Merge pull request #132 from gomate-community/pipeline

feature@support deepresearch-cli

Author: yanqiangmiffy

README.md

@@ -21,6 +21,53 @@ The core design of TrustRAG lies in its **high configurability and modularity**,
 
 ![framework.png](resources%2Fframework.png)
 
+## DeepResearch Features
+
+The DeepResearch framework achieves deep information search and processing through layered queries, recursive iteration, and intelligent decision-making. This process includes the following key steps:
+
+1. Intent Understanding
+   After the user inputs a query, the system parses it into multiple sub-queries to more precisely understand the user's needs.
+
+2. Processing Condition Judgment
+   The system determines whether to continue execution based on the following conditions:
+   1. **Whether the token budget is exceeded**
+   2. **Whether the action depth is exceeded**
+   > If these conditions are met, the query is terminated and the answer is returned directly; otherwise, it enters the recursive execution step.
+
+3. Recursive Execution Steps
+   During recursive execution, the system performs information retrieval, model reasoning, and context processing tasks
+   **Information Retrieval**
+   - **Get current question**
+   - **Build question execution sequence**
+   - **Recursive traversal**
+   - **Depth-first search**
+   - **Model reasoning**
+     > The system performs model reasoning, judging the next action through system prompts and context understanding.
+
+4. Action Type Determination
+   Based on the reasoning results, the system decides the next type of action to execute:
+   - **answer**: Answer action
+   - **reflect**: Reflection action
+   - **search**: Search action
+   - **read**: Reading action
+   - **coding**: Code action
+
+   > These actions affect the context and continuously update the system state.
+
+5. Result Feedback
+   Based on the final action type, the system performs the corresponding task and returns the results to the user, completing the entire process.
+
+DeepResearch process diagram:
+
+![DeepSearch.png](resources/DeepSearch.png)
+
+Run the CLI tool:
+```bash
+cd trustrag/modules/deepsearch
+cp .env.example .env #Configure LLM API and search
+python pipeline.py
+```
+
 ## ✨Key Features
 
 **“Reliable input, Trusted output”**

README_zh.md

@@ -62,6 +62,13 @@ DeepResearch流程示意图如下：
 
 ![DeepSearch.png](resources/DeepSearch.png)
 
+运行cli工具：
+```bash
+cd trustrag/modules/deepsearch
+cp .env.example .env #配置LLM API以及搜索
+python pipeline.py
+```
+
 ## ✨主要特色
 
 **“Reliable input,Trusted output”**

docs/deep-research.md

@@ -0,0 +1,45 @@
+## DeepResearch功能
+
+DeepResearch 框架通过分层查询、递归迭代以及智能决策等步骤，实现深度信息搜索和处理。本流程主要包含以下关键步骤：
+
+1. 意图理解（Intent Understanding）
+用户输入查询后，系统会将其解析为多个子查询，以便更精确地理解用户需求。
+
+2. 处理条件判断
+系统会根据如下条件判断是否继续执行：
+   1. **Token 预算是否超出**
+   2. **动作深度是否超出**
+  >如果满足上述条件，则终止查询并直接返回答案；否则进入递归执行步骤。
+
+3. 递归执行步骤
+在递归执行过程中，系统执行信息检索、模型推理及上下文处理等任务
+**信息检索**
+- **获取当前问题**
+- **构建问题执行序列**
+- **递归遍历**
+- **深度优先搜索**
+-**模型推理**
+  >系统进行模型推理，通过系统提示和上下文理解来判断下一步动作。
+4. 动作类型判定
+根据推理结果，系统决定下一步执行的动作类型：
+- **answer**：回答动作
+- **reflect**：反思动作
+- **search**：搜索动作
+- **read**：阅读动作
+- **coding**：代码动作
+
+  >这些动作会影响上下文，并不断更新系统状态。
+
+5. 结果反馈
+根据最终的动作类型，系统执行相应的任务，并将结果返回给用户，完成整个流程。
+
+DeepResearch流程示意图如下：
+
+![DeepSearch.png](resources/DeepSearch.png)
+
+运行cli工具：
+```bash
+cd trustrag/modules/deepsearch
+cp .env.example .env #配置LLM API以及搜索
+python pipeline.py
+```