polygres-retrieval-design

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Polygres Retrieval Design

Polygres检索设计

Produce a reviewable plan before retrieval configuration or application work. This skill is advisory: it must not mutate a project directly.

在检索配置或应用开发前生成可评审的计划。本技能仅提供咨询建议：不得直接修改项目。

Workflow

工作流程

Extract the user outcome, authorization, latency, freshness, and expected result shape from the prompt and inspected application. Ask only for a missing fact that changes the recommendation; representative questions are useful evidence, not a mandatory interview.
Resolve project mode, then inspect the supplied schema, verified row identifiers, sample data, and existing retrieval configuration. Label missing evidence as unresolved; never infer production facts from a table or column name.
Select the smallest sufficient strategy using
```
references/strategy-selection.md
```
. Reject an unsupported strategy rather than inventing a capability.
For graph retrieval, apply
```
references/graph-modeling.md
```
. For an existing vector configuration, TSVector, or fuzzy retrieval, apply
```
references/vector-and-text-design.md
```
.
For pgContext collections, point synchronization, registered filters, or Context retrieval modes, apply
```
references/context-design.md
```
.
For multi-stage retrieval or RAG, apply
```
references/hybrid-and-rag-plan.md
```
.
Write the result with
```
references/plan-template.md
```
. When called by
```
$polygres-data-pipeline
```
, return the selected and omitted components plus exact public-interface handoffs in a machine-readable section so the caller can continue without another interview.
In a design-only request, stop before mutation. In an active data-pipeline setup, return control to the orchestrating skill; its consolidated review and approval govern the implementation.

从提示信息和待检查应用中提取用户目标、权限、延迟要求、数据新鲜度以及预期结果形态。仅询问会影响推荐结果的缺失信息；代表性问题可作为有用依据，但无需进行强制性访谈。
确定项目模式，然后检查提供的 schema、已验证的行标识符、样本数据和现有检索配置。将缺失的依据标记为未解决；切勿从表名或列名推断生产环境事实。
使用
```
references/strategy-selection.md
```
选择最小且足够的策略。对于不支持的策略直接拒绝，而非凭空创造能力。
针对图检索，应用
```
references/graph-modeling.md
```
。对于现有向量配置、TSVector或模糊检索，应用
```
references/vector-and-text-design.md
```
。
针对pgContext集合、点同步、已注册过滤器或Context检索模式，应用
```
references/context-design.md
```
。
针对多阶段检索或RAG，应用
```
references/hybrid-and-rag-plan.md
```
。
使用
```
references/plan-template.md
```
撰写结果。当被
```
$polygres-data-pipeline
```
调用时，在机器可读部分返回已选和未选组件以及精确的公共接口交接内容，以便调用方无需再次访谈即可继续操作。
在仅设计的请求中，在修改操作前停止。在活跃的数据管道设置中，将控制权交还给编排技能；其综合评审和批准将指导实施工作。

Design rules

设计规则

Prefer relational retrieval for exact predicates, joins, aggregates, and transactions that do not need a retrieval index.
For a synced project, keep exact SQL, transactions, and source mutations in the source PostgreSQL database. Use only the supported Runtime retrieval and retrieval-configuration surfaces on Polygres.
Treat graph, vector, text, and Context indexes and point mappings as derived project resources whose readiness and refresh behavior must be validated.
Use exact schema identifiers and stable row ID values from verified data. Do not use invented row IDs or fuzzy-match schema names.
Bound graph direction, depth, fan-out, result count, and cycle behavior.
Record the embedding model, dimensions, metric, input construction, and response to a dimension mismatch or empty embedding.
Default new semantic retrieval plans to a Polygres AI Context collection. Decide explicitly whether distinct embeddings belong as named vectors in one collection or require separate collection-level source and policy settings. Record any existing pgvector configuration that must remain compatible or needs a migration plan. Never treat those resources as interchangeable. On a synced project, use only an existing synchronized source table and embedding column; do not plan
```
add_column
```
or
```
new_table
```
on the target.
State TSVector language/configuration choices and fuzzy thresholds.
For hybrid retrieval, define stage order, provenance, deduplication, authorization, and token budget.
Include rebuild or reindex triggers, readiness checks, and rollback or fallback behavior.

对于不需要检索索引的精确谓词、连接、聚合和事务，优先使用关系型检索。
对于同步项目，将精确SQL、事务和源修改保留在源PostgreSQL数据库中。仅在Polygres上使用受支持的Runtime检索和检索配置界面。
将图、向量、文本和Context索引以及点映射视为派生项目资源，必须验证其就绪状态和刷新行为。
使用来自已验证数据的精确schema标识符和稳定行ID值。不得使用虚构的行ID或模糊匹配的schema名称。
限制图的方向、深度、扇出、结果数量和循环行为。
记录embedding模型、维度、度量标准、输入构造方式，以及对维度不匹配或空embedding的处理方式。
新的语义检索计划默认使用Polygres AI Context集合。明确决定不同的embedding应作为命名向量放在一个集合中，还是需要单独的集合级源和策略设置。记录任何必须保持兼容或需要迁移计划的现有pgvector配置。切勿将这些资源视为可互换的。在同步项目中，仅使用现有的同步源表和embedding列；不得在目标端计划
```
add_column
```
或
```
new_table
```
操作。
说明TSVector语言/配置选择和模糊阈值。
对于混合检索，定义阶段顺序、来源、去重、权限和token预算。
包含重建或重新索引触发器、就绪检查以及回退或 fallback 行为。

Boundaries

边界限制

This skill must not mutate a project, generate secrets, or claim that a plan has been applied.
Do not write mutating command examples. Route approved configuration work to
```
$polygres-cli
```
and approved Python integration to
```
$polygres-sdk
```
.
Filters are not an authorization boundary. Apply access control before data enters retrieval and again when results are resolved.
If required columns, stable IDs, or a usable sample are absent, make a reversible provisional recommendation when possible. Stop only when the missing evidence makes every safe recommendation invalid.

本技能不得修改项目、生成密钥或声称计划已被应用。
不得编写修改性命令示例。将已批准的配置工作路由到
```
$polygres-cli
```
，已批准的Python集成路由到
```
$polygres-sdk
```
。
过滤器不是权限边界。在数据进入检索前应用访问控制，并在结果解析时再次应用。
如果缺少必要的列、稳定ID或可用样本，尽可能做出可逆的临时建议。仅当缺失的依据导致所有安全建议都无效时才停止操作。

Completion

完成要求

Return the completed reviewable plan, the unresolved assumptions, the recommended public-surface handoffs, and the explicit approval needed before any mutation or implementation.

返回已完成的可评审计划、未解决的假设、推荐的公共界面交接内容，以及在进行任何修改或实施前所需的明确批准。