In Progress
智能招投标AI平台
面向招投标全流程，打通商机采集、招标文件理解、知识检索、投标内容生成与人工复核链路。
系统快照
架构、信号与证据会先于正文呈现。- Doc pipeline
- Structured
- Review loop
- Human-in-loop
Problem
招投标业务从商机获取、招标文件理解、知识检索到投标内容生成都高度依赖人工，信息分散、复杂文档处理成本高，历史经验也难以复用。系统不仅要生成文本，还要理解文档结构、约束条件和业务事实。
Constraints
- 文档包含表格、图片、代码块和扫描件，不能简单按长度切分。
- 生成内容必须可追溯到来源。
- 关键结论需要人工复核，不能完全自动提交。
System Design
系统从业务流程出发拆分商机采集、文件资产管理、文档解析、RAG 服务与 Agent 服务等能力域，明确服务边界、数据契约和上下游链路，并将共性能力沉淀为可复用服务。
Agent Runtime 围绕 Run 生命周期、Runtime Context、Tool Governance、Memory / Skill、LongTask、Subagent 等机制设计统一运行控制面，解决复杂 Agent 任务中的上下文治理、工具安全、长任务执行与多智能体协同问题。
文档智能链路打通复杂文档从解析、结构化、知识入库到检索的完整过程，将 Parser Artifacts 转换为可追溯的结构化内容，并建设 Section-aware Chunk、异步索引、图文混合检索、Rerank 与 Grounded Evidence。
Key Decisions
把投标系统做成“可控生成流程”比做成“万能写作助手”更可靠。每一步都需要中间结果、引用依据和人工确认入口。
Impact
- 完成核心 AI 技术底座建设，打通从商机采集、文件处理、文档解析、知识索引与检索到 Agent 智能应用的完整链路。
- 将 Agent、RAG、文档解析等能力从业务内功能沉淀为独立平台服务，降低后续业务场景的重复建设与接入成本。
- 为招投标 AI 能力持续扩展提供统一技术底座。
Reflection
后续重点会放在评测体系、模板管理、引用质量和多人协作流程上。
Problem
Bidding workflows involve opportunity discovery, tender documents, historical proposals, qualification materials, and strict formatting requirements.
The system needs to understand document structure, constraints, and business facts rather than only generate text.
Constraints
Documents contain tables, images, code blocks, and scanned pages, so they cannot be chunked only by length.
Generated content must be traceable to sources.
Important conclusions need human review before submission.
System Design
The system separates opportunity collection, file assets, document parsing, RAG services, and Agent services into clear capability domains.
The Agent Runtime provides a shared control plane for run lifecycle, runtime context, tool governance, memory and skills, long tasks, and subagent collaboration.
The document pipeline converts parser artifacts into section-aware chunks, then supports async indexing, mixed media retrieval, reranking, and grounded evidence.
Key Decisions
A controlled generation workflow is more reliable than a universal writing assistant.
Every step needs intermediate results, citations, and an entry point for human confirmation.
Impact
The project built the core AI foundation from opportunity collection and document processing to knowledge indexing, retrieval, and Agent applications.
Agent, RAG, and document parsing capabilities were extracted from the business workflow into reusable platform services.
Reflection
Future work should focus on evaluation, template management, citation quality, and multi-person collaboration.