fix(search): narrow adaptive deep reranking

Author: simonsysun

CHANGELOG.md

@@ -11,7 +11,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `seeklink get PATH:LINE -C N` prints a grep-style context window around a search hit, returning `N` lines before and after the requested line while preserving direct filesystem reads and path-escape protection.
 - `seeklink search --json` and `seeklink status --json` emit stable machine-readable stdout for agents that should not scrape the human text format.
 - `seeklink search --rerank-k N` and `seeklink search --no-rerank` let callers trade precision for latency per query without changing the global reranker configuration.
-- `seeklink search --rerank-k auto` chooses a 5- or 20-candidate reranker budget from query shape, keeping exact title / alias and English queries fast while giving CJK and filtered queries deeper reranking.
+- `seeklink search --rerank-k auto` chooses a 5- or 20-candidate reranker budget from query shape, keeping exact title / alias, English, and ordinary CJK queries fast while giving filtered and CJK technical queries deeper reranking.
 - The blind-test runner now accepts `--rerank-k N`, `--rerank-k auto`, and `--no-rerank`, and records requested plus resolved reranking metadata in result JSON for latency / quality sweeps.
 
 ### Fixed

docs/blind-test.md

@@ -100,8 +100,8 @@ For each `(query, config)` pair (recorded by the runner):
   when reranking is disabled, or `"auto"` for query-sensitive routing)
 - `resolved_rerank_k` — actual numeric candidate budget used for this query
   (`5` or `20` for `auto`, `0` when reranking is disabled)
-- `rerank_k_reason` — why `auto` chose that budget (`title`, `cjk`,
-  `filter`, `default`, or `fixed`)
+- `rerank_k_reason` — why `auto` chose that budget (`title`,
+  `cjk_technical`, `filter`, `default`, or `fixed`)
 - `recall_at_10` — fraction of `expected_paths` in top-10
 - `mrr` — reciprocal rank of first expected hit in top-10 (0 if none)
 

seeklink/search.py

@@ -22,6 +22,27 @@
 RerankK = int | Literal["auto"]
 AUTO_RERANK_FAST_K = 5
 AUTO_RERANK_DEEP_K = 20
+_CJK_TECHNICAL_RERANK_TERMS = (
+    "向量",
+    "嵌入",
+    "文档",
+    "切块",
+    "算法",
+    "搜索",
+    "检索",
+    "融合",
+    "模型",
+    "数据库",
+    "fsrs",
+    "bm25",
+    "rrf",
+    "rag",
+    "anki",
+    "hybrid",
+    "chunk",
+    "embedding",
+    "vector",
+)
 
 
 @dataclass(slots=True)
@@ -36,6 +57,11 @@ def _contains_cjk(text: str) -> bool:
     return any("\u3400" <= ch <= "\u9fff" for ch in text)
 
 
+def _contains_technical_rerank_term(text: str) -> bool:
+    folded = text.casefold()
+    return any(term in folded for term in _CJK_TECHNICAL_RERANK_TERMS)
+
+
 def _resolve_rerank_k_with_reason(
     query: str,
     rerank_k: RerankK,
@@ -46,10 +72,10 @@ def _resolve_rerank_k_with_reason(
     """Resolve a numeric rerank budget for one query.
 
     The default CLI path still passes an integer. The explicit "auto" mode is
-    a conservative policy from the 22-query pilot: English and title/alias
-    lookups got most of the reranker benefit by reranking only the top 5, while
-    CJK / mixed queries without a title hit needed deeper candidates to recover
-    recall.
+    a conservative policy from the 22-query pilot: English, title/alias, and
+    ordinary CJK lookups got most of the reranker benefit by reranking only the
+    top 5, while CJK / mixed technical queries needed deeper candidates to
+    recover recall.
     """
     if isinstance(rerank_k, int):
         return rerank_k, "fixed"
@@ -61,8 +87,8 @@ def _resolve_rerank_k_with_reason(
         return AUTO_RERANK_DEEP_K, "filter"
     if title_ranks:
         return AUTO_RERANK_FAST_K, "title"
-    if _contains_cjk(query):
-        return AUTO_RERANK_DEEP_K, "cjk"
+    if _contains_cjk(query) and _contains_technical_rerank_term(query):
+        return AUTO_RERANK_DEEP_K, "cjk_technical"
     return AUTO_RERANK_FAST_K, "default"
 
 

tests/test_search.py

@@ -62,7 +62,15 @@ def test_title_match_uses_fast_budget(self):
             title_ranks={10: 1},
         ) == 5
 
-    def test_cjk_without_title_match_uses_deep_budget(self):
+    def test_cjk_without_title_match_uses_fast_budget(self):
+        assert _resolve_rerank_k(
+            "学完就忘怎么办",
+            "auto",
+            has_filter=False,
+            title_ranks={},
+        ) == 5
+
+    def test_cjk_technical_query_uses_deep_budget(self):
         assert _resolve_rerank_k(
             "把文档切块放进向量库",
             "auto",