ck draft

Author: tpfz

backend/modules/evaluation/infra/repo/experiment/ck/expt_turn_result_filter.go

@@ -376,7 +376,7 @@ func (d *exptTurnResultFilterDAOImpl) buildKeywordSearchConditions(ctx context.C
 // buildBaseSQL 构建基础SQL语句
 func (d *exptTurnResultFilterDAOImpl) buildBaseSQL(ctx context.Context, whereSQL, keywordCond string, args *[]interface{}) string {
 	sql := "SELECT  etrf.item_id, etrf.status FROM " + getClickHouseDatabaseName() + ".expt_turn_result_filter etrf"
-	sql += " WHERE 1=1"
+	sql += " FINAL WHERE 1=1"
 	if keywordCond != "" {
 		// 将 evalSetSyncCkDate 插入到 args 切片的第一个位置
 		newArgs := make([]interface{}, 0, len(*args))

release/deployment/docker-compose/conf/model_config.yaml

@@ -1,159 +1,31 @@
 models:
-  # reasoning model
   - id: 1
-    workspace_id: 0 # In the future, there will be the concept of public/private workspaces. Public models are managed by the public workspace, private models by the private workspace. Currently, all models belong to the public workspace, and the public workspace id is temporarily set to 0.
-    name: "deepseek-r1-distill-qwen-32b-250120"
+    name: "doubao"
     frame: "eino"
     protocol: "ark"
     protocol_config:
-      api_key: "a715a14b-3b93-47da-8bc9-844c12fecff7"
-      model: "ep-20250304143659-5bcjt"
+      api_key: "***"
+      model: "***"
     param_config:
       param_schemas:
         - name: "temperature"
           label: "temperature"
-          desc: "Increasing temperature will make the model output more diverse and creative. Conversely, lowering the temperature will make the output more compliant with instructions but reduce diversity. It is recommended not to adjust together with 'Top p'."
+          desc: "Increasing temperature makes model output more diverse and creative, while decreasing it makes output more focused on instructions but less diverse. It's recommended not to adjust this simultaneously with 'Top p'."
           type: "float"
           min: "0"
           max: "1.0"
           default_val: "0.7"
         - name: "max_tokens"
           label: "max_tokens"
-          desc: "Controls the maximum length of model output tokens. Typically, 100 tokens are about 150 Chinese characters."
+          desc: "Controls the maximum number of tokens in model output. Typically, 100 tokens equals about 150 Chinese characters."
           type: "int"
           min: "1"
-          max: "8192"
+          max: "4096"
           default_val: "2048"
-  # multimodal model
-  - id: 2
-    workspace_id: 0 # In the future, there will be the concept of public/private workspaces. Public models are managed by the public workspace, private models by the private workspace. Currently, all models belong to the public workspace, and the public workspace id is temporarily set to 0.
-    name: "doubao-1.5-vision-pro-32k"
-    desc: ""
-    ability:
-      max_context_tokens: 65536
-      max_input_tokens: 65536
-      max_output_tokens: 8192
-      function_call: false
-      json_mode: false
-      multi_modal: true
-      ability_multi_modal:
-        image: true
-        ability_image:
-          url_enabled: true
-          binary_enabled: true
-          max_image_size: 20 # unit MB
-          max_image_count: 20
-    frame: "eino"
-    protocol: "ark"
-    protocol_config:
-      base_url: "https://ark.cn-beijing.volces.com/api/v3"
-      api_key: "a715a14b-3b93-47da-8bc9-844c12fecff7"
-      model: "ep-20250304145131-ndcct"
-      protocol_config_ark:
-        region: "cn-beijing"
-    scenario_configs:
-      default:
-        scenario: "default"
-        quota:
-          qpm: 0
-          tpm: 0
-        unavailable: false
-      evaluator:
-        scenario: "evaluator"
-        quota:
-          qpm: 0
-          tpm: 0
-        unavailable: false
-    param_config:
-      param_schemas:
-        - name: "temperature"
-          label: "temperature"
-          desc: "Increasing temperature will make the model output more diverse and creative. Conversely, lowering the temperature will make the output more compliant with instructions but reduce diversity. It is recommended not to adjust together with 'Top p'."
-          type: "float"
-          min: "0"
-          max: "1.0"
-          default_val: "0.7"
-        - name: "max_tokens"
-          label: "max_tokens"
-          desc: "Controls the maximum length of model output tokens. Typically, 100 tokens are about 150 Chinese characters."
-          type: "int"
-          min: "1"
-          max: "8192"
-          default_val: "2048"
-  # fc model
-  - id: 3
-    workspace_id: 0 # In the future, there will be the concept of public/private workspaces. Public models are managed by the public workspace, private models by the private workspace. Currently, all models belong to the public workspace, and the public workspace id is temporarily set to 0.
-    name: "doubao-1.5-lite-32k"
-    desc: ""
-    ability:
-      max_context_tokens: 32000
-      max_input_tokens: 32000
-      max_output_tokens: 12000
-      function_call: true
-      json_mode: false
-      multi_modal: false
-    frame: "eino"
-    protocol: "ark"
-    protocol_config:
-      base_url: "https://ark.cn-beijing.volces.com/api/v3"
-      api_key: "a715a14b-3b93-47da-8bc9-844c12fecff7"
-      model: "ep-20250227201314-frn9m"
-      protocol_config_ark:
-        region: "cn-beijing"
-    scenario_configs:
-      default:
-        scenario: "default"
-        quota:
-          qpm: 0
-          tpm: 0
-        unavailable: false
-      evaluator:
-        scenario: "evaluator"
-        quota:
-          qpm: 0
-          tpm: 0
-        unavailable: false
-    param_config:
-      param_schemas:
-        - name: "temperature"
-          label: "temperature"
-          desc: "Increasing temperature will make the model output more diverse and creative. Conversely, lowering the temperature will make the output more compliant with instructions but reduce diversity. It is recommended not to adjust together with 'Top p'."
-          type: "float"
-          min: "0"
-          max: "1.0"
-          default_val: "0.1"
         - name: "top_p"
           label: "top_p"
-          desc: "The model will consider token results within top_p probability mass."
+          desc: "Selects the minimum token set with cumulative probability reaching top_p during generation, excluding tokens outside the set, balancing diversity and reasonableness."
           type: "float"
-          min: "0"
+          min: "0.001"
           max: "1.0"
-          default_val: "0.1"
-        - name: "max_tokens"
-          label: "max_tokens"
-          desc: "Controls the maximum length of model output tokens. Typically, 100 tokens are about 150 Chinese characters."
-          type: "int"
-          min: "1"
-          max: "8192"
-          default_val: "2048"
-        - name: "top_k"
-          label: "top_k"  # Displayed as a name on the front end
-          desc: "Only sample from the top k tokens with the highest probability to limit the candidate range and improve generation stability." # Displayed as a description on the front end
-          type: "int" # Required. Must be float, int, bool, string
-          min: "1"
-          max: "100"
-          default_val: "50"
-        - name: "frequency_penalty"
-          label: "frequency_penalty"  # Displayed as a name on the front end
-          desc: "Penalizes generated tokens, with higher frequency resulting in higher penalties, suppressing repetitive content." # Displayed as a description on the front end
-          type: "float" # Required. Must be float, int, bool, string
-          min: "0"
-          max: "2.0"
-          default_val: "0"
-        - name: "presence_penalty"
-          label: "presence_penalty"  # Displayed as a name on the front end
-          desc: "Penalizes all tokens that have appeared, preventing the same content from appearing repeatedly, increasing content diversity." # Displayed as a description on the front end
-          type: "float" # Required. Must be float, int, bool, string
-          min: "0"
-          max: "2.0"
-          default_val: "0"
+          default_val: "0.7"

release/deployment/helm-chart/charts/app/bootstrap/init/clickhouse/init-sql/evaluation.sql

@@ -1,7 +1,7 @@
 -- Copyright (c) 2025 coze-dev Authors
 -- SPDX-License-Identifier: Apache-2.0
 
--- Create expt_turn_result_filter table for kubernetes environment
+-- Create expt_turn_result_filter table for docker environment
 CREATE TABLE IF NOT EXISTS expt_turn_result_filter
 (
     `space_id` String,
@@ -24,9 +24,8 @@ CREATE TABLE IF NOT EXISTS expt_turn_result_filter
     INDEX idx_expt_id expt_id TYPE bloom_filter() GRANULARITY 1,
     INDEX idx_item_id item_id TYPE bloom_filter() GRANULARITY 1,
     INDEX idx_turn_id turn_id TYPE bloom_filter() GRANULARITY 1
-)
-ENGINE = ReplicatedReplacingMergeTree('/clickhouse/stone_dataengine_commercial/cozeloop_evaluation/{shard}', '{replica}')
-PARTITION BY created_date
-ORDER BY (expt_id, cityHash64(item_id), turn_id)
-SAMPLE BY cityHash64(item_id)
-SETTINGS index_granularity = 8192;
+    )
+    ENGINE = ReplacingMergeTree(updated_at)
+    PARTITION BY created_date
+    ORDER BY (expt_id, item_id, turn_id)
+    SETTINGS index_granularity = 8192;