Summary
Built an enterprise RAG-QA system with multi-source ingestion, hybrid retrieval, citation-backed answers, human clarification, governance, and regression evaluation.
Business Value
Turned scattered documents into an enterprise AI QA platform with continuous sync, traceable citations, configurable agent workflows, and inference governance.
Engineering Depth
Demonstrates integrated engineering across hybrid retrieval, LangGraph orchestration, agent autonomy, tool ecosystems, semantic caching, model health, layered instructions, hallucination checks, SDK delivery, and evaluation gates.
Evidence
中软国际项目归档
Delivery record · Confidence Medium · Verified 2026-03-13
- Evidence level: strict review (core sections only show verifiable metrics)
- Source type: Project delivery record
- Source link: no public link provided, review against delivery records
- Verified at: 2026-03-13 (96 days ago, fresh evidence)
Rationale: Medium confidence: missing a public source link.
本地仓库 README / AI_HIGHLIGHTS / docs/reference / core modules / tests / recent commits
Repository · Confidence High · Verified 2026-06-10
- Evidence level: strict review (core sections only show verifiable metrics)
- Source type: Repository / code records
- Source link: public link provided for independent review
- Verified at: 2026-06-10 (7 days ago, fresh evidence)
Rationale: High confidence: organized under strict evidence rules, traceable to repository or code records, includes an accessible source link, verified 7 days ago.
View evidenceBackground
企业资料分散在本地目录、Notion 等多源，业务希望把制度、FAQ、项目文档沉淀为可检索、可引用、可持续同步，并能支持多步 Agent 问答的知识平台。
Challenge
传统单路检索难以同时覆盖标题、关键词与语义混合场景；Agent 模式如果缺少工具边界、反思、自主拆解、缓存治理和安全检测，也容易变成不可解释的高成本黑盒。
Action and Results
Solution
- 搭建混合检索：结构、全文、向量三路召回配合 query rewrite、加权 RRF 与 rerank，提升复杂问题场景下的证据命中率。
- 重构 Agent 运行时：在 Gateway 与检索层引入 LangGraph，支持 checkpoint、interrupt/resume、
step_events、工具注册中心、任务拆解 DAG、反思闭环与三层记忆。 - 补齐平台治理：建设多知识库、多源连接器、batch dry-run/jobs、token-aware chunk、retrieve/debug、人工接管队列、五层指令合并、6 类场景模板、RAG 幻觉检测、企业聊天 v2 与 Python SDK。
- 强化推理与评测闭环：引入三层语义缓存、模型健康熔断、复杂度驱动路由、请求合并和 readyz/trace 诊断，并通过 retrieval fixture、local ingest fixture、22 个后端测试文件、9 个前端测试文件和 400+ 测试项约束回归。
Result
形成面向中文企业场景的 AI 问答平台，支持 grounded answer、引用溯源、多源同步、可解释检索调试、可恢复执行、Agent 自主决策、人工接管、推理治理和最小评测回归闭环。
Key Signals
Combined structured, full-text, and vector retrieval. Used weighted RRF and reranking to balance hit rate and traceability, then exposed retrieve/debug evidence paths. Connected retrieval and QA into a LangGraph runtime with checkpoint recovery, step_events, verification metadata, and offline evaluation gates. Tech Stack
FastAPILangGraphPostgreSQLQdrantRAGVue 3FastEmbedDockerLangChainAgent RuntimePython SDK