From 4d000c94f48fcbe1d95fc255aa7842be03dad4d0 Mon Sep 17 00:00:00 2001
From: tangwang <tangwang@hsyl>
Date: Thu, 2 Apr 2026 11:29:03 +0800
Subject: [PATCH] 融合公式调参

---
 config/config.yaml                           | 20 ++++++++++----------
 scripts/evaluation/eval_framework/prompts.py | 45 ++++++++++++++++++++++-----------------------
 2 files changed, 32 insertions(+), 33 deletions(-)

diff --git a/config/config.yaml b/config/config.yaml
index 0b2f9c2..7a38319 100644
--- a/config/config.yaml
+++ b/config/config.yaml
@@ -285,7 +285,7 @@ coarse_rank:
   input_window: 700
   output_window: 240
   fusion:
-    es_bias: 0.1
+    es_bias: 10.0
     es_exponent: 0.05
     text_bias: 0.1
     text_exponent: 0.35
@@ -293,10 +293,10 @@ coarse_rank:
     # 因为es的打分已经给了trans进行了折扣，所以这里不再继续折扣
     text_translation_weight: 1.0
     knn_text_weight: 1.0
-    knn_image_weight: 1.0
-    knn_tie_breaker: 0.1
+    knn_image_weight: 2.0
+    knn_tie_breaker: 0.3
     knn_bias: 0.6
-    knn_exponent: 0.2
+    knn_exponent: 0.4
 
 # 精排配置（轻量 reranker）
 fine_rank:
@@ -324,21 +324,21 @@ rerank:
   #   max(knn_text_weight * text_knn, knn_image_weight * image_knn)
   #   + knn_tie_breaker * 另一侧较弱信号
   fusion:
-    es_bias: 0.1
+    es_bias: 10.0
     es_exponent: 0.05
-    rerank_bias: 1.0e-05
+    rerank_bias: 0.1
     rerank_exponent: 1.15
-    fine_bias: 1.0e-05
+    fine_bias: 0.1
     fine_exponent: 1.0
     text_bias: 0.1
     text_exponent: 0.25
     # base_query_trans_* 相对 base_query 的权重（见 search/rerank_client 中文本 dismax 融合）
     text_translation_weight: 0.8
     knn_text_weight: 1.0
-    knn_image_weight: 1.0
-    knn_tie_breaker: 0.1
+    knn_image_weight: 2.0
+    knn_tie_breaker: 0.3
     knn_bias: 0.6
-    knn_exponent: 0.2
+    knn_exponent: 0.4
 
 # 可扩展服务/provider 注册表（单一配置源）
 services:
diff --git a/scripts/evaluation/eval_framework/prompts.py b/scripts/evaluation/eval_framework/prompts.py
index bee800c..f7dea90 100644
--- a/scripts/evaluation/eval_framework/prompts.py
+++ b/scripts/evaluation/eval_framework/prompts.py
@@ -130,21 +130,18 @@ Typical situations:
 - The core product type matches, but the product clearly violates an explicit and important requirement in the query, with little or no acceptable substitutability.
 
 Typical examples:
-- Query: “pants”
-  Product: “shoes”
-  → Wrong product type.
-- Query: “slim-fit pants”
-  Product: “loose wide-leg pants”
-  → Clear contradiction in fit, with extremely low substitutability.
-- Query: “sleeveless dress”
-  Product: “long-sleeve dress”
-  → Clear contradiction in sleeve type.
-- Query: “jeans”
-  Product: “sweatpants”
-  → Different core category, with significantly different style and wearing scenario.
-- Query: “boots”
-  Product: “sneakers”
-  → Different core category, different function, and different usage scenario.
+
+1. **Different core product type with no substitutability**
+- Query: "pants", Product: "shoes"
+- Query: "boots", Product: "sneakers"
+
+2. **Close core product type but clear conflict on key attributes**
+- Query: "slim-fit pants", Product: "loose wide-leg pants"
+- Query: "sleeveless dress", Product: "long-sleeve dress"
+- Query: "loose-fitting hoodie", Product: "slim-fit T-shirt"
+
+3. **Same broad category but significant difference in style or usage scenario**
+- Query: "jeans", Product: "sweatpants / dress pants"
 
 ## Judgment Principles
 
@@ -277,16 +274,18 @@ _CLASSIFY_TEMPLATE_ZH = """你是一个服饰电商搜索系统中的相关性�
 - 核心商品类型匹配，但商品明显违背了查询中一个明确且重要的要求，且几乎不具备可接受的替代性。
 
 典型情况：
+
+**1. 核心品类不同，且无替代性**
 - 查询：“裤子”，商品：“鞋子”
-  → 商品类型错误。
-- 查询：“修身裤”，商品：“宽松阔腿裤”
-  → 与版型要求明显冲突，替代性极低。
-- 查询：“无袖连衣裙”，商品：“长袖连衣裙”
-  → 与袖型要求明显冲突。
-- 查询：“牛仔裤”，商品：“运动裤”
-  → 核心品类不同（牛仔裤 vs 运动裤），风格和场景差异大。
 - 查询：“靴子”，商品：“运动鞋”
-  → 核心品类不同，功能和适用场景差异大。
+
+**2. 核心品类相近，但关键属性明显冲突**
+- 查询：“紧身裤”，商品：“阔腿裤”
+- 查询：“无袖连衣裙”，商品：“长袖连衣裙”
+- 查询：“宽松卫衣”，商品：“修身T恤”
+
+**3. 核心品类同属大类，但风格、场景差异巨大**
+- 查询：“牛仔裤”，商品：“运动裤 / 西裤”
 
 ## 判断原则
 
--
libgit2 0.21.2