Auto-sync from GitHub Actions

mcp_server.py

@@ -50,147 +50,61 @@ mcp = FastMCP(
 @mcp.tool()
 async def search_tags(
     query: str,
-    use_segmentation: bool = True,
-    top_k: int = 5,
-    limit: int = 80,
-    popularity_weight: float = 0.15,
+    search_mode: str = "full_scene",
+    category: str = "all",
     show_nsfw: bool = True,
     include_wiki: bool = False,
-    category: str = "all",
-    group_mode: str = "off",
-    max_per_group: int = 2,
 ) -> str:
     """
 Search Danbooru tags using natural language and return a ready-to-use prompt.
+Only supported for general, copyright, and character tag searches; **artists and meta tags are not supported.**
 
 ## Args
 - query: Natural language description (Chinese recommended).
-- use_segmentation: Split multi-concept input into segments for separate retrieval. True for scene descriptions, False for single-concept queries.
-- top_k: Candidates recalled per segment. Semantics change with use_segmentation — see guide below.
-- limit: Max tags returned.
-- popularity_weight: Influence of tag post count on ranking (0.0–1.0). Default 0.15.
+- search_mode: Preset strategy. Pick the one that matches your intent.
+    "full_scene"       — Full scene → prompt (e.g. "一个穿着白色水手服的少女在雨中奔跑")
+    "concept_explore"  — Vague concept exploration, broad recall (e.g. "赛博朋克服装", "兔耳朵", "中国风汉服")
+    "subject_describe" — Describe a subject to find matching tags (e.g. "EVA中蓝发的驾驶员", "两侧有开口，前方有拉绳的运动短裤")
+    "precise_lookup"   — Precise lookup / spell fix (e.g. "selafuku", "thighhigh")
+- category: Filter to a specific tag category. Default "all".
+    "all"       — All (通用 + 版权 + 人物)
+    "general"   — Visual attributes, clothing, pose, background, etc.
+    "character" — Named characters from any series
+    "copyright" — Specific anime/game/franchise titles
 - show_nsfw: Include NSFW tags. Default True.
 - include_wiki: Append wiki description to each result. Default False.
-- category: Filter results to a specific tag category. Default "all".
-    "all"       —  All categories (通用 + 版权 + 人物 )
-    "general"   —  General: visual attributes, clothing, pose, background, etc.
-    "copyright" —  Copyright: specific anime/game/franchise titles
-    "character" —  Character: named characters from any series
-    Use this when you know what kind of tag you need — e.g. looking for a
-    character name vs. describing a scene visually.
-- group_mode: Tag group processing mode. Default "off".
-    "off"     — No group processing (backward compatible)
-    "expand"  — Boost same-group tags for concept exploration
-    "diverse" — Limit tags per group for scene diversity
-- max_per_group: Max tags per group in diverse mode. Default 2.
+    Set True when tags are unfamiliar and need disambiguation.
 
 ## Query writing guide
 
-The `query` parameter supports explicit delimiter control for precise segmentation.
-
-### Explicit delimiters
-
 Use **spaces, newlines, Chinese commas (，), or Chinese dunhao (、)** to manually separate concepts.
 Each delimiter-bounded segment ≤7 characters stays atomic — the engine respects your intent.
 
-| Query style | Example | When to use |
-|---|---|---|
-| Concept list (spaces) | `运动社团 校队 比赛 运动会` | You know the exact concepts to search |
-| Concept list (dun hao) | `反乌托邦、赛博朋克、蒸汽朋克` | Same, with Chinese list punctuation |
-| Natural sentence | `一个穿着白色水手服的少女在雨中奔跑` | Scene description, let the engine auto-split |
-| Mixed | `运动社团 一个穿水手服的少女` | Mix concepts with descriptive phrases |
-
-Segments >7 characters are still auto-split by jieba, but the raw segment is kept as an additional query
-to preserve clause-level semantics (multi-granularity retrieval).
-
-### Recommendations
-
-1. **Concept lists → use explicit delimiters:** Group independent concepts with spaces or dunhao.
-   `运动社团 校队 比赛 体育祭 田径部` is better than `运动社团校队比赛体育祭田径部`.
-
-2. **Scene descriptions → write naturally:** Natural Chinese with Chinese commas works well for full scenes.
-   `一个穿着白色水手服，蓝色短裙的少女在雨中奔跑` — commas here are grammatical, not delimiters.
-
-3. **Precise lookup → turn off segmentation:** For finding a specific character or copyright title,
-   set `use_segmentation=False` and combine with `category` filter.
-   e.g. `query="EVA中蓝发的零号机驾驶员"` with `category="character"` and `use_segmentation=False`.
-
-4. **Category filtering:** Use `category` to narrow results. Looking for a character?
-   `category="character"`. Building a scene prompt? `category="general"`.
-
-## Parameter guide
-
-### Step 1 — Decide use_segmentation + top_k together
-
-top_k means "candidates per segment"; its effect depends on whether segmentation is on.
-
-Multi-concept input (scene description) → use_segmentation=True
-
-| Sub-scenario          | top_k | Reason                                                   |
-|-----------------------|-------|----------------------------------------------------------|
-| Full scene → prompt   | 5     | Many segments; low top_k distributes result slots fairly |
-| Vague concept explore | 80    | Few segments; high top_k needed for broad recall         |
-
-Single-concept input → use_segmentation=False
-
-top_k acts as total candidate pool size. Use 20 for all single-concept cases.
+| Query style | Example |
+|---|---|
+| Concept list (spaces) | `运动社团 校队 比赛 运动会` |
+| Concept list (dun hao) | `反乌托邦、赛博朋克、蒸汽朋克` |
+| Natural sentence | `一个穿着白色水手服的少女在雨中奔跑` |
+| Mixed | `运动社团 一个穿水手服的少女` |
 
-| Sub-scenario                  | top_k |
-|-------------------------------|-------|
-| Describe subject / find tag   | 20    |
-| Precise lookup / spell fix    | 20    |
-
-### Step 2 — Decide limit independently
-
-| Goal                         | limit |
-|------------------------------|-------|
-| Full prompt for image gen    | 80    |
-| Concept exploration          | 20–80 |
-| Precise lookup / role search | 10–20 |
-
-### Auxiliary params
-
-popularity_weight (default 0.15, rarely needs changing):
-- Higher (0.3+): favor common, well-established tags
-- Lower (0.0): surface niche/rare tags
-
-include_wiki (default False):
-- True: The meaning of the tag is important — disambiguation, explaining tags to users, exploring unfamiliar domains, or when you are unsure of the tag's meaning
-- False: Prompt generation (Wiki is irrelevant to the downstream task), tags are known
-
-### Quick reference
-
-| Scenario                      | use_segmentation | top_k | limit |
-|-------------------------------|------------------|-------|-------|
-| Full scene → prompt (default) | True             | 5     | 80    |
-| Vague concept exploration     | True             | 80    | 80    |
-| Describe subject / find tag   | False            | 20    | 20    |
-| Precise lookup / spell fix    | False            | 20    | 10    |
-
-### Workflow
-
-After search_tags, pass selected tags to get_related_tags to discover complementary tags via co-occurrence (accessories, character features, scene atmosphere).
-Supports chained exploration / iterative loops – take the interesting tags from the returned results as input to call get_related_tags again,
-and use the results from get_related to feed back into a new round of search, enabling multi-hop deep traversal along the co-occurrence graph.
-
-## Examples
-
-Precise lookup / spell fix — e.g. "selafuku", "thighhigh", "twintail"
-→ use_segmentation=False, top_k=20, limit=10
-
-Vague concept exploration — e.g. "兔耳朵", "赛博朋克服装", "假肢"
-→ use_segmentation=True, top_k=80, limit=80
-
-Describe subject / find tag — e.g. "EVA中蓝发的零号机驾驶员", "命运石之门中的助手"
-→ use_segmentation=False, top_k=20, limit=20
+## Workflow
 
-Full scene → prompt — e.g. "一个穿着白色水手服，蓝色短裙的少女在雨中的城市里奔跑"
-→ use_segmentation=True, top_k=5, limit=80
+After search_tags, pass selected tags to get_related_tags to discover complementary tags via co-occurrence.
+Chain freely: search_tags → get_related_tags → get_related_tags → search_tags for multi-hop exploration.
 
 ## Returns
+
 JSON with: prompt (comma-separated tags), keywords, results.
 Each result: tag, cn_name, category, final_score, count[, wiki if include_wiki=True].
     """
+    _SEARCH_MODE_PRESETS: dict[str, dict] = {
+        "precise_lookup":   {"top_k": 10, "limit": 10, "popularity_weight": 0.15, "use_segmentation": False, "group_mode": "off",    "max_per_group": 2},
+        "concept_explore":  {"top_k": 80, "limit": 80, "popularity_weight": 0.15, "use_segmentation": True,  "group_mode": "expand",  "max_per_group": 2},
+        "subject_describe": {"top_k": 20, "limit": 20, "popularity_weight": 0.15, "use_segmentation": False, "group_mode": "off",    "max_per_group": 2},
+        "full_scene":       {"top_k": 5,  "limit": 80, "popularity_weight": 0.15, "use_segmentation": True,  "group_mode": "diverse", "max_per_group": 2},
+    }
+    preset = _SEARCH_MODE_PRESETS.get(search_mode, _SEARCH_MODE_PRESETS["full_scene"])
+
     _CATEGORY_MAP: dict[str, list[str]] = {
         "all":       ["General", "Character", "Copyright", "Artist", "Meta"],
         "general":   ["General"],
@@ -199,20 +113,20 @@ Each result: tag, cn_name, category, final_score, count[, wiki if include_wiki=T
     }
     target_categories = _CATEGORY_MAP.get(
         category,
-        _CATEGORY_MAP["all"],  # unrecognized value → fall back to all
+        _CATEGORY_MAP["all"],
     )
 
     tagger = await DanbooruTagger.get_instance()
     request = SearchRequest(
         query=query,
-        top_k=top_k,
-        limit=limit,
-        popularity_weight=popularity_weight,
+        top_k=preset["top_k"],
+        limit=preset["limit"],
+        popularity_weight=preset["popularity_weight"],
         show_nsfw=show_nsfw,
-        use_segmentation=use_segmentation,
+        use_segmentation=preset["use_segmentation"],
         target_categories=target_categories,
-        group_mode=group_mode,
-        max_per_group=max_per_group,
+        group_mode=preset["group_mode"],
+        max_per_group=preset["max_per_group"],
     )
     response = await tagger.search_async(request)
     # 计数：每次 MCP 搜索调用均计入搜索、成功、复制；访问不变
@@ -259,6 +173,7 @@ async def get_related_tags(
 ) -> str:
     """
 Return co-occurrence-based tag recommendations for a given tag list (NPMI scoring).
+Only supported for general, copyright, and character tag searches; **artists and meta tags are not supported.**
 
 This tool surfaces tags that frequently appear alongside the seeds in
 Danbooru, mixing categories (General / Character / Copyright) by design.
@@ -370,7 +285,7 @@ JSON array sorted by aggregated NPMI score (descending). Each result:
             item["wiki"] = r.wiki
         output.append(item)
 
-    payload = output
+    payload = {"results": output}
     if corrections:
         correction_notes = [
             f"{bad} → {good}" for bad, good in corrections.items()