refactor: add single tool (#39)

Author: qin-ctx

build.sh

@@ -14,10 +14,12 @@ if ! command -v python3 &> /dev/null; then
     exit 1
 fi
 
+USE_UV=false
 # 2. Check for uv and install dependencies
 if command -v uv &> /dev/null; then
     echo "--> Using uv to install dependencies..."
     uv sync
+    USE_UV=true
 else
     echo "--> uv not found, using pip to install from pyproject.toml..."
     python3 -m pip install -e .
@@ -26,7 +28,11 @@ fi
 # 3. Install PyInstaller if not present
 if ! python3 -c "import PyInstaller" 2>/dev/null; then
     echo "--> PyInstaller not found. Installing..."
-    python3 -m pip install pyinstaller
+    if [ "$USE_UV" = true ]; then
+        uv pip install pyinstaller
+    else
+        python3 -m pip install pyinstaller
+    fi
 fi
 
 # 4. Clean up previous builds

config/prompts_en.yaml

@@ -76,8 +76,8 @@ chat_workflow:
          Criteria: Query suggests need to access information already stored in the system
 
       ## Classification Decision Flow
-      1. Determine if it involves historically stored data/memory in the system ’ qa_analysis
-      2. Determine if it is simple social interaction ’ simple_chat
+      1. Determine if it involves historically stored data/memory in the system � qa_analysis
+      2. Determine if it is simple social interaction � simple_chat
       ## Pattern Recognition Guidance
       - **Time pattern**: Containing time words usually points to qa_analysis
       - **Subject pattern**: First-person queries (I, my) usually involve personal historical data
@@ -312,68 +312,38 @@ chat_workflow:
     # Tool result validation and filtering
     tool_result_validation:
       system: |
-        You are the tool result validation expert of the OpenContext intelligent context management system, responsible for evaluating the quality and relevance of tool call results.
+        You are the tool result filtering expert of the OpenContext intelligent context management system. Your task is simple: filter results that are relevant to the user's question from tool-returned results.
 
-        ## Core Tasks
-        Evaluate the tool call results just executed, determining:
-        1. Which results are valuable for answering user questions (relevance)
-        2. Whether tool parameters are correct (validity)
-        3. Whether certain tools need to be retried with different parameters (improvement suggestions)
-
-        ## Evaluation Dimensions
-
-        ### Relevance Judgment
+        ## Relevance Judgment Criteria
         - **High relevance**: Directly contains information needed to answer the question
-        - **Medium relevance**: Contains some useful information, but not sufficient
+        - **Medium relevance**: Contains some useful information, helpful for answering
         - **Low relevance**: Related to the question but not very useful
         - **Not relevant**: Completely unrelated information
 
         **Only keep high and medium relevance results**
 
-        ### Parameter Validity
-        - Whether tool parameters are reasonable (query terms, time range, etc.)
-        - Whether returned results meet expectations
-        - Whether results are useless due to incorrect parameters
-
-        ### Improvement Suggestions
-        - If a tool parameter is inappropriate, suggest how to adjust
-        - If information is insufficient, suggest what tool should be called in the next round
-        - Avoid repeatedly calling the same tool with the same parameters
-
-        ## Output Format
-        Strictly output in JSON format:
+        ## Output Format (Strictly Follow)
+        Must strictly output in the following JSON format:
         ```json
         {
-          "relevant_result_ids": ["result_id_1", "result_id_3", "result_id_5"],
-          "feedback": "Brief feedback explanation",
-          "retry_suggestions": [
-            {
-              "tool_name": "Tool name",
-              "reason": "Why retry is needed",
-              "suggested_params": {"param": "Suggested parameter value"}
-            }
-          ]
+          "relevant_result_ids": ["result_id_1", "result_id_2", "result_id_3"]
         }
         ```
 
-        **Important Principles**:
-        - relevant_result_ids only includes IDs of high and medium relevance results
-        - feedback summarizes overall quality in 1-2 sentences
-        - retry_suggestions are only provided when retry is really needed
-        - If all results are not relevant, return empty relevant_result_ids: []
+        **Important Requirements**:
+        - Field name must be `relevant_result_ids` (not relevant_results)
+        - Value must be a string array, containing only result_id values
+        - Do not add other fields
+        - If all results are not relevant, return empty array: `{"relevant_result_ids": []}`
       user: |
-        Please evaluate the following tool call results:
+        Please filter results that are relevant to the user's question from the following tool results.
 
         **User Question**: {original_query}
         **Enhanced Query**: {enhanced_query}
 
-        **Tool Call Situation**:
-        {tool_calls}
-
-        **Tool Return Results**:
+        **Tool Results**:
         {tool_results}
-
-        Please analyze the relevance of each result and return JSON format evaluation results.
+        ```
 
     sufficiency_evaluation:
       system: |
@@ -890,7 +860,7 @@ generation:
          - Keep general activities concise but complete overview, ensure all activities are reflected
          - Explain user's specific operations and goals
          - Use natural friendly tone, avoid excessive emoji use, maximum 1-2
-         - Reflect activity coherence and logic, description in three layers: Main activity ’ Specific operation ’ Goal result
+         - Reflect activity coherence and logic, description in three layers: Main activity � Specific operation � Goal result
 
       3. **Context ID Requirements**:
          - Select at most 5 most valuable context IDs to return

config/prompts_zh.yaml

@@ -312,68 +312,38 @@ chat_workflow:
     # 工具结果验证与过滤
     tool_result_validation:
       system: |
-        你是OpenContext智能上下文管理系统的工具结果验证专家,负责评估工具调用结果的质量和相关性。
+        你是OpenContext智能上下文管理系统的工具结果过滤专家。你的任务很简单：从工具返回的结果中，筛选出与用户问题相关的结果。
 
-        ## 核心任务
-        评估刚刚执行的工具调用结果,判断:
-        1. 哪些结果对回答用户问题有价值(相关性)
-        2. 工具参数是否正确(有效性)
-        3. 是否需要用不同参数重试某些工具(改进建议)
-
-        ## 评估维度
-
-        ### 相关性判断
+        ## 相关性判断标准
         - **高相关**: 直接包含回答问题所需的信息
-        - **中相关**: 包含部分有用信息,但不够充分
-        - **低相关**: 与问题相关但不太有用
+        - **中相关**: 包含部分有用信息，对回答有帮助
+        - **低相关**: 与问题相关但用处不大
         - **不相关**: 完全无关的信息
 
         **只保留高相关和中相关的结果**
 
-        ### 参数有效性
-        - 工具参数是否合理(查询词、时间范围等)
-        - 返回结果是否符合预期
-        - 是否因为参数错误导致结果无用
-
-        ### 改进建议
-        - 如果某个工具参数不当,建议如何调整
-        - 如果信息不足,建议下轮应该调用什么工具
-        - 避免重复调用相同工具和参数
-
-        ## 输出格式
-        严格输出JSON格式:
+        ## 输出格式（严格遵守）
+        必须严格按照以下JSON格式输出：
         ```json
         {
-          "relevant_result_ids": ["result_id_1", "result_id_3", "result_id_5"],
-          "feedback": "简要反馈说明",
-          "retry_suggestions": [
-            {
-              "tool_name": "工具名称",
-              "reason": "为什么需要重试",
-              "suggested_params": {"param": "建议的参数值"}
-            }
-          ]
+          "relevant_result_ids": ["result_id_1", "result_id_2", "result_id_3"]
         }
         ```
 
-        **重要原则**:
-        - relevant_result_ids 只包含高相关和中相关结果的ID
-        - feedback 用1-2句话概括整体质量
-        - retry_suggestions 只在确实需要重试时才提供
-        - 如果所有结果都不相关,返回空的 relevant_result_ids: []
+        **重要要求**：
+        - 字段名必须是 `relevant_result_ids`（不是 relevant_results）
+        - 值必须是字符串数组，只包含 result_id 的值
+        - 不要添加其他字段
+        - 如果所有结果都不相关，返回空数组：`{"relevant_result_ids": []}`
       user: |
-        请评估以下工具调用结果:
+        请从以下工具结果中筛选出与用户问题相关的结果。
 
         **用户问题**: {original_query}
         **增强查询**: {enhanced_query}
 
-        **工具调用情况**:
-        {tool_calls}
-
-        **工具返回结果**:
+        **工具结果**:
         {tool_results}
-
-        请分析每个结果的相关性,返回JSON格式的评估结果。
+        ```
 
     sufficiency_evaluation:
       system: |

opencontext/context_consumption/completion/completion_service.py

@@ -21,7 +21,7 @@
 from opencontext.context_consumption.completion.completion_cache import get_completion_cache
 from opencontext.utils.logging_utils import get_logger
 from opencontext.models.enums import CompletionType
-from opencontext.tools.retrieval_tools.text_search_tool import TextSearchTool
+from opencontext.tools.retrieval_tools.semantic_context_tool import SemanticContextTool
 
 logger = get_logger(__name__)
 
@@ -56,7 +56,7 @@ def __init__(self):
         self.chat_client = None
         self.cache = get_completion_cache()  # Use a dedicated cache manager
         self.prompt_manager = None  # Prompt manager
-        self.text_search_tool = None  # TextSearchTool instance
+        self.semantic_search_tool = None  # SemanticContextTool instance
 
         # Completion configuration
         self.max_context_length = 500  # Maximum context length
@@ -73,9 +73,9 @@ def _initialize(self):
             self.storage = get_storage()
 
             self.prompt_manager = get_prompt_manager()
-            
-            # Initialize TextSearchTool
-            self.text_search_tool = TextSearchTool()
+
+            # Initialize SemanticContextTool
+            self.semantic_search_tool = SemanticContextTool()
 
             logger.info("CompletionService initialized successfully")
 
@@ -330,18 +330,17 @@ def _get_reference_suggestions(self, context: Dict[str, Any]) -> List[Completion
         suggestions = []
 
         try:
-            if not self.text_search_tool:
+            if not self.semantic_search_tool:
                 return suggestions
-            
+
             # Use context for vector search
             search_text = context.get('context_before', '')
             if len(search_text) < 10:
                 return suggestions
-            
-            # Use TextSearchTool for semantic search
-            search_results = self.text_search_tool.execute(
+
+            # Use SemanticContextTool for semantic search
+            search_results = self.semantic_search_tool.execute(
                 query=search_text,
-                context_type='vault',  # Only search note content
                 top_k=5
             )
 

opencontext/context_consumption/context_agent/core/llm_context_strategy.py

@@ -67,7 +67,7 @@ async def analyze_and_plan_tools(
         response = await generate_for_agent_async(
             messages=messages,
             tools=self.all_tools,
-            thinking="disabled",
+            # thinking="disabled",
         )
 
         # Extract tool calls from the response
@@ -231,7 +231,7 @@ async def execute_tool_calls_parallel(
         for call in tool_calls:
             function_name = call.get("function", {}).get("name")
             function_args = call.get("function", {}).get("arguments", {})
-            
+            # self.logger.info(f"Tool call {function_name} args {function_args}")
             if function_name:
                 task = self.tools_executor.run_async(function_name, function_args)
                 tasks.append((function_name, task))
@@ -358,12 +358,12 @@ async def validate_and_filter_tool_results(
         user_template = prompts.get("user", "")
 
         # Format tool calls summary
-        tool_calls_summary = []
-        for call in tool_calls:
-            tool_calls_summary.append({
-                "tool_name": call.get("function", {}).get("name"),
-                "parameters": call.get("function", {}).get("arguments", {})
-            })
+        # tool_calls_summary = []
+        # for call in tool_calls:
+        #     tool_calls_summary.append({
+        #         "tool_name": call.get("function", {}).get("name"),
+        #         "parameters": call.get("function", {}).get("arguments", {})
+        #     })
 
         # Format tool results summary
         results_summary = []
@@ -378,7 +378,7 @@ async def validate_and_filter_tool_results(
         user_prompt = user_template.format(
             original_query=intent.original_query,
             enhanced_query=intent.enhanced_query,
-            tool_calls=json.dumps(tool_calls_summary, ensure_ascii=False, indent=2),
+            # tool_calls=json.dumps(tool_calls_summary, ensure_ascii=False, indent=2),
             tool_results=json.dumps(results_summary, ensure_ascii=False, indent=2),
         )
 
@@ -387,10 +387,10 @@ async def validate_and_filter_tool_results(
         messages = [{"role": "system", "content": system_prompt}]
 
         # Add user's chat history to give LLM context awareness
-        if existing_context.chat_history:
-            recent_messages = existing_context.chat_history[-10:]  # Last 10 messages
-            for msg in recent_messages:
-                messages.append({"role": msg.role, "content": msg.content})
+        # if existing_context.chat_history:
+        #     recent_messages = existing_context.chat_history[-10:]  # Last 10 messages
+        #     for msg in recent_messages:
+        #         messages.append({"role": msg.role, "content": msg.content})
 
         messages.append({"role": "user", "content": user_prompt})
 
@@ -405,17 +405,27 @@ async def validate_and_filter_tool_results(
             validation_result = parse_json_from_response(response)
 
             # Extract relevant result IDs
-            relevant_ids = set(validation_result.get("relevant_result_ids", []))
 
-            # Filter relevant context items
-            relevant_items = [
-                item for item in tool_results
-                if item.id in relevant_ids
-            ]
+            # Fallback: if no valid IDs found, return all results to avoid data loss
+            if 'relevant_result_ids' not in validation_result:
+                self.logger.warning(
+                    "No relevant_result_ids found in validation response. "
+                    "Returning all results as fallback."
+                )
+                relevant_items = tool_results
+            else:
+                relevant_ids = set(validation_result.get("relevant_result_ids", []))
+                # Filter relevant context items
+                relevant_items = [
+                    item for item in tool_results
+                    if item.id in relevant_ids
+                ]
+                # self.logger.info(f"Filtered to {len(relevant_items)}/{len(tool_results)} relevant items")
+
             # Build validation message for conversation history
             validation_message = {
                 "role": "assistant",
-                "content": f"Tool validation result:\n{json.dumps(validation_result, ensure_ascii=False, indent=2)}"
+                "content": f"Filtered {len(relevant_items)}/{len(tool_results)} relevant results"
             }
 
             return relevant_items, validation_message