05-rag-documents.jsonl

{"doc_id":"rag-001","title":"Ingress 排障手册","source":"internal-wiki","updated_at":"2026-04-10","content":"当 Ingress 返回 502 时，应优先检查 upstream Pod 是否健康、Service 端口映射是否正确、EndpointSlice 是否正常生成。","tags":["k8s","ingress","ops"]}
{"doc_id":"rag-002","title":"LoRA 微调说明","source":"ml-team-note","updated_at":"2026-04-08","content":"LoRA 通过在冻结基座模型的前提下训练低秩增量矩阵，降低显存和训练成本。产物通常是 adapter 配置和 adapter 权重文件。","tags":["llm","training","lora"]}
{"doc_id":"rag-003","title":"数据脱敏规范","source":"security-policy","updated_at":"2026-04-01","content":"进入训练或检索系统前，必须处理手机号、身份证号、邮箱、住址、银行卡号等敏感信息，并保留脱敏审计记录。","tags":["security","data-governance"]}