a3216 commited on
Commit
e0e5f2c
·
verified ·
1 Parent(s): 5346e6e

sync from GitHub 943c198: feat(tex): 公式改用固定槽位模板(去掉 list-item 内的 for),ensure 失败带 HTTP 状态码

Browse files
Dockerfile CHANGED
@@ -24,10 +24,14 @@ ENV MPLCONFIGDIR=/tmp/matplotlib
24
  # 预热 mathtext 字体缓存 + 在构建期就验证公式渲染可用:
25
  # 否则第一个含公式的回复要等字体缓存建立(数秒)。断言失败会让镜像构建失败,
26
  # 避免部署出一个「公式静默不渲染」的后端。
27
- # 注意:build_texs 吃的是「正文」,公式必须带 $...$ 分隔符;
28
- # 外层用 shell 单引号包住 python -c,避免 $ 被 shell 变量展开。
 
 
 
 
29
  RUN mkdir -p /tmp/matplotlib \
30
- && python -c 'from app.services.formula import build_texs, is_available; assert is_available(), "matplotlib unavailable"; r = build_texs("$a^2 + b^2 = c^2$"); assert r, "formula render failed"; print("formula render ok:", r[0]["w"], "x", r[0]["h"])'
31
 
32
  ENV PORT=7860
33
  ENV PYTHONUNBUFFERED=1
 
24
  # 预热 mathtext 字体缓存 + 在构建期就验证公式渲染可用:
25
  # 否则第一个含公式的回复要等字体缓存建立(数秒)。断言失败会让镜像构建失败,
26
  # 避免部署出一个「公式静默不渲染」的后端。
27
+ # 注意:ensure_batch 吃的是「不带 $ 的 LaTeX 正文」(与端上 /v1/xtc/tex/ensure
28
+ # 的入参一致);这里还顺带断言两件事——
29
+ # 1) 单批 12 条必须全部渲染回来(_MAX_FORMULAS 若被改小,端上一次请求里的
30
+ # 多余公式会被截掉并永久降级成纯文本);
31
+ # 2) 下发的 id 必须能在 LRU 里直接命中(降级渲染时 id 用实际 scale 计算,
32
+ # 键不一致会让每次回源取图都重渲染一遍)。
33
  RUN mkdir -p /tmp/matplotlib \
34
+ && python -c 'from app.services.formula import ensure_batch, get_png, is_available; assert is_available(), "matplotlib unavailable"; items = [{"latex": "x_%d^2 + y_%d^2 = r_%d^2" % (i, i, i), "key": "k%d" % i} for i in range(12)]; r = ensure_batch(items); assert len(r) == 12, "formula batch truncated: %d/12" % len(r); assert get_png(r[0]["id"]), "formula cache id mismatch"; print("formula render ok:", r[0]["id"], r[0]["w"], "x", r[0]["h"], "batch", len(r))'
35
 
36
  ENV PORT=7860
37
  ENV PYTHONUNBUFFERED=1
_chem_check.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ r"""化学公式渲染覆盖自检(保留作回归用)。
2
+
3
+ 背景:\ce{...}(mhchem 宏包语法)mathtext 不认,formula.py 里有一层手写转换
4
+ (_convert_mhchem / _replace_ce)。这里把常见化学写法一条条过一遍,看哪些真能
5
+ 渲出来、转换后的 mathtext 长什么样、哪些按设计「不渲染」(端上退回 LaTeX 原文)。
6
+
7
+ 注意:端上发过来的是**去掉分隔符的公式正文**($$ / \[ \] / $ 在前端就被剥掉了),
8
+ 所以这里不要拿带 \[ \] 的整段正文去测——那条路径根本不会走到后端。
9
+ """
10
+ from __future__ import annotations
11
+
12
+ import os
13
+ import sys
14
+
15
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
16
+
17
+ from app.services import formula as F # noqa: E402
18
+
19
+ # (名称, 公式正文, 期望结果) expect: 'ok' 渲染成功 / 'none' 按设计不渲染
20
+ CASES = [
21
+ ("氢气燃烧(参考公式 10)", r"\ce{2H2 + O2 -> 2H2O}", "ok"),
22
+ ("合成氨(参考公式 11)", r"\ce{N2 + 3H2 <=> 2NH3}", "ok"),
23
+ ("可逆号 <->", r"\ce{A <-> B}", "ok"),
24
+ ("括号化合物 + 上标电荷", r"\ce{Ca(OH)2 -> Ca^2+ + 2OH-}", "ok"),
25
+ ("醋酸电离", r"\ce{CH3COOH <=> CH3COO- + H+}", "ok"),
26
+ ("络合平衡", r"\ce{Fe^3+ + 3SCN- <=> Fe(SCN)3}", "ok"),
27
+ ("高锰酸钾氧化", r"\ce{MnO4- + 8H+ + 5e- -> Mn^2+ + 4H2O}", "ok"),
28
+ ("呼吸作用(多系数)", r"\ce{C6H12O6 + 6O2 -> 6CO2 + 6H2O}", "ok"),
29
+ ("纯分子式", r"\ce{H2SO4}", "ok"),
30
+ ("带状态符号", r"\ce{H2O(l) -> H2O(g)}", "ok"),
31
+ ("同位素(嵌套花括号)", r"\ce{^{14}C -> ^{14}N + e-}", "ok"),
32
+ ("气体上升号 ^", r"\ce{2Na + 2H2O -> 2NaOH + H2 ^}", "ok"),
33
+ ("沉淀号 v", r"\ce{Ag+ + Cl- -> AgCl v}", "ok"),
34
+ ("两个 \\ce 连着写", r"\ce{H2} + \ce{O2}", "ok"),
35
+ ("平衡常数(普通公式 12)", r"K_c = \frac{[C]^c [D]^d}{[A]^a [B]^b}", "ok"),
36
+ ("化学式 + 单位文字混排", r"\ce{2H2(g) + O2(g) -> 2H2O(l)} \quad \Delta H = -571.6\ \text{kJ/mol}", "ok"),
37
+ ("中文混进公式 → 不渲染(否则画成空心方块)", r"燃烧的方程式是 \ce{2H2 + O2 -> 2H2O},放热。", "none"),
38
+ ("中文注解块级公式 → 不渲染", "热力学第一定律:\\Delta U = Q + W", "none"),
39
+ ("矩阵 → 不渲染(mathtext 硬限制)", r"\begin{pmatrix} a & b \\ c & d \end{pmatrix}", "none"),
40
+ ("\\ce 花括号没闭合 → 不渲染", r"\ce{2H2 + O2", "none"),
41
+ ]
42
+
43
+ ok = 0
44
+ bad = 0
45
+ for name, tex, expect in CASES:
46
+ lines = F.normalize_latex(tex)
47
+ info = F.render(tex, fs=12, scale=3)
48
+ got = "ok" if (info and lines) else "none"
49
+ hit = (got == expect)
50
+ if hit:
51
+ ok += 1
52
+ else:
53
+ bad += 1
54
+ print("[%s] %-34s %s" % ("OK " if hit else "BAD ", name, "渲染成功" if got == "ok" else "不渲染 → 退回原文"))
55
+ if lines:
56
+ print(" 转换后:%s" % " | ".join(lines))
57
+ if got == "ok":
58
+ print(" 体积 %5dB design=%sx%s" % (info["bytes"], info["w"], info["h"]))
59
+ else:
60
+ print(" 预处理:%s" % ("判定不可渲染" if not lines else "mathtext 解析失败"))
61
+
62
+ print("\n合计 %d 条:符合预期 %d / 不符合 %d" % (len(CASES), ok, bad))
_ref_test.py ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """用用户给的 13 条参考公式验证识别/预处理/渲染覆盖率。"""
2
+ import io
3
+ import os
4
+ import sys
5
+
6
+ sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
7
+
8
+ from PIL import Image # noqa: E402
9
+
10
+ from app.services import formula # noqa: E402
11
+
12
+ SAMPLES = [
13
+ ("1 欧拉恒等式", r"\[e^{i\pi} + 1 = 0\]"),
14
+ ("2 高斯积分", r"\[\int_{-\infty}^{\infty} e^{-x^2}\,dx = \sqrt{\pi}\]"),
15
+ ("3 巴塞尔问题", r"\[\sum_{n=1}^{\infty} \frac{1}{n^2} = \frac{\pi^2}{6}\]"),
16
+ ("4 傅里叶变换", r"\[\hat{f}(\xi) = \int_{-\infty}^{\infty} f(x) e^{-2\pi i x \xi}\,dx\]"),
17
+ ("5 质能方程", r"\[E = mc^2\]"),
18
+ ("6 万有引力", r"\[F = G \frac{m_1 m_2}{r^2}\]"),
19
+ ("7 薛定谔方程", "\\[i\\hbar \\frac{\\partial}{\\partial t}\\Psi(\\mathbf{r},t)\n"
20
+ "= -\\frac{\\hbar^2}{2m}\\nabla^2 \\Psi(\\mathbf{r},t)\n"
21
+ "+ V(\\mathbf{r},t)\\Psi(\\mathbf{r},t)\\]"),
22
+ ("8 麦克斯韦方程组", "\\[\\begin{aligned}\n\\nabla \\cdot \\mathbf{E} &= \\frac{\\rho}{\\varepsilon_0}, &\n"
23
+ "\\nabla \\cdot \\mathbf{B} &= 0, \\\\\n\\nabla \\times \\mathbf{E} &= -\\frac{\\partial \\mathbf{B}}{\\partial t}, &\n"
24
+ "\\nabla \\times \\mathbf{B} &= \\mu_0 \\mathbf{J}\n+ \\mu_0 \\varepsilon_0 \\frac{\\partial \\mathbf{E}}{\\partial t}.\n\\end{aligned}\\]"),
25
+ ("9 热力学第一定律", r"\[\Delta U = Q - W\]"),
26
+ ("10 氢气燃烧", r"\[\ce{2H2 + O2 -> 2H2O}\]"),
27
+ ("11 合成氨", r"\[\ce{N2(g) + 3H2(g) <=> 2NH3(g)} \qquad \Delta H < 0\]"),
28
+ ("12 平衡常数", r"\[K_c = \frac{[\ce{NH3}]^2}{[\ce{N2}][\ce{H2}]^3}\]"),
29
+ ("13 理想气体", r"\[PV = nRT\]"),
30
+ ]
31
+
32
+ # 端上识别(JS)会先把分隔符剥掉,这里模拟同样的输入
33
+ def strip_delim(s):
34
+ t = s.strip()
35
+ for a, b in (("\\[", "\\]"), ("$$", "$$"), ("\\(", "\\)")):
36
+ if t.startswith(a) and t.endswith(b):
37
+ return t[len(a):len(t) - len(b)].strip()
38
+ return t
39
+
40
+
41
+ ok = bad = 0
42
+ for name, raw in SAMPLES:
43
+ body = strip_delim(raw)
44
+ lines = formula.normalize_latex(body)
45
+ info = formula.render(body, fs=16)
46
+ if info:
47
+ ok += 1
48
+ img = Image.open(io.BytesIO(info["png"]))
49
+ print("[OK ] %-16s %d 行 png=%dx%d design=%sx%s %5dB %s" % (
50
+ name, len(lines or []), img.size[0], img.size[1],
51
+ info["w"], info["h"], info["bytes"], img.mode))
52
+ with open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "_ref_%s.png" % name.split()[0]), "wb") as fp:
53
+ fp.write(info["png"])
54
+ else:
55
+ bad += 1
56
+ print("[FAIL] %-16s %s" % (name, "normalize 返回 None" if not lines else "渲染失败"))
57
+
58
+ print("\n合计 %d 条:成功 %d / 失败 %d" % (len(SAMPLES), ok, bad))
_tex_e2e_test.py ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """公式渲染后端:真实 HTTP 端到端自检(临时脚本,验证完可删)。
2
+
3
+ 与 _ref_test.py 的区别:那个只调服务层函数,这个起 uvicorn(完整 app.main:app),
4
+ 真发 HTTP 请求(stdlib urllib,本机 Python 3.14 上 httpx 0.28.1 解析 URL 就崩),
5
+ 覆盖端上实际会走的两条路径:
6
+
7
+ POST /v1/xtc/tex/ensure (带 x-xtc-access-key,批量渲染)
8
+ GET /v1/xtc/tex/{id}.png (capability URL,端上 request.download 直接取)
9
+
10
+ 检查项:
11
+ 1. 鉴权:没带 access key 的 /ensure → 401
12
+ 2. 批大小:12 条一次全部渲染回来(旧上限 4 会把后 8 条静默截断)
13
+ 3. 降级:矩阵这类 mathtext 排不了的公式不出现在响应里(端上据此保留原文)
14
+ 4. 图片本身:PNG 签名、image/png、<=50KB、RGBA 且真有透明像素、
15
+ 像素尺寸 = design 尺寸 * scale
16
+ 5. 内容寻址:同一批两次请求 id 完全一致;重复取图字节一致
17
+ 6. 体积兜底:超 50KB 的公式自动降 scale,且下发 id 能直接命中缓存(不重复渲染)
18
+ 7. 防滥用:id 与 f/fs/scale 不一致 → 403;未知 id 无 f → 404;限流 → 429
19
+ 8. 参数校验:fs 越界 / 缺 items → 400
20
+
21
+ 跑法:cd d:\\AI\\xtc-backend-hf && python _tex_e2e_test.py
22
+ """
23
+ from __future__ import annotations
24
+
25
+ import io
26
+ import json
27
+ import os
28
+ import sys
29
+ import threading
30
+ import time
31
+ import urllib.error
32
+ import urllib.request
33
+
34
+ HERE = os.path.dirname(os.path.abspath(__file__))
35
+ os.environ.setdefault("XTC_DB_PATH", os.path.join(os.path.dirname(HERE), "_tex_e2e.db"))
36
+ # 端上就是带 access key 调的;这里设一个 key 顺便验证鉴权链路
37
+ os.environ["XTC_ACCESS_KEY"] = "e2e-access-key-0123456789"
38
+ os.environ["XTC_JWT_SECRET"] = "e2e-jwt-secret-0123456789"
39
+ os.environ.setdefault("XTC_ADMIN_KEY", "e2e-admin-key-0123456789")
40
+
41
+ sys.path.insert(0, HERE)
42
+
43
+ import uvicorn # noqa: E402
44
+
45
+ from app.main import app # noqa: E402
46
+ from app.services import formula as F # noqa: E402
47
+ from app.api import tex as TEX_API # noqa: E402
48
+
49
+ PORT = 8813
50
+ BASE = "http://127.0.0.1:%d" % PORT
51
+ KEY = "e2e-access-key-0123456789"
52
+
53
+ PASS = 0
54
+ FAIL = 0
55
+
56
+
57
+ def check(name: str, ok: bool, detail: str = "") -> None:
58
+ global PASS, FAIL
59
+ if ok:
60
+ PASS += 1
61
+ print(" [PASS] %s %s" % (name, detail))
62
+ else:
63
+ FAIL += 1
64
+ print(" [FAIL] %s %s" % (name, detail))
65
+
66
+
67
+ def http(method: str, path: str, *, body=None, headers=None):
68
+ """返回 (status, headers_dict, content_bytes)。"""
69
+ url = path if path.startswith("http") else BASE + path
70
+ data = None
71
+ hdrs = {k: v for k, v in (headers or {}).items()}
72
+ if body is not None:
73
+ data = json.dumps(body).encode("utf-8")
74
+ hdrs["Content-Type"] = "application/json"
75
+ req = urllib.request.Request(url, data=data, headers=hdrs, method=method)
76
+ try:
77
+ with urllib.request.urlopen(req, timeout=120) as resp:
78
+ return resp.status, {k.lower(): v for k, v in resp.headers.items()}, resp.read()
79
+ except urllib.error.HTTPError as e:
80
+ return e.code, {k.lower(): v for k, v in (e.headers or {}).items()}, e.read()
81
+
82
+
83
+ def post_json(path, body, headers=None):
84
+ st, _h, raw = http("POST", path, body=body, headers=headers)
85
+ try:
86
+ return st, json.loads(raw.decode("utf-8"))
87
+ except Exception:
88
+ return st, {}
89
+
90
+
91
+ def start_server() -> uvicorn.Server:
92
+ cfg = uvicorn.Config(app, host="127.0.0.1", port=PORT, log_level="warning", access_log=False)
93
+ srv = uvicorn.Server(cfg)
94
+ threading.Thread(target=srv.run, daemon=True).start()
95
+ for _ in range(300):
96
+ if srv.started:
97
+ return srv
98
+ time.sleep(0.05)
99
+ raise RuntimeError("uvicorn 没起来")
100
+
101
+
102
+ def png_info(raw: bytes):
103
+ from PIL import Image
104
+
105
+ img = Image.open(io.BytesIO(raw))
106
+ return img.mode, img.size, (img.getchannel("A").getextrema() if img.mode == "RGBA" else None)
107
+
108
+
109
+ def qs(url: str) -> dict:
110
+ return dict(p.split("=", 1) for p in url.split("?", 1)[1].split("&"))
111
+
112
+
113
+ def main() -> int:
114
+ srv = start_server()
115
+ print("uvicorn 已启动:%s\n" % BASE)
116
+ try:
117
+ # ---------- 1. 鉴权 ----------
118
+ print("1) 鉴权")
119
+ st, _ = post_json("/v1/xtc/tex/ensure", {"items": [{"latex": "a^2", "key": "k"}]})
120
+ check("无 access key → 401", st == 401, "got %d" % st)
121
+ st, _ = post_json("/v1/xtc/tex/ensure", {"items": [{"latex": "a^2", "key": "k"}]},
122
+ {"x-xtc-access-key": KEY})
123
+ check("带 access key → 200", st == 200, "got %d" % st)
124
+
125
+ # ---------- 2. 批大小:12 条 ----------
126
+ print("2) 批大小(端上单条消息上限 12)")
127
+ items = [{"latex": "x_{%d}^{2} + y_{%d}^{2} = r_{%d}^{2}" % (i, i, i), "key": "key%d" % i}
128
+ for i in range(F._MAX_FORMULAS)]
129
+ st, body = post_json("/v1/xtc/tex/ensure", {"fs": 12, "scale": 3, "items": items},
130
+ {"x-xtc-access-key": KEY})
131
+ got = body.get("items") or []
132
+ check("12 条全部返回", len(got) == 12, "返回 %d 条" % len(got))
133
+ check("key 原样回传", len(got) == 12 and all(got[i].get("key") == "key%d" % i for i in range(12)))
134
+ check("有 id/url/w/h", all(g.get("id") and g.get("url") and g.get("w") and g.get("h") for g in got))
135
+ if got:
136
+ print(" 示例:%s -> %sx%s %s" % (got[0]["id"], got[0]["w"], got[0]["h"], got[0]["url"][:70]))
137
+
138
+ # ---------- 3. 降级 ----------
139
+ print("3) 渲染不了的公式必须不出现在响应里(端上据此保留 LaTeX 原文)")
140
+ mix = [
141
+ {"latex": "a^2 + b^2 = c^2", "key": "ok1"},
142
+ {"latex": "\\begin{pmatrix} a & b \\\\ c & d \\end{pmatrix}", "key": "matrix"},
143
+ {"latex": "\\int_0^1 x\\,dx = \\frac{1}{2}", "key": "ok2"},
144
+ {"latex": "\\begin{array}{cc} 1 & 2 \\end{array}", "key": "array"},
145
+ {"latex": "x" * 600, "key": "toolong"},
146
+ ]
147
+ _st, body = post_json("/v1/xtc/tex/ensure", {"items": mix}, {"x-xtc-access-key": KEY})
148
+ keys = [it.get("key") for it in (body.get("items") or [])]
149
+ check("矩阵/array/超长都不下发", keys == ["ok1", "ok2"], "返回 %s" % keys)
150
+
151
+ # ---------- 4. 取图 ----------
152
+ print("4) GET 取图")
153
+ url = got[0]["url"]
154
+ scale = int(qs(url)["scale"])
155
+ st, hdrs, raw = http("GET", url)
156
+ check("200 + image/png", st == 200 and hdrs.get("content-type") == "image/png",
157
+ "%d %s" % (st, hdrs.get("content-type")))
158
+ check("PNG 签名", raw[:8] == b"\x89PNG\r\n\x1a\n")
159
+ check("单张 <=50KB", len(raw) <= 50 * 1024, "%d 字节" % len(raw))
160
+ check("Cache-Control immutable", "immutable" in (hdrs.get("cache-control") or ""))
161
+ mode, size, alpha = png_info(raw)
162
+ check("RGBA 真彩色", mode == "RGBA", mode)
163
+ check("真有透明像素", alpha is not None and alpha[0] == 0 and alpha[1] == 255, str(alpha))
164
+ want_w = round(got[0]["w"] * scale)
165
+ want_h = round(got[0]["h"] * scale)
166
+ check("像素尺寸 = design*scale", abs(size[0] - want_w) <= 1 and abs(size[1] - want_h) <= 1,
167
+ "png %s vs 期望 %sx%s" % (size, want_w, want_h))
168
+
169
+ # ---------- 5. 内容寻址 ----------
170
+ print("5) 内容寻址 / 幂等")
171
+ _st, body2 = post_json("/v1/xtc/tex/ensure", {"fs": 12, "scale": 3, "items": items},
172
+ {"x-xtc-access-key": KEY})
173
+ ids2 = [it["id"] for it in (body2.get("items") or [])]
174
+ check("同内容同 id", ids2 == [g["id"] for g in got])
175
+ _st, _h, raw2 = http("GET", url)
176
+ check("重复取图字节一致", raw2 == raw)
177
+
178
+ # ---------- 6. 体积兜底 + 缓存键对齐 ----------
179
+ # 真造一条 >50KB 的公式得 1000+ 字符,会先被 _MAX_LATEX_LEN(512) 拦掉;
180
+ # 这里临时把体积上限压到 4KB(该公式 scale=3 约 7.1KB、scale=2 约 3.1KB),
181
+ # 等价地逼出「降 scale 重渲染」这条路。
182
+ print("6) 体积兜底(超过体积上限自动降 scale,下发 id 必须直接命中缓存)")
183
+ old_max_bytes = F._MAX_BYTES
184
+ F._MAX_BYTES = 4 * 1024
185
+ try:
186
+ big = "\\sum_{i=0}^{n} \\frac{x^{i}}{i!} + \\int_{0}^{\\infty} e^{-t} t^{x-1} dt"
187
+ _st, bbody = post_json("/v1/xtc/tex/ensure",
188
+ {"fs": 12, "scale": 3, "items": [{"latex": big, "key": "big"}]},
189
+ {"x-xtc-access-key": KEY})
190
+ finally:
191
+ F._MAX_BYTES = old_max_bytes
192
+ big_items = bbody.get("items") or []
193
+ check("被判定超体积的公式仍渲染成功", len(big_items) == 1)
194
+ if big_items:
195
+ big_scale = int(qs(big_items[0]["url"])["scale"])
196
+ check("确实降了 scale", big_scale < 3, "scale=%d" % big_scale)
197
+ used_before = TEX_API._window_used
198
+ st, _h, rb = http("GET", big_items[0]["url"])
199
+ check("降级图取图 200 且 <=50KB", st == 200 and len(rb) <= 50 * 1024, "%d 字节" % len(rb))
200
+ check("命中缓存、没再消耗限流额度", TEX_API._window_used == used_before,
201
+ "%d -> %d" % (used_before, TEX_API._window_used))
202
+
203
+ # ---------- 7. 防滥用 ----------
204
+ # 注意:LRU 命中时不校验 f(id 本身就是 capability,端上取图就是这个语义),
205
+ # 所以这里必须拿一个「从没渲染过」的 id 去撞校验分支。
206
+ print("7) 防滥用")
207
+ from urllib.parse import quote
208
+ never = "\\gamma_{7} = \\delta_{7} \\cdot \\theta_{7}"
209
+ never_id = F.formula_id(never, 12, 3, F.DEFAULT_FG)
210
+ st, _h, _b = http("GET", "/v1/xtc/tex/%s.png?f=%s&fs=12&scale=3&fg=%s" % (
211
+ never_id, quote("\\rho_{9} = \\sigma_{9}", safe=""), quote(F.DEFAULT_FG, safe="")))
212
+ check("id 与内容不一致 → 403", st == 403, "got %d" % st)
213
+ st, _h, _b = http("GET", "/v1/xtc/tex/0123456789abcdef0123.png")
214
+ check("未知 id 且无 f → 404", st == 404, "got %d" % st)
215
+
216
+ old_budget, old_used, old_start = TEX_API._RENDER_BUDGET, TEX_API._window_used, TEX_API._window_start
217
+ TEX_API._RENDER_BUDGET = 1
218
+ TEX_API._window_used = 0
219
+ TEX_API._window_start = time.time()
220
+ try:
221
+ def fresh_url(i):
222
+ from urllib.parse import quote
223
+ tex = "\\alpha_{%d} = \\beta_{%d}" % (900 + i, 900 + i)
224
+ tid = F.formula_id(tex, 12, 3, F.DEFAULT_FG)
225
+ return "/v1/xtc/tex/%s.png?f=%s&fs=12&scale=3&fg=%s" % (
226
+ tid, quote(tex, safe=""), quote(F.DEFAULT_FG, safe=""))
227
+
228
+ a = http("GET", fresh_url(1))[0]
229
+ b = http("GET", fresh_url(2))[0]
230
+ check("限流:第 1 个放行 / 第 2 个 429", a == 200 and b == 429, "%d / %d" % (a, b))
231
+ finally:
232
+ TEX_API._RENDER_BUDGET, TEX_API._window_used, TEX_API._window_start = old_budget, 0, old_start
233
+
234
+ # ---------- 8. 参数校验 ----------
235
+ print("8) 参数校验")
236
+ st, _ = post_json("/v1/xtc/tex/ensure", {"fs": 99, "items": []}, {"x-xtc-access-key": KEY})
237
+ check("fs 越界 → 400", st == 400, "got %d" % st)
238
+ st, _ = post_json("/v1/xtc/tex/ensure", {"nope": 1}, {"x-xtc-access-key": KEY})
239
+ check("缺 items → 400", st == 400, "got %d" % st)
240
+ st, _h, _b = http("GET", "/v1/xtc/tex/%s.png" % got[0]["id"])
241
+ check("有 id 时不需要 f(LRU 命中直接返回)", st == 200, "got %d" % st)
242
+ finally:
243
+ srv.should_exit = True
244
+ time.sleep(0.3)
245
+
246
+ print("\n合计:pass %d / fail %d" % (PASS, FAIL))
247
+ return 1 if FAIL else 0
248
+
249
+
250
+ if __name__ == "__main__":
251
+ raise SystemExit(main())
app/api/pseudo_stream.py CHANGED
@@ -23,7 +23,6 @@ from ..media.imagefix import (
23
  from ..providers.policy import is_image_model
24
  from ..services import pseudo_store
25
  from ..services.config_store import load_config
26
- from ..services.formula import build_texs
27
  from ..utils.thought import extract_thought_and_answer
28
  from ._common import (
29
  CORS_HEADERS,
@@ -256,11 +255,6 @@ async def _run_pseudo_session(
256
  streamed_thought = "".join(thought_parts)
257
  tag_thought, text = extract_thought_and_answer(full_text)
258
  thought = streamed_thought or tag_thought
259
- # 公式渲染:把回复里的 $...$ 公式在服务端渲染成小尺寸 PNG,只回「公式清单」
260
- # (正文 text 保持不变,端上按出现顺序切分正文即可还原图文混排)。
261
- # 放线程池:mathtext 解析是 CPU 密集,直接在事件循环里跑会阻塞其他请求。
262
- # 失败/依赖缺失返回空清单,端上按纯文本渲染,不影响主链路。
263
- texs = await asyncio.to_thread(build_texs, text)
264
  # 一次性写入完整回复 + emit done。前端轮询到 done 直接读 body.text。
265
  pseudo_store.update_session_state(
266
  session_id,
@@ -270,7 +264,6 @@ async def _run_pseudo_session(
270
  )
271
  pseudo_store.append_event(
272
  session_id, type="done", delta="", finish_reason=last_finish or "stop",
273
- extra={"texs": texs} if texs else None,
274
  )
275
  record_usage(
276
  access_key=access_key, provider=provider.id, model=clean_model,
 
23
  from ..providers.policy import is_image_model
24
  from ..services import pseudo_store
25
  from ..services.config_store import load_config
 
26
  from ..utils.thought import extract_thought_and_answer
27
  from ._common import (
28
  CORS_HEADERS,
 
255
  streamed_thought = "".join(thought_parts)
256
  tag_thought, text = extract_thought_and_answer(full_text)
257
  thought = streamed_thought or tag_thought
 
 
 
 
 
258
  # 一次性写入完整回复 + emit done。前端轮询到 done 直接读 body.text。
259
  pseudo_store.update_session_state(
260
  session_id,
 
264
  )
265
  pseudo_store.append_event(
266
  session_id, type="done", delta="", finish_reason=last_finish or "stop",
 
267
  )
268
  record_usage(
269
  access_key=access_key, provider=provider.id, model=clean_model,
app/api/tex.py CHANGED
@@ -14,11 +14,12 @@ from __future__ import annotations
14
  import logging
15
  import time
16
 
17
- from fastapi import APIRouter, Query, Response
18
 
 
19
  from ..errors import HttpError
20
  from ..services import formula
21
- from ._common import CORS_HEADERS
22
 
23
  logger = logging.getLogger(__name__)
24
 
@@ -42,6 +43,40 @@ def _allow_render() -> bool:
42
  return True
43
 
44
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45
  @router.get("/{tex_id}.png")
46
  async def get_tex_png(
47
  tex_id: str,
 
14
  import logging
15
  import time
16
 
17
+ from fastapi import APIRouter, Depends, Query, Request, Response
18
 
19
+ from ..auth import require_access_key
20
  from ..errors import HttpError
21
  from ..services import formula
22
+ from ._common import CORS_HEADERS, ok_with_cors, read_request_body
23
 
24
  logger = logging.getLogger(__name__)
25
 
 
43
  return True
44
 
45
 
46
+ @router.post("/ensure")
47
+ async def ensure_texs(
48
+ request: Request,
49
+ _key: str = Depends(require_access_key),
50
+ ) -> dict:
51
+ """端上「前端识别公式 → 请求生成图片」的批量入口。
52
+
53
+ 请求:{"items": [{"latex": "...", "key": "端上自算的标识"}]}
54
+ 返回:{"items": [{"key", "id", "url", "w", "h"}]}(渲染失败的条目不下发,
55
+ 端上按「该处保留 LaTeX 原文」降级处理)。
56
+
57
+ 端上先查本地缓存(文件名 t_<id>.png):命中的公式不会进这个请求,
58
+ 只有确实缺图的才来渲染,渲染结果按内容寻址进进程内 LRU。
59
+ """
60
+ body = await read_request_body(request)
61
+ items = body.get("items") if isinstance(body, dict) else None
62
+ if not isinstance(items, list):
63
+ raise HttpError("items is required", status=400, code="bad_request")
64
+ try:
65
+ fs = int(body.get("fs") or formula.DEFAULT_FS)
66
+ scale = int(body.get("scale") or formula.DEFAULT_SCALE)
67
+ except (TypeError, ValueError):
68
+ raise HttpError("invalid fs/scale", status=400, code="bad_request")
69
+ if fs < 8 or fs > 48 or scale < 1 or scale > 4:
70
+ raise HttpError("fs/scale out of range", status=400, code="bad_request")
71
+ fg = str(body.get("fg") or formula.DEFAULT_FG)[:16]
72
+
73
+ # 渲染是 CPU 密集:放线程池,避免阻塞事件循环(mathtext 解析 + PIL 编码)
74
+ import asyncio
75
+
76
+ rendered = await asyncio.to_thread(formula.ensure_batch, items, fs, scale, fg)
77
+ return ok_with_cors({"items": rendered})
78
+
79
+
80
  @router.get("/{tex_id}.png")
81
  async def get_tex_png(
82
  tex_id: str,
app/api/xtc.py CHANGED
@@ -27,7 +27,6 @@ from ..media.imagefix import (
27
  )
28
  from ..providers.policy import is_image_model, is_model_allowed
29
  from ..services.config_store import load_config
30
- from ..services.formula import build_texs
31
  from ..utils.sse import sse_data, sse_done
32
  from ..utils.thought import (
33
  extract_thought_and_answer,
@@ -319,7 +318,7 @@ async def xtc_chat(
319
  usage=(openai_resp or {}).get("usage"),
320
  ok=True,
321
  )
322
- return await _build_xtc_response(
323
  openai_resp,
324
  provider=provider.id,
325
  model=clean_model,
@@ -358,7 +357,7 @@ def _build_upstream_body(body: dict, clean_model: str, messages: list) -> dict:
358
  return out
359
 
360
 
361
- async def _build_xtc_response(
362
  openai_resp: dict,
363
  *,
364
  provider: str,
@@ -397,11 +396,6 @@ async def _build_xtc_response(
397
  "usage": openai_resp.get("usage") or {},
398
  "finish_reason": choice.get("finish_reason"),
399
  }
400
- # 公式渲染:只回公式清单(正文 text 不变),端上按出现顺序切分正文还原图文混排。
401
- # 放线程池避免 mathtext 阻塞事件循环;失败/依赖缺失返回空清单。
402
- texs = await asyncio.to_thread(build_texs, text)
403
- if texs:
404
- payload["texs"] = texs
405
  if image_fix_mode:
406
  payload["image_fix_mode"] = image_fix_mode
407
  payload["image_fixed"] = image_fixed
 
27
  )
28
  from ..providers.policy import is_image_model, is_model_allowed
29
  from ..services.config_store import load_config
 
30
  from ..utils.sse import sse_data, sse_done
31
  from ..utils.thought import (
32
  extract_thought_and_answer,
 
318
  usage=(openai_resp or {}).get("usage"),
319
  ok=True,
320
  )
321
+ return _build_xtc_response(
322
  openai_resp,
323
  provider=provider.id,
324
  model=clean_model,
 
357
  return out
358
 
359
 
360
+ def _build_xtc_response(
361
  openai_resp: dict,
362
  *,
363
  provider: str,
 
396
  "usage": openai_resp.get("usage") or {},
397
  "finish_reason": choice.get("finish_reason"),
398
  }
 
 
 
 
 
399
  if image_fix_mode:
400
  payload["image_fix_mode"] = image_fix_mode
401
  payload["image_fixed"] = image_fixed
app/services/formula.py CHANGED
@@ -36,33 +36,17 @@ DEFAULT_BG = "#1f2937"
36
 
37
  _MAX_BYTES = 50 * 1024
38
  _MAX_LATEX_LEN = 512 # 单条公式长度上限(防超长输入拖垮解析)
39
- _MAX_FORMULAS = 4 # 单条消息最多渲染几条公式(与端上固定槽位一致)
 
 
 
 
40
  _MAX_PIXELS = 4096 * 4096 # 输出像素总量上限
41
 
42
- # 公式抽取:$$...$$ 优先于 $...$;\(...\) 与 \[...\] 一并支持
43
- _MATH_PATTERNS = (
44
- re.compile(r"\$\$(.+?)\$\$", re.S),
45
- re.compile(r"\\\[(.+?)\\\]", re.S),
46
- re.compile(r"\\\((.+?)\\\)", re.S),
47
- re.compile(r"(?<![\\$])\$([^$\n]+?)\$(?!\$)"),
48
- )
49
-
50
  _rendered: "LRUCache" = LRUCache(maxsize=512)
51
  _render_lock = threading.Lock()
52
  _matplotlib_ok: Optional[bool] = None
53
 
54
- # 行内 $...$ 的合理性过滤:货币/价格写法("$5 和 $10")会配对成假公式。
55
- # 显式分隔符($$、\[、\()不做过滤——那是作者明确标注的公式。
56
- _CJK_RE = re.compile(r"[\u2e80-\u9fff\uff00-\uffef]")
57
- _MATH_HINT_RE = re.compile(r"[\\^_=+\-*/<>()\[\]{}|]")
58
-
59
-
60
- def _plausible_inline(body: str) -> bool:
61
- """行内 $...$ 是否像公式:含数学符号且不含中文,避免把价格写法当公式。"""
62
- if _CJK_RE.search(body):
63
- return False
64
- return bool(_MATH_HINT_RE.search(body))
65
-
66
 
67
  def is_available() -> bool:
68
  """matplotlib 是否可用(导入失败则整个公式功能关闭)。"""
@@ -87,68 +71,200 @@ def formula_id(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg:
87
  return hashlib.sha1(raw.encode("utf-8")).hexdigest()[:20]
88
 
89
 
90
- # ===== 抽取 =====
 
 
91
 
92
- def extract_formulas(text: str) -> list[dict]:
93
- """按出现顺序抽取公式,返回 [{latex, start, end}]。
 
 
 
 
 
 
94
 
95
- 只在「能整条渲染」的前提下才返回:调用方拿到清单后用 render() 兜底,
96
- 渲染失败的条目会被剔除(正文里保留 LaTeX 原文,不影响可读性)。
97
- """
98
- s = str(text or "")
99
- if not s or "$" not in s and "\\(" not in s and "\\[" not in s:
100
- return []
101
- hits: list[tuple[int, int, str]] = []
102
- taken: list[tuple[int, int]] = []
103
-
104
- def overlaps(a: int, b: int) -> bool:
105
- for x, y in taken:
106
- if a < y and x < b:
107
- return True
108
- return False
109
-
110
- for pat in _MATH_PATTERNS:
111
- inline = pat is _MATH_PATTERNS[-1]
112
- for m in pat.finditer(s):
113
- body = (m.group(1) or "").strip()
114
- if not body or len(body) > _MAX_LATEX_LEN:
115
- continue
116
- if inline and not _plausible_inline(body):
117
- continue
118
- a, b = m.start(), m.end()
119
- if overlaps(a, b):
120
- continue
121
- taken.append((a, b))
122
- hits.append((a, b, body))
123
 
124
- hits.sort(key=lambda h: h[0])
125
- return [{"latex": body, "start": a, "end": b} for a, b, body in hits]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
126
 
127
 
128
  # ===== 渲染 =====
129
 
130
- def _render_png(latex: str, fs: int, scale: int, fg: str, bg: str = "") -> tuple[bytes, int, int]:
131
- """渲染一张 PNG,返回 (bytes, w_px, h_px)。"""
 
 
 
 
132
  from matplotlib.figure import Figure
133
  from PIL import Image
134
 
135
  dpi = 72.0 * scale
136
- fig = Figure(figsize=(2, 2), dpi=dpi)
137
- fig.patch.set_alpha(0)
138
- fig.text(0, 0, "$" + latex + "$", fontsize=fs, color=fg)
139
- buf = io.BytesIO()
140
- # pad 1 design px:四周留一点边,端上圆角/描边不至于切到字形
141
- fig.savefig(buf, format="png", dpi=dpi, transparent=True,
142
- bbox_inches="tight", pad_inches=1.0 / 72.0)
143
- fig.clear()
144
- buf.seek(0)
145
-
146
- img = Image.open(buf).convert("RGBA")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
147
  if bg:
148
- out = Image.new("RGB", img.size, bg)
149
- out.paste(img, (0, 0), img)
150
  else:
151
- out = img
152
  out.info = {} # 丢掉元数据
153
  png = io.BytesIO()
154
  out.save(png, format="PNG", compress_level=9)
@@ -175,7 +291,10 @@ def render(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str
175
  if s in tried or s < 1:
176
  continue
177
  tried.append(s)
178
- png, pw, ph = _render_png(body, fs, s, fg)
 
 
 
179
  if pw * ph > _MAX_PIXELS:
180
  return None
181
  if len(png) <= _MAX_BYTES or s == 1:
@@ -188,6 +307,13 @@ def render(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str
188
  }
189
  with _render_lock:
190
  _rendered[key] = info
 
 
 
 
 
 
 
191
  return info
192
  except Exception:
193
  logger.info("[formula] render failed: %s", body[:80], exc_info=False)
@@ -195,41 +321,52 @@ def render(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str
195
  return None
196
 
197
 
198
- # ===== 供聊天响应使用 =====
 
 
 
 
 
199
 
200
- def build_texs(text: str, fs: int = DEFAULT_FS) -> list[dict]:
201
- """把正文里的公式渲染成清单:[{id, latex, url, w, h, tex}](最多 _MAX_FORMULAS 条)。
202
 
203
- url 为相对路径,端上拼 baseUrl;端上用 request.download 拉取后落盘缓存。
204
- 渲染失败的条目被剔除,正文里保留 LaTeX 原文。
 
 
 
 
 
 
 
 
 
205
  """
206
  if not is_available():
207
  return []
208
  out: list[dict] = []
209
- for hit in extract_formulas(text)[:_MAX_FORMULAS]:
210
- info = render(hit["latex"], fs=fs)
 
 
 
 
 
211
  if not info:
212
  continue
 
213
  out.append({
214
- "id": formula_id(hit["latex"], fs, info["scale"], DEFAULT_FG),
215
- "latex": hit["latex"],
216
- "url": "/v1/xtc/tex/" + formula_id(hit["latex"], fs, info["scale"], DEFAULT_FG) + ".png",
 
 
217
  "w": info["w"],
218
  "h": info["h"],
219
- "tex": True,
220
  })
221
  return out
222
 
223
 
224
- def get_png(tex_id: str) -> Optional[bytes]:
225
- """按 id 取已渲染的 PNG(LRU 命中才返回;不再重新渲染,避免公开计算入口)。"""
226
- entry = _rendered.get(str(tex_id or ""))
227
- if not entry:
228
- return None
229
- return entry.get("png")
230
-
231
 
232
- def render_png(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str = DEFAULT_FG) -> Optional[bytes]:
233
- """按公式内容渲染取 PNG(端点调用:内容寻址,可安全地按需渲染)。"""
234
- info = render(latex, fs=fs, scale=scale, fg=fg)
235
- return info.get("png") if info else None
 
36
 
37
  _MAX_BYTES = 50 * 1024
38
  _MAX_LATEX_LEN = 512 # 单条公式长度上限(防超长输入拖垮解析)
39
+ # 单次批量请求最多渲染几条。必须 >= 端上单条消息的公式上限(tex.js MAX_PER_MSG=12),
40
+ # 否则端上一次请求里的多余条目会被这里截掉,端上又拿不到结果 → 那些公式被
41
+ # 永久降级成 LaTeX 原文(旧值 4 就是这个 bug 的根源)。
42
+ # 端上另外做了分批(每批 <= _ENSURE_BATCH),两端上限不一致时也不会静默丢。
43
+ _MAX_FORMULAS = 12
44
  _MAX_PIXELS = 4096 * 4096 # 输出像素总量上限
45
 
 
 
 
 
 
 
 
 
46
  _rendered: "LRUCache" = LRUCache(maxsize=512)
47
  _render_lock = threading.Lock()
48
  _matplotlib_ok: Optional[bool] = None
49
 
 
 
 
 
 
 
 
 
 
 
 
 
50
 
51
  def is_available() -> bool:
52
  """matplotlib 是否可用(导入失败则整个公式功能关闭)。"""
 
71
  return hashlib.sha1(raw.encode("utf-8")).hexdigest()[:20]
72
 
73
 
74
+ # ===== LaTeX 预处理 =====
75
+ # 模型/文档里的公式五花八门,mathtext 只认最基础的数学语法,这里先把常见写法
76
+ # 归一化成 mathtext 可解析的形式(否则整条公式渲染失败、只能降级成文本)。
77
 
78
+ # 文档级命令:不是公式,出现在裸公式块里应当忽略
79
+ _PREAMBLE_CMDS = (
80
+ "usepackage", "documentclass", "begin{document}", "end{document}",
81
+ "title", "author", "date", "maketitle", "tableofcontents", "newpage",
82
+ "clearpage", "centering", "setlength", "geometry", "pagestyle",
83
+ "section", "subsection", "label", "ref", "cite", "includegraphics",
84
+ "hline", "toprule", "midrule", "bottomrule", "today",
85
+ )
86
 
87
+ # 多行环境(\\ 分行、& 分列):去掉环境包裹与对齐符后按行渲染
88
+ _MULTILINE_ENVS = (
89
+ "aligned", "align", "align*", "gathered", "gather", "gather*",
90
+ "split", "cases", "dcases", "eqnarray", "eqnarray*",
91
+ )
92
+ # 不支持的环境(矩阵/表格需要列对齐,mathtext 排不了)→ 直接判定渲染失败
93
+ _UNSUPPORTED_ENVS = ("matrix", "pmatrix", "bmatrix", "vmatrix", "array", "tabular", "table")
94
+
95
+ _re_begin = re.compile(r"\\begin\{([^}]*)\}")
96
+ _re_comment = re.compile(r"(?<!\\)%[^\n]*")
97
+ # 公式正文里的中日韩字符(含全角标点):mathtext 没有中文字形,画出来是一排空心
98
+ # 方块,比退回 LaTeX 原文更难读。端上对行内 $..$ 已经排除中文,但 $$..$$ / \[..\]
99
+ # 没有排除(模型很爱在块级公式里写中文注解),这里兜底判为「不渲染」。
100
+ _re_cjk = re.compile(r"[\u2e80-\u9fff\u3000-\u303f\uff00-\uffef]")
101
+
102
+
103
+ def is_preamble_latex(body: str) -> bool:
104
+ """整块是导言区/文档级命令(不是公式),用于裸公式识别时排除。"""
105
+ s = str(body or "").strip()
106
+ if not s:
107
+ return True
108
+ for cmd in _PREAMBLE_CMDS:
109
+ if ("\\" + cmd) in s or ("\\begin{document}") in s:
110
+ return True
111
+ return False
 
 
 
112
 
113
+
114
+ def _convert_mhchem(body: str) -> str:
115
+ """\\ce{2H2 + O2 -> 2H2O} 这类 mhchem 写法 → mathtext 能排的等价写法。
116
+
117
+ mhchem 是宏包语法,mathtext 不支持,这里只覆盖化学方程式的常见写法:
118
+ 箭头、可逆号、元素后的数字下标、电荷上标、气体/沉淀号。
119
+ """
120
+ b = str(body or "")
121
+ b = b.replace("<=>", " \\rightleftharpoons ").replace("<->", " \\rightleftharpoons ")
122
+ b = b.replace("->", " \\rightarrow ").replace("<-", " \\leftarrow ")
123
+ # 元素符号/括号后的数字 → 下标(H2 → H_{2},NH3 → NH_{3},(g) 不动)
124
+ b = re.sub(r"(?<=[A-Za-z\)\]])([0-9]+)(?![0-9])", r"_{\1}", b)
125
+ # 电荷上标:^2- / ^+ / ^3+ → ^{2-}
126
+ b = re.sub(r"\^([0-9]*[+\-])", r"^{\1}", b)
127
+ # 气体上升号 ^ / 沉淀号 v(mhchem 写法,mathtext 没有这两个记号)→ 箭头。
128
+ # 只认「独立成词」的那个字符:^{14} 的 ^、hv(光量子)里的 v 都不能动。
129
+ b = re.sub(r"(?:^|(?<=\s))\^(?=\s|$)", r" \\uparrow ", b)
130
+ b = re.sub(r"(?:^|(?<=\s))v(?=\s|$)", r" \\downarrow ", b)
131
+ return b
132
+
133
+
134
+ def _replace_ce(s: str) -> str:
135
+ """把 \\ce{...} 换掉。手写扫描而不是正则:\\ce 里会出现嵌套花括号
136
+ (同位素 ^{14}C、\\ce{Fe(SCN)3} 之类),正则 [^{}]* 匹配不到就会原样漏给
137
+ mathtext → 整条公式渲染失败。
138
+ """
139
+ out = []
140
+ i = 0
141
+ while i < len(s):
142
+ j = s.find("\\ce{", i)
143
+ if j < 0:
144
+ out.append(s[i:])
145
+ break
146
+ out.append(s[i:j])
147
+ k = j + 4
148
+ depth = 1
149
+ buf = []
150
+ while k < len(s) and depth > 0:
151
+ ch = s[k]
152
+ if ch == "{":
153
+ depth += 1
154
+ elif ch == "}":
155
+ depth -= 1
156
+ if depth > 0:
157
+ buf.append(ch)
158
+ k += 1
159
+ if depth != 0:
160
+ # 花括号没闭合:原样保留,交给 mathtext 报错(端上退回原文)
161
+ out.append(s[j:])
162
+ break
163
+ out.append(_convert_mhchem("".join(buf)))
164
+ i = k
165
+ return "".join(out)
166
+
167
+
168
+ def normalize_latex(body: str) -> Optional[list]:
169
+ """把一条公式归一化成「待渲染的行列表」;无法渲染返回 None。
170
+
171
+ 返回的每行都是不带 $ 的 mathtext 表达式,多行由调用方按行堆叠渲染。
172
+ """
173
+ s = str(body or "").strip()
174
+ if not s:
175
+ return None
176
+ # 去掉 LaTeX 注释(\% 转义的不算注释)
177
+ s = _re_comment.sub(" ", s)
178
+ # mhchem 化学式
179
+ s = _replace_ce(s)
180
+
181
+ # 环境名处理:多行环境去壳,不支持的环境直接失败
182
+ envs = _re_begin.findall(s)
183
+ for env in envs:
184
+ base = env.strip().rstrip("*")
185
+ if base in _UNSUPPORTED_ENVS:
186
+ return None
187
+ if base not in [e.rstrip("*") for e in _MULTILINE_ENVS]:
188
+ # 未知环境(如 \text、自定义环境)交给 mathtext 自然报错
189
+ pass
190
+ s = re.sub(r"\\begin\{[^}]*\}|\\end\{[^}]*\}", " ", s)
191
+ s = s.replace("&", " ")
192
+
193
+ # 换行:\\ 是 LaTeX 显式换行;源码里的裸换行也按换行处理——手表屏宽只有
194
+ # 240px,长公式横向铺满后必须整体缩小才放得下,按源码本就断行的地方拆开
195
+ # 反而更清楚(每行各自居中堆叠)。
196
+ s = s.replace("\\\\", "\n")
197
+ lines = []
198
+ for raw in s.split("\n"):
199
+ ln = " ".join(raw.split())
200
+ if ln:
201
+ lines.append(ln)
202
+ if not lines:
203
+ return None
204
+ # 含中文 → 不渲染(mathtext 会画成空心方块),端上退回 LaTeX 原文
205
+ for ln in lines:
206
+ if _re_cjk.search(ln):
207
+ return None
208
+ # 全部行都只是导言区命令 → 不是公式
209
+ if all(is_preamble_latex(ln) for ln in lines):
210
+ return None
211
+ return lines
212
 
213
 
214
  # ===== 渲染 =====
215
 
216
+ def _render_lines_rgba(lines: list, fs: int, scale: int, fg: str):
217
+ """按行渲染并垂直堆叠为一张 RGBA 图(多行公式:换行符/aligned/cases)。
218
+
219
+ 逐行单独出图再拼接,避免在 figure 里做坐标换算(bbox_inches='tight'
220
+ 会重新裁剪,figure 的分数坐标无法预知最终尺寸)。行内居中。
221
+ """
222
  from matplotlib.figure import Figure
223
  from PIL import Image
224
 
225
  dpi = 72.0 * scale
226
+ parts = []
227
+ for ln in lines:
228
+ fig = Figure(figsize=(2, 2), dpi=dpi)
229
+ fig.patch.set_alpha(0)
230
+ fig.text(0, 0, "$" + ln + "$", fontsize=fs, color=fg)
231
+ buf = io.BytesIO()
232
+ fig.savefig(buf, format="png", dpi=dpi, transparent=True,
233
+ bbox_inches="tight", pad_inches=0.0)
234
+ fig.clear()
235
+ buf.seek(0)
236
+ parts.append(Image.open(buf).convert("RGBA"))
237
+
238
+ gap = max(1, int(round(fs * scale * 0.35)))
239
+ width = max(p.size[0] for p in parts)
240
+ height = sum(p.size[1] for p in parts) + gap * (len(parts) - 1)
241
+ out = Image.new("RGBA", (width, height), (0, 0, 0, 0))
242
+ y = 0
243
+ for p in parts:
244
+ out.paste(p, ((width - p.size[0]) // 2, y), p)
245
+ y += p.size[1] + gap
246
+ return out
247
+
248
+
249
+ def _render_png(body: str, fs: int, scale: int, fg: str, bg: str = "") -> Optional[tuple]:
250
+ """渲染一条公式(可多行),返回 (bytes, w_px, h_px);无法渲染返回 None。"""
251
+ from PIL import Image
252
+
253
+ lines = normalize_latex(body)
254
+ if not lines:
255
+ return None
256
+
257
+ img = _render_lines_rgba(lines, fs, scale, fg)
258
+ # 四周补 1 design px 边距(避免端上圆角/描边切到字形或根号上划线)
259
+ pad = max(1, int(round(scale)))
260
+ canvas = Image.new("RGBA", (img.size[0] + pad * 2, img.size[1] + pad * 2), (0, 0, 0, 0))
261
+ canvas.paste(img, (pad, pad), img)
262
+
263
  if bg:
264
+ out = Image.new("RGB", canvas.size, bg)
265
+ out.paste(canvas, (0, 0), canvas)
266
  else:
267
+ out = canvas
268
  out.info = {} # 丢掉元数据
269
  png = io.BytesIO()
270
  out.save(png, format="PNG", compress_level=9)
 
291
  if s in tried or s < 1:
292
  continue
293
  tried.append(s)
294
+ got = _render_png(body, fs, s, fg)
295
+ if not got:
296
+ return None
297
+ png, pw, ph = got
298
  if pw * ph > _MAX_PIXELS:
299
  return None
300
  if len(png) <= _MAX_BYTES or s == 1:
 
307
  }
308
  with _render_lock:
309
  _rendered[key] = info
310
+ # 体积兜底降级后,下发给端上的 id 用的是「实际 scale」算出来的
311
+ # (见 ensure_batch),端上之后就拿这个 id 回源取图。这里同时按
312
+ # 实际 scale 的 id 入缓存,否则 /v1/xtc/tex/{id}.png 每次都 LRU
313
+ # 未命中,白白重渲染一遍(限流额度也被重复消耗)。
314
+ actual = formula_id(body, fs, s, fg)
315
+ if actual != key:
316
+ _rendered[actual] = info
317
  return info
318
  except Exception:
319
  logger.info("[formula] render failed: %s", body[:80], exc_info=False)
 
321
  return None
322
 
323
 
324
+ def get_png(tex_id: str) -> Optional[bytes]:
325
+ """按 id 取已渲染的 PNG(LRU 命中才返回),未命中返回 None。"""
326
+ entry = _rendered.get(str(tex_id or ""))
327
+ if not entry:
328
+ return None
329
+ return entry.get("png")
330
 
 
 
331
 
332
+ def render_png(latex: str, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str = DEFAULT_FG) -> Optional[bytes]:
333
+ """按公式内容渲染取 PNG(端点调用:内容寻址,可安全地按需渲染)。"""
334
+ info = render(latex, fs=fs, scale=scale, fg=fg)
335
+ return info.get("png") if info else None
336
+
337
+
338
+ def ensure_batch(items: list, fs: int = DEFAULT_FS, scale: int = DEFAULT_SCALE, fg: str = DEFAULT_FG) -> list:
339
+ """端上「前端识别 → 请求生成」入口:批量渲染公式,返回可下载清单。
340
+
341
+ items: [{latex, key}](key 由端上自算,用于把结果对回原文里的公式)
342
+ 返回: [{key, id, url, w, h}],渲染失败的条目被剔除并由端上降级为纯文本。
343
  """
344
  if not is_available():
345
  return []
346
  out: list[dict] = []
347
+ for it in (items or [])[:_MAX_FORMULAS]:
348
+ if not isinstance(it, dict):
349
+ continue
350
+ latex = str(it.get("latex") or "").strip()
351
+ if not latex or len(latex) > _MAX_LATEX_LEN:
352
+ continue
353
+ info = render(latex, fs=fs, scale=scale, fg=fg)
354
  if not info:
355
  continue
356
+ tid = formula_id(latex, fs, info["scale"], fg)
357
  out.append({
358
+ "key": str(it.get("key") or ""),
359
+ "id": tid,
360
+ # 带 f=:Space 重启后进程内 LRU 丢失,端上仍能按需取到同一张图
361
+ "url": "/v1/xtc/tex/%s.png?f=%s&fs=%d&scale=%d&fg=%s" % (
362
+ tid, _urlquote(latex), int(fs), int(info["scale"]), _urlquote(fg)),
363
  "w": info["w"],
364
  "h": info["h"],
 
365
  })
366
  return out
367
 
368
 
369
+ def _urlquote(s: str) -> str:
370
+ from urllib.parse import quote
 
 
 
 
 
371
 
372
+ return quote(str(s or ""), safe="")