能力组合Capability Combination
AI 数据库三层能力模型A Three-layer AI Database Capability Model
AI 应用的复杂性不是来自某一种数据类型，而是多模数据、工程执行和可信治理同时存在。川序以关系数据为权威事实，按需组合全文、向量与图等能力；数据库适配层保持统一产品契约，不要求三种数据库同时部署。AI application complexity does not come from one data type. It comes from multimodal data, durable execution, and trusted governance occurring together. Chuanxu keeps relational facts authoritative and composes full-text, vector, graph, and other capabilities as needed; adapters retain one product contract without requiring all three databases together.
AI 原生数据能力AI-native Data
关系、JSON、全文、向量与属性图，让不同形态的 Agent 上下文可共同检索和关联。Relational, JSON, full-text, vector, and property graph capabilities retrieve and connect different forms of Agent context.
企业级工程能力Enterprise Engineering
ACID、MVCC、索引、分区、存储过程、调度、文件与备份恢复，让长期运行的状态可正确处理。ACID, MVCC, indexes, partitioning, procedures, scheduling, files, and recovery keep long-running state correct.
可信运行能力Trusted Operations
身份、权限、隔离、审计、加密与高可用，为数据访问和故障处置建立可验证边界。Identity, authorization, isolation, audit, encryption, and availability establish verifiable access and failure boundaries.
AI 原生数据能力AI-native Data Capabilities
每种能力都解决不同的 Agent 问题Each Capability Solves a Different Agent Problem
| 能力Capability | 在 Agent 中的作用Role for Agents | 川序中的使用Use in Chuanxu |
|---|
| 关系与 SQLRelational and SQL | 表达权威事实、约束、状态和范围，支持精确过滤与一致更新。Represent authoritative facts, constraints, state, and scope for precise filtering and consistent updates. | 身份、组织、Agent 归属、任务、审批、审计与当前记忆指针。Identity, organization, Agent ownership, tasks, approval, audit, and current-memory pointers. |
| JSON | 保存模型输出、工具参数和演进中的上下文结构，避免为每次变化重建表结构。Store model output, tool parameters, and evolving context without rebuilding a schema for every change. | 规格、事件载荷、工具参数、候选整理结果与扩展属性。Specifications, event payloads, tool parameters, candidate consolidation results, and extensible properties. |
| 全文检索Full-text Search | 召回精确术语、错误码、实体名称与长文档中的关键表达。Recall exact terms, error codes, entity names, and key expressions in long documents. | 知识、记忆、Skill 内容、审计摘要和运行记录的词项检索。Term retrieval across knowledge, memory, Skills, audit summaries, and runtime records. |
| 向量检索Vector Retrieval | 发现语义相近但不使用相同词语的经验、知识或上下文。Find semantically similar experience, knowledge, or context expressed with different terms. | 长期记忆、知识检索与候选上下文排序。Long-term memory, knowledge retrieval, and candidate-context ranking. |
| 属性图Property Graph | 沿关系扩展上下文，分析依赖、协作、谱系和影响路径。Expand context along relationships and analyze dependencies, collaboration, lineage, and impact paths. | 知识关联、组织责任、任务依赖、Graph Engineering 与记忆谱系投影。Knowledge relationships, organization accountability, task dependencies, Graph Engineering, and memory-lineage projections. |
文件与空间能力可以作为特定场景扩展，但不是所有 Agent 场景的基础刚需。能力的价值不在堆叠，而在于能够在同一事务、授权与运维边界内组合使用。File and spatial capabilities can extend specific scenarios, but are not baseline requirements for every Agent workload. Value comes from composing capabilities within one transaction, authorization, and operations boundary, not from feature accumulation.
多模融合混合检索Multimodal Hybrid Retrieval
不是“只找相似”，而是按事实、文本、语义与关系共同决定上下文Not Just Similarity: Facts, Text, Semantics, and Relationships Select Context Together
这里的“多模”指关系、JSON 文档、全文、向量和图关系等数据与检索模态，不等同于图像或音频模型。融合检索先在数据库中缩小授权范围，再把不同信号组合排序，只将相关的 Top-N 上下文交给模型。Here, “multimodal” means relational, JSON-document, full-text, vector, and graph data and retrieval modes. It does not mean image or audio models. Hybrid retrieval first narrows the authorized database scope, then combines signals and sends only relevant Top-N context to the model.
授权范围与结构化过滤 → 全文精确召回 → 向量语义排序 → 图关系扩展 → 可解释的 Top-N 上下文Authorized scope and structured filters → full-text precision recall → vector semantic ranking → graph relationship expansion → explainable Top-N context
先过滤，后排序Filter Before Ranking
工作区、主体、数据分类、有效期和显式授权先定义候选范围，避免让语义相似度跨越安全边界。Workspace, principal, classification, validity, and explicit grants define the candidate scope before semantic ranking, preventing similarity from crossing a security boundary.
精确与语义互补Precision Meets Semantics
全文擅长固定术语和错误码；向量擅长近义表达。两者组合可减少只靠关键词漏召回或只靠相似度误召回。Full-text handles fixed terminology and error codes; vectors handle paraphrase. Together they reduce both keyword misses and similarity-only false matches.
关系补足上下文Relationships Complete Context
图可从命中的知识、任务或记忆扩展必要邻接关系，例如依赖、责任链和版本谱系。Graphs can extend necessary adjacent relationships from matched knowledge, tasks, or memory, such as dependencies, accountability, and version lineage.
结果可复核Results Can Be Reviewed
结构化条件、匹配信号与关系路径可以保留为查询和审计证据；图用于理解和检索扩展，不构成授权来源。Structured conditions, matching signals, and relationship paths can remain query and audit evidence. Graphs support understanding and retrieval expansion; they never grant authority.
企业级工程能力Enterprise Engineering
让长周期、多 Agent 工作保持正确而不是只完成一次调用Keep Long-running Multi-Agent Work Correct, Not Merely Invoked Once
ACID · MVCC · 约束Constraints
保证领取、状态切换、审批和记忆版本发布的一致性，避免重复执行、覆盖和错误回退。Keep claiming, state transitions, approval, and memory-version publication consistent, avoiding duplicate execution, overwrite, and invalid fallback.
索引 · 分区 · 存储过程Indexes · Partitioning · Procedures
支撑规模化检索、冷热数据组织和靠近数据的受控处理，减少在应用侧拼接状态。Support retrieval at scale, hot/cold data organization, and controlled processing near the data instead of assembling state in applications.
调度 · 租约 · 重试Scheduling · Leases · Retries
让持久作业在 Worker 异常后可重新领取、超时、重试和恢复，明确创建、入队与执行的边界。Let durable jobs be reclaimed, timed out, retried, and recovered after Worker failure, with clear boundaries between definition, enqueueing, and execution.
可信运行与运维Trusted Operations
安全、审计与恢复是 Agent 生产运行的基础能力Security, Audit, and Recovery Are Baseline Agent Production Capabilities
数据库身份与最小权限Database Identity and Least Privilege
独立身份、角色与行级/服务端策略在真实数据访问边界执行授权；提示词和 API 规范不能替代这一边界。Independent identities, roles, and row/server policy enforce authorization at the real data boundary; prompts and API contracts cannot replace it.
加密与审计证据Encryption and Audit Evidence
敏感配置受加密保护，关键访问、变更和副作用保留可查询、可复核的审计记录与引用。Sensitive configuration is encrypted, while key access, changes, and side effects retain queryable, reviewable audit records and references.
备份恢复与高可用Backup, Recovery, and Availability
Agent 核心状态在数据库中持久化，进程或节点异常后可由替代实例恢复；数据库自身的备份恢复和高可用设计避免底座不可访问。Core Agent state persists in the database and can be recovered by replacement instances after process or node failure; database backup, recovery, and availability design protects the foundation itself.
从能力到产品From Capability to Product
数据库能力在川序中形成一条管理闭环Database Capabilities Form One Management Loop in Chuanxu
| 产品目标Product Outcome | 数据库能力组合Database Combination | 形成的结果Result |
|---|
| 可观测Observable | 关系、JSON、全文、向量、图与审计Relational, JSON, full-text, vector, graph, and audit | 看见 Agent、上下文、协作关系、调用轨迹和证据。See Agents, context, collaboration relationships, activity trails, and evidence. |
| 可编排Orchestrated | 事务、约束、任务、租约、调度、图关系Transactions, constraints, jobs, leases, scheduling, and graph relationships | 让任务计划、分支、协作关卡、Loop 和 Graph Run 可持久化执行和调整。Persist and adjust task plans, branches, collaboration gates, Loops, and Graph Runs. |
| 可运维Operable | 身份、权限、加密、审计、备份恢复与高可用Identity, authorization, encryption, audit, backup/recovery, and availability | 在明确边界内控制访问、追溯操作并从异常恢复受管状态。Control access within explicit boundaries, trace operations, and recover managed state after failure. |
