CASE-LEVEL KICKOFF · C0-001 · 2026-10-09
Reality → searchable context.本地向量化、结构化事实与可追溯来源。
Reality does not become a vector database. Selected observations produce derived representations for retrieval; the original Reality, factual records, ownership and execution authority remain outside that approximation.Reality 不被数据库取代。我们向量化的是获准观察中的可检索表达，不是把人的身份、物理事实、权限和现场状态都压成相似度。
Preserve the observation first.先保存观察，不先把现实压成向量。
Capture only the permitted input. Keep source kind, device/Runtime reference, capture time and uncertainty, units, coordinate frame, format and original content hash. A replay, a simulation and a physical sample never become interchangeable.
先保留原始图像、音频、测量或文档及来源、时间、单位、坐标与哈希。对内容的判断可以更新，原始来源不能被派生解释覆盖。
Derive text and structured facts.本地提取文本与结构化事实。
Use a specified parser, OCR/ASR or feature extractor where appropriate. Record preprocessing and model versions, parent hashes, uncertainty and test quality. Preserve exact numerical readings in structured fields rather than only prose.
有需要才进行 OCR、语音或特征提取；记录方法版本与父级哈希。数值、单位、时间及校准关系留在结构化记录中，不只留一句模型描述。
Embed selected, approved content.只向量化需要检索且获准使用的内容。
Chunk with source locators. Pin embedding model, preprocessing, dimension, metric, normalization, modality and language profile. CPU is a valid baseline; NPU acceleration is optional and requires an independently matched model/runtime.
按来源位置分片，固定嵌入模型、预处理、维度、距离方法与语言模态。本地向量化是待实现能力，不假定 T736 NPU 已能执行，也不要求全部原始数据都向量化。
Keep four distinct kinds of records.分开原件、事实、向量和工作记录。
Use durable originals; structured facts and lineage; rebuildable derived vectors; and authoritative identity/authority/work/operation records. A particular database engine is deliberately not selected by this directory.
原始对象库；结构化事实与来源；可重建向量索引；身份授权工作及执行记录。它们可共用基础存储，但不能混为同一种事实或权限。
Filter, retrieve, then read the source.先隔离权限及时效，再检索并回看来源。
Query and indexed vectors must belong to a compatible embedding space. Apply tenant, permission and freshness rules independently of similarity; check access again when retrieving source objects. Combine exact lookup, metadata and semantic retrieval as the work requires.
相似度不是权限、当前可用性、计量值或合规结论。读取原件时再次检查权限；无来源、过期或模型空间不匹配时拒绝或重建，不静默混用。
Turn context into a proposal—not automatic action.检索上下文先支持建议，不直接驱动执行。
The Host assembles the permitted context for a model or human. Retrieved documents remain untrusted evidence, not instructions that can expand tool privileges. Destination policy, current local state and the accepted work decide whether an action can proceed.
资料与传感内容是证据，不能夹带指令扩大权限。另一现场重新核对任务、授权、当前状态、预算及占用后，才能进入有界执行。
Observe the new Reality and retain lineage.对新 Reality 再感知，并接回连续记录。
Tie an independent observation or human receipt to the operation and original work. Distinguish accepted results, failed attempts and uncertain effects. Revoke or rebuild stale indexes when source versions, permissions or retention change.
失败、未知效果与通过均保留自己的状态。原始资料、权限、版本或保存期变化后，原件与向量索引要同步撤销、失效或重建。
Portable data contracts / 可交接的数据契约
The example is a labeled synthetic text fixture, not a captured sensor reading. No synthetic numerical vectors are presented as real embeddings.
| Required relationship | Why it matters |
|---|---|
| observation_id → original locator + SHA256 | Read the same source again / 回看相同来源 |
| transformation → parent hash + method/revision | Explain and rebuild a derived claim / 解释并重建派生表达 |
| embedding-space identity → query + index | Avoid cross-model dimension/semantic mismatch / 避免模型空间混用 |
| tenant + scope + expiry → retrieved source | Prevent unauthorized or stale context / 避免越权及过期上下文 |
| operation → effect observation → acceptance | Keep a proposed action distinct from a real outcome / 分开建议与现实结果 |
External technical basis / 外部技术依据
Microsoft documents vector and nonvector fields within an index and metadata-based filtering. These references support this design distinction; they do not commit HSIMC to Azure AI Search or prove the local node has implemented it.
Continue through HSIMC回到同一个 HSIMC 工作与现实架构
This public directory defines reference components, inputs, outputs and acceptance targets. It does not grant execution authority, bind a Runtime, create a vector database, activate an MCP endpoint, reserve a supplier or accept an order. 目录不等于已验证的整套系统；当前实例、权限、可用时间和交付均需各自证据。