feat: Optimize loop iterations by caching template compilation

apps/application/workflow/loop_workflow_manage.py

@@ -7,11 +7,12 @@
 @desc:
 """
 
-from typing import Dict, Optional, Callable
+from typing import Dict, Callable
 
-from application.workflow.common import Workflow, WorkflowType, Node
+from application.workflow.common import Workflow, WorkflowType
 from application.workflow.i_node import INode
 from application.workflow.workflow_manage import WorkflowManage, CallBack
+from common.utils.prompt_template import render_prompt
 
 
 class LoopWorkFlowManage(WorkflowManage):
@@ -34,13 +35,10 @@ def get_parent_context(self, node_id, key):
         return self.parent_workflow_manage.get_context(node_id, key)
 
     def generate_prompt(self, prompt):
-        prompt = self.workflow.reset_prompt(prompt)
-        prompt = self.parent_workflow_manage.workflow.reset_prompt(prompt)
+        input_template = self.workflow.reset_prompt(prompt)
+        input_template = self.parent_workflow_manage.workflow.reset_prompt(input_template)
         context = {**self.context, **self.parent_workflow_manage.context}
-        from langchain_core.prompts import PromptTemplate
-
-        prompt_template = PromptTemplate.from_template(prompt, template_format="jinja2")
-        return prompt_template.format(context=context)
+        return render_prompt(input_template, context)
 
     def get_reference_field(self, node_id, fields):
         """

apps/application/workflow/workflow_manage.py

@@ -12,13 +12,12 @@
 import threading
 from typing import List, Dict, Optional, Callable
 
-from langchain_core.prompts import PromptTemplate
-
 from application.workflow.common import Workflow, WorkflowType, Node, get_node_parameters
 from application.workflow.i_node import INode, Signal
-from application.workflow.message.struct.content import Content, Position
+from application.workflow.message.struct.content import Content
 
 from application.workflow.status import Status
+from common.utils.prompt_template import render_prompt
 
 
 class CallBack:
@@ -223,9 +222,8 @@ def generate_prompt(self, prompt):
         @param prompt: 提示词
         @return: 处理后的提示词
         """
-        prompt = self.workflow.reset_prompt(prompt)
-        prompt_template = PromptTemplate.from_template(prompt, template_format="jinja2")
-        return prompt_template.format(context=self.context)
+        input_template = self.workflow.reset_prompt(prompt)
+        return render_prompt(input_template, self.context)
 
     def get_reference_field(self, node_id, fields):
         """

apps/common/utils/prompt_template.py

@@ -0,0 +1,34 @@
+# coding=utf-8
+
+import hashlib
+
+from jinja2.sandbox import SandboxedEnvironment
+
+from common.cache.mem_cache import MemCache
+
+# 缓存 reset_prompt 后的模板编译产物（只依赖配置，同一工作流运行期间恒定）
+template_cache = MemCache(
+    "workflow_template_cache",
+    {
+        "TIMEOUT": 3600,  # 缓存有效期为 1 小时
+        "OPTIONS": {
+            "MAX_ENTRIES": 1000,  # 最多缓存 1000 个条目
+            "CULL_FREQUENCY": 10,  # 达到上限时，删除约 1/10 的缓存
+        },
+    },
+)
+
+
+def render_prompt(input_template, context):
+    """
+    渲染提示词，编译一次后缓存编译产物，之后每轮仅渲染
+    @param input_template: reset_prompt 处理后的模板字符串
+    @param context:        渲染模板所需的上下文数据
+    @return: 渲染后的提示词字符串
+    """
+    key = f"workflow_template::{hashlib.sha256(input_template.encode('utf-8')).hexdigest()}"
+    template = template_cache.get(key)
+    if template is None:
+        template = SandboxedEnvironment().from_string(input_template)
+        template_cache.set(key, template)
+    return template.render(context=context)