billing: 帳單、付款、退款\nshipping: 寄送、包裹遺失損壞\ntech: 產品故障不會動"),
+    "是否為退款請求": ("I want my money back for the broken headphones.",
+                    "是非題", "這是一則退款請求嗎？", ""),
+    "內容審核": ("這家餐廳難吃到我想給一星，老闆還敢嗆客人。",
+                    "是非題", "這則評論含有針對店家的惡意攻擊嗎？", ""),
+}
+
+
+def load_preset(name):
+    s, m, q, o = PRESETS[name]
+    return s, m, q, o
+
+
+CSS = """
+.gradio-container { max-width: 900px !important; }
+.bar textarea { font-family: monospace !important; white-space: pre !important; }
+"""
+
+with gr.Blocks(title="Jev-style 判別式決策 POC") as demo:
+    gr.Markdown("# ⚡ Jev-style 判別式決策 POC\n"
+                "輸入一段**狀態文字**＋**問題**，模型直接輸出**機率分佈**——不生成任何文字（`output_tokens=0`）。"
+                "這就是 System 1 判別引擎：傳統 LLM Agent 每步 1–3 秒還可能寫壞 JSON，這裡一次前向傳播就給出完整機率分佈。")
+    with gr.Row():
+        preset = gr.Dropdown(list(PRESETS.keys()), label="範例一鍵填入", value="客服訊息分類")
+    state = gr.Textbox(label="狀態文字（state）", lines=3)
+    mode = gr.Radio(["是非題", "多選題"], label="題型", value="多選題")
+    question = gr.Textbox(label="問題（instructions）", lines=1)
+    options = gr.Textbox(label="候選選項（多選題用，每行「名稱: 描述」）", lines=4)
+    btn = gr.Button("送出判別", variant="primary")
+    out_summary = gr.Textbox(label="判別結果")
+    out_bars = gr.Textbox(label="機率分佈", elem_classes="bar", lines=6)
+    out_meta = gr.Textbox(label="效能", lines=1)
+
+    preset.change(load_preset, preset, [state, mode, question, options])
+    btn.click(decide, [state, mode, question, options],
+              [out_summary, out_bars, out_meta])
+    demo.load(load_preset, preset, [state, mode, question, options])
+
+if __name__ == "__main__":
+    # Spaces 啟動時先備好後端（背景執行，不擋 UI）
+    threading.Thread(target=ensure_backend, daemon=True).start()
+    demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)),
+                css=CSS)

requirements.txt

@@ -0,0 +1,2 @@
+gradio>=6.0
+requests>=2.31.0