← 工作项目← Work

Smart-GSA 智能审编系统Smart-GSA Curation System

2025.10 – 至今 · 中国科学院北京基因组研究所(国家生物信息中心)Oct 2025 – present · Beijing Institute of Genomics, CAS

面向国家级基因组数据库(GSA)的独立审编后台,承接日均 500 TB 数据吞吐与数百条元数据提交的审核工作。The standalone curation back office for a national genomics database (GSA) — hundreds of metadata submissions reviewed every day, against 500 TB of daily data throughput.

JavaSpring BootSpring AI RAGAgentEvaluation

1自研 AgentLoopA hand-written agent loop

7 个事件 + 3 个钩子7 events + 3 hooks

2将三级审核任务抽象为 Skills + ToolsThree tiers of review, expressed as Skills + Tools

将 13 条审编 SOP 聚合为 4 份 Skills,按需调用13 curation SOPs consolidated into 4 Skills, read on demand

将业务逻辑整理成七个工具:看数据 · 跑脚本 · 查库 · 检索 · 交结论Business logic split into seven tools: read data · run checks · query the DB · retrieve · submit findings

3RAGFlow:将向量库作为工具进行调用RAGFlow: the vector store is called as a tool

将三类信息向量化:ENA/SRA 官方语义定义 · 关联文献的 Methods 片段 · 通过审核的范例Three kinds of material are indexed: official ENA/SRA field semantics · Methods sections of linked papers · descriptions that passed review

结构化的依据不进向量库 —— 物种谱系、受控词表、字段必填性走工具校验Structured criteria stay out of the index — taxonomy, controlled vocabularies and required-field rules are checked by tools

4beforeStop 钩子The beforeStop hook

结束前强制提交审核结果Forces the review result to be submitted before the loop may end

5beforeTool 钩子The beforeTool hook

逐次调用拦截:拦住重复的工具调用 + 工具调用预算Intercepts every call: blocks duplicate tool calls, enforces the call budget

6五种结束状态Five terminal states

DONE 不再请求工具 · CAPPED 撞熔断机制 · TRUNCATED 模型输出上限 · REFUSED 模型拒答 · FAILED 调模型失败DONE no more tool calls · CAPPED circuit breaker tripped · TRUNCATED output limit hit · REFUSED model declined · FAILED model call failed

AgentToolUse任务TaskExcel 解析Excel parsing上下文 + 压缩模块Context + compaction大模型请求LLM requestStop ReasonbeforeTool工具调用Tool callbeforeStop结束End审核 SOP SkillsCuration SkillsToolsRAGFlow