fix: route ai polish through available models

2026-06-04 10:13:31 +08:00
parent 3f5dfdc465
commit 88d598303c
7 changed files with 67 additions and 23 deletions
--- a/RULES.md
+++ b/RULES.md
@@ -150,7 +150,8 @@
 - `LOCAL_ASR_BIN` / `LOCAL_ASR_MODEL` / `LOCAL_ASR_TIMEOUT_SECONDS`：本机 ASR 兜底，默认使用 `/opt/homebrew/bin/mlx_whisper` + `mlx-community/whisper-tiny`，用于当前 SKG 网关 `/audio/transcriptions` 不可用时生成真实逐句时间轴
 - `TRANSLATE_MODEL`：字幕翻译模型，默认 `gemini-2.5-flash`
 - `GPT_TEXT_MODEL`：GPT 文本 / 视觉默认模型，默认 `gpt-4o`；用于兜底修正旧 Gemini 覆盖值
- `REWRITE_MODEL`：通用改写/分镜描述模型，默认 `gpt-4o`；如果旧环境仍写 `gemini-*`，后端会自动改用 `GPT_TEXT_MODEL`
+- `REWRITE_MODEL`：通用改写/分镜描述主模型，当前用于 AI 润色时默认 `gpt-4o-mini`；如果主模型不可用，`/prompt/polish` 会继续尝试 `REWRITE_MODEL_FALLBACKS`
+- `REWRITE_MODEL_FALLBACKS`：AI 润色备用模型列表，逗号分隔，默认 `gpt-4o-mini,gemini-2.5-flash`；只有全部失败时才允许返回本地模板 fallback
 - `VISION_MODEL`：关键帧画面理解模型，默认 `gpt-4o`；如果旧环境仍写 `gemini-*`，后端会自动改用 `GPT_TEXT_MODEL`
 - `AUDIO_REWRITE_MODEL`：后续音频口播改写模型，默认跟随 `REWRITE_MODEL`；如果旧环境仍写 `gemini-*`，后端会自动改用 `REWRITE_MODEL`
 - `AUDIO_PRODUCT_BRIEF`：音频口播改写时注入的 SKG 产品卖点
--- a/api/.env.example
+++ b/api/.env.example
@@ -30,7 +30,8 @@ LOCAL_ASR_MODEL=mlx-community/whisper-tiny
 LOCAL_ASR_TIMEOUT_SECONDS=180
 TRANSLATE_MODEL=gemini-2.5-flash
 GPT_TEXT_MODEL=gpt-4o
-REWRITE_MODEL=gpt-4o
+REWRITE_MODEL=gpt-4o-mini
+REWRITE_MODEL_FALLBACKS=gemini-2.5-flash
 VISION_MODEL=gpt-4o
 PRODUCT_VIEW_MODEL=gpt-image-2
 IMAGE_BASE_URL=https://ai.skg.com/ezlink/v1
--- a/api/main.py
+++ b/api/main.py
@@ -107,6 +107,11 @@ def gpt_model_env(name: str, default: str | None = None) -> str:


 REWRITE_MODEL = gpt_model_env("REWRITE_MODEL")
+REWRITE_MODEL_FALLBACKS = [
+    model.strip()
+    for model in os.getenv("REWRITE_MODEL_FALLBACKS", "gpt-4o-mini,gemini-2.5-flash").split(",")
+    if model.strip()
+]
 VISION_MODEL = gpt_model_env("VISION_MODEL")
 IMAGE_BASE_URL = os.getenv("IMAGE_BASE_URL", LLM_BASE_URL).strip()
 IMAGE_API_KEY = os.getenv("IMAGE_API_KEY", LLM_API_KEY).strip()
@@ -5914,6 +5919,7 @@ class CreativeCopyResp(BaseModel):
 class PromptPolishReq(BaseModel):
    text: str
    system_prompt: str = ""
+    model: str = ""
    mode: Literal["image", "video", "general", "chat"] = "image"
    target_language: Literal["en", "zh", "keep"] = "en"

@@ -6320,11 +6326,26 @@ def _prompt_polish_fallback(req: PromptPolishReq) -> PromptPolishResp:
    return PromptPolishResp(model="fallback", text=_sanitize_polished_prompt(req, intent, _basic_polished_prompt(req, intent)))


-def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output: str, *, allow_llm: bool = False) -> str:
+def _prompt_polish_model_candidates(req: PromptPolishReq) -> list[str]:
+    requested = (req.model or "").strip()
+    candidates = [requested, REWRITE_MODEL, *REWRITE_MODEL_FALLBACKS]
+    seen: set[str] = set()
+    out: list[str] = []
+    for model in candidates:
+        model = model.strip()
+        key = model.lower()
+        if model and key not in seen:
+            out.append(model)
+            seen.add(key)
+    return out
+
+
+def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output: str, *, allow_llm: bool = False, model: str | None = None) -> str:
    out = _sanitize_polished_prompt(req, intent, output)
    issue = _polished_prompt_issue(intent, out)
    if not issue or not allow_llm or not LLM_API_KEY:
        return out
+    repair_model = (model or REWRITE_MODEL).strip() or REWRITE_MODEL
    repair_prompt = (
        "Repair the rewritten generation prompt so it follows the source input exactly.\n"
        f"Issue to fix: {issue}.\n"
@@ -6340,7 +6361,7 @@ def _repair_polished_prompt(req: PromptPolishReq, intent: PromptIntent, output:
    )
    try:
        resp = llm().chat.completions.create(
-            model=REWRITE_MODEL,
+            model=repair_model,
            messages=[
                {"role": "system", "content": "You repair generation prompts by removing contradictions and preserving only source intent."},
                {"role": "user", "content": repair_prompt},
@@ -6410,23 +6431,28 @@ def polish_prompt(req: PromptPolishReq) -> PromptPolishResp:
        prompt += f"\nUser-selected polishing guidance:\n{user_system[:1000]}\n"
    prompt += f"\nSource input:\n{intent.cleaned_text[:2500]}"

-    try:
-        resp = llm().chat.completions.create(
-            model=REWRITE_MODEL,
-            messages=[
-                {"role": "system", "content": "You are a neutral professional prompt editor. Preserve source intent exactly and never inject SKG or unrelated brands, products, platforms, people, or marketing context."},
-                {"role": "user", "content": prompt},
-            ],
-            temperature=0.45,
-            max_tokens=900,
-        )
-        out = _clean_prompt_output(resp.choices[0].message.content or "")
-        if not out:
-            out = _prompt_polish_fallback(req).text
-        return PromptPolishResp(model=REWRITE_MODEL, text=_repair_polished_prompt(req, intent, out, allow_llm=True))
-    except Exception as e:
-        print(f"[prompt polish fallback] {e}", flush=True)
-        return _prompt_polish_fallback(req)
+    model_errors: list[str] = []
+    for model in _prompt_polish_model_candidates(req):
+        try:
+            resp = llm().chat.completions.create(
+                model=model,
+                messages=[
+                    {"role": "system", "content": "You are a neutral professional prompt editor. Preserve source intent exactly and never inject SKG or unrelated brands, products, platforms, people, or marketing context."},
+                    {"role": "user", "content": prompt},
+                ],
+                temperature=0.45,
+                max_tokens=900,
+            )
+            out = _clean_prompt_output(resp.choices[0].message.content or "")
+            if not out:
+                raise RuntimeError("empty prompt polish response")
+            return PromptPolishResp(model=model, text=_repair_polished_prompt(req, intent, out, allow_llm=True, model=model))
+        except Exception as e:
+            message = str(e).replace("\n", " ")[:400]
+            model_errors.append(f"{model}: {message}")
+            print(f"[prompt polish model fallback] model={model} error={message}", flush=True)
+    print(f"[prompt polish fallback] {' | '.join(model_errors)}", flush=True)
+    return _prompt_polish_fallback(req)


@app.post("/translate")
--- a/deploy/.env.local.example
+++ b/deploy/.env.local.example
@@ -54,7 +54,8 @@ AI_HTTP_PROXY=

 # Text/vision/audio model names
 GPT_TEXT_MODEL=gpt-4o
-REWRITE_MODEL=gpt-4o
+REWRITE_MODEL=gpt-4o-mini
+REWRITE_MODEL_FALLBACKS=gemini-2.5-flash
 VISION_MODEL=gpt-4o
 TRANSLATE_MODEL=gemini-2.5-flash
 ASR_BASE_URL=https://ai.skg.com/azure/v1
--- a/deploy/.env.production.example
+++ b/deploy/.env.production.example
@@ -56,7 +56,8 @@ FASTER_WHISPER_DEVICE=cpu
 FASTER_WHISPER_COMPUTE_TYPE=int8
 TRANSLATE_MODEL=gemini-2.5-flash
 GPT_TEXT_MODEL=gpt-4o
-REWRITE_MODEL=gpt-4o
+REWRITE_MODEL=gpt-4o-mini
+REWRITE_MODEL_FALLBACKS=gemini-2.5-flash
 VISION_MODEL=gpt-4o
 PRODUCT_VIEW_MODEL=gpt-image-2
 IMAGE_BASE_URL=https://ai.skg.com/ezlink/v1
--- a/docs/source-analysis.html
+++ b/docs/source-analysis.html
@@ -1417,6 +1417,19 @@ ProductRefStateItem {
              <p><strong>影响：</strong>Postgres 里的 <code>canvas_projects</code> 重新成为主存储；刷新、换浏览器或本地缓存异常时，不应再把服务端项目缩小或清空。旧项目首次迁移仍可用，但迁移动作变为非破坏性。</p>
            </div>
          </article>
+          <article class="change">
+            <header>
+              <h3>2026-06-04 · AI 润色不再静默套模板</h3>
+              <span class="tag amber">API</span>
+              <span class="tag violet">Canvas</span>
+              <span class="tag cyan">Model</span>
+            </header>
+            <div class="body">
+              <p><strong>问题：</strong>当前网关分组对 <code>gpt-4o</code> 返回“无可用渠道”，而 <code>/prompt/polish</code> 捕获异常后直接返回本地 <code>fallback</code>，用户看到的是固定尾巴模板，不是真正的模型润色；同时前端 <code>useChat({ model: 'gpt-4o-mini' })</code> 没有把 <code>model</code> 发给后端，配置实际上未生效。</p>
+              <p><strong>改动：</strong><code>PromptPolishReq</code> 新增 <code>model</code> 字段，<code>web/canvas-app/src/hooks/useApi.js</code> 会把前端选择的模型传到 <code>/prompt/polish</code>；后端按“请求模型 → <code>REWRITE_MODEL</code> → <code>REWRITE_MODEL_FALLBACKS</code>”依次尝试，当前本地默认 <code>REWRITE_MODEL=gpt-4o-mini</code>、备用 <code>gemini-2.5-flash</code>。只有全部模型失败时才返回本地模板 <code>model=fallback</code>，并在日志里记录每个失败模型。</p>
+              <p><strong>影响：</strong>画布底部和文本节点的 AI 润色会优先走真实模型输出，不再把固定 “Clear main subject...” 或 “Cinematic motion...” 当作正常润色结果；如果未来网关主模型不可用，接口会自动降级到备用模型，而不是立刻套模板。</p>
+            </div>
+          </article>
          <article class="change">
            <header>
              <h3>2026-05-26 · AI 润色改为意图分类和冲突校验</h3>
--- a/web/canvas-app/src/hooks/useApi.js
+++ b/web/canvas-app/src/hooks/useApi.js
@@ -200,6 +200,7 @@ export const useChat = (options = {}) => {
        body: JSON.stringify({
          text: content,
          system_prompt: options.systemPrompt || '',
+          model: options.model || '',
          mode,
          target_language: options.targetLanguage || (mode === 'chat' ? 'keep' : 'en')
        })