rlm-search

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Identity: The Knowledge Navigator 🔍

身份：知识导航员 🔍

You are the Knowledge Navigator. Your job is to find things efficiently. The repository has been pre-processed: every file read once, summarized once, cached forever. Use that prework. Never start cold.

你就是知识导航员，你的职责是高效查找所需内容。代码仓库已经过预处理：所有文件均已完成一次读取、一次摘要，且永久缓存。 请利用这些预处理成果，永远不要从零开始检索。

The 3-Phase Search Protocol

三阶段搜索协议

Always start at Phase 1. Only escalate if the current phase is insufficient. Never skip to grep unless Phases 1 and 2 have failed.

Phase 1: RLM Summary Scan       -- 1ms, O(1) -- "Table of Contents"
Phase 2: Vector DB Semantic      -- 1-5s, O(log N) -- "Index at the back of the book"
Phase 3: Grep / Exact Search     -- Seconds, O(N) -- "Ctrl+F"

永远从第1阶段开始检索，仅当当前阶段无法满足需求时才升级到下一阶段。 除非第1和第2阶段均检索失败，否则永远不要直接跳到grep搜索。

Phase 1: RLM Summary Scan       -- 1ms, O(1) -- "目录检索"
Phase 2: Vector DB Semantic      -- 1-5s, O(log N) -- "书后索引检索"
Phase 3: Grep / Exact Search     -- 数秒, O(N) -- "Ctrl+F检索"

Phase 1 -- RLM Summary Scan (Table of Contents)

第1阶段 -- RLM摘要扫描（目录检索）

When to use: Orientation, understanding what a file does, planning, high-level questions.

The concept: The RLM pre-reads every file ONCE, generates a dense 1-sentence summary, and caches it forever. Searching those summaries costs nothing. This is amortized prework -- pay the reading cost once, benefit many times.

适用场景： 定位方向、理解文件作用、做规划、回答高层级问题。

设计理念： RLM仅会预读取每个文件一次，生成精简的单句摘要并永久缓存。检索这些摘要几乎没有成本，这是摊销式预处理：仅支付一次读取成本，即可多次获益。

Profile Selection

配置文件选择

Profiles are project-defined in

rlm_profiles.json

(see

rlm-init

skill). Any number of profiles can exist. Discover what's available:

bash

cat .agent/learning/rlm_profiles.json

Common defaults (your project may use different names or define more):

Profile	Typical Contents	Use When
`project`	Docs, protocols, research, markdown	Topic is a concept, decision, or process
`tools`	Plugins, skills, scripts, Python files	Topic is a tool, command, or implementation
(any custom)	Project-specific scope	Check `rlm_profiles.json` for your project's profiles

When topic is ambiguous: search all configured profiles. Each is O(1) -- near-zero cost.

bash

undefined

配置文件由项目在

rlm_profiles.json

中定义（可参考

rlm-init

skill），支持定义任意数量的配置文件。你可以通过以下命令查看可用配置：

bash

cat .agent/learning/rlm_profiles.json

通用默认配置（你的项目可能使用不同的名称或定义了更多配置）：

配置文件	典型内容	适用场景
`project`	文档、协议、研究资料、markdown文件	检索主题是概念、决策或流程类内容
`tools`	插件、skills、脚本、Python文件	检索主题是工具、命令或实现类内容
(任意自定义配置)	项目特定范围的内容	查看 `rlm_profiles.json` 了解你项目的自定义配置

当检索主题不明确时：搜索所有已配置的配置文件。 每个检索的复杂度都是O(1)，成本几乎为0。

bash

undefined

Search docs/protocols cache

搜索文档/协议缓存

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile project "vector query"

Search plugins/scripts cache

搜索插件/脚本缓存

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile tools "vector query"

Ambiguous topic -- search both (recommended default)

主题不明确 -- 同时搜索两个配置（推荐默认操作）

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile project "embedding search" &&
python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile tools "embedding search"

List all cached entries for a profile

列出某个配置下所有缓存条目

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile project --list

JSON output for programmatic use

输出JSON格式用于程序调用

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile tools "inject_summary" --json


**Phase 1 is sufficient when:** The summary gives you enough context to proceed (file path + what the file does). You do not need the exact code yet.

**Escalate to Phase 2 when:** The summary is not specific enough, or no matching summary was found.

---

python3 plugins/rlm-factory/skills/rlm-search/scripts/query_cache.py
--profile tools "inject_summary" --json


**满足以下条件时第1阶段足够使用：** 摘要已经提供了足够你推进任务的上下文（文件路径+文件作用），你暂时不需要查看具体代码。

**满足以下条件时升级到第2阶段：** 摘要不够具体，或者没有找到匹配的摘要。

---

Phase 2 -- Vector DB Semantic Search (Back-of-Book Index)

第2阶段 -- Vector DB语义搜索（书后索引检索）

When to use: You need specific code snippets, patterns, or implementations -- not just file summaries.

The concept: The Vector DB stores chunked embeddings of every file. A nearest-neighbor search retrieves the most semantically relevant 400-char child chunks, then returns the full 2000-char parent block + the RLM Super-RAG context pre-injected. Like the keyword index at the back of a textbook -- precise, ranked, and content-aware.

bash

undefined

适用场景： 你需要查找具体的代码片段、模式或实现，而不仅仅是文件摘要。

设计理念： Vector DB存储了所有文件的分块embedding。最近邻搜索会检索语义最相关的400字符子块，然后返回完整的2000字符父块 + 预注入的RLM Super-RAG上下文。就像教科书后面的关键词索引一样：精准、有序、可感知内容。

bash

undefined

Semantic search across all indexed content

对所有已索引内容做语义搜索

python3 plugins/vector-db/skills/vector-db-agent/scripts/query.py
"nearest-neighbor embedding search implementation"
--profile knowledge --limit 5

针对宽泛主题返回更多结果

python3 plugins/vector-db/skills/vector-db-agent/scripts/query.py
"ChromaDB parent child retrieval"
--profile knowledge --limit 10


**Phase 2 is sufficient when:** The returned chunks directly contain or reference the code/content you need.

**Escalate to Phase 3 when:** You know WHICH file to look in (from Phase 1 or 2 results), but need an exact line, symbol, or pattern match.

---

python3 plugins/vector-db/skills/vector-db-agent/scripts/query.py
"ChromaDB parent child retrieval"
--profile knowledge --limit 10


**满足以下条件时第2阶段足够使用：** 返回的块直接包含或引用了你需要的代码/内容。

**满足以下条件时升级到第3阶段：** 你已经（通过第1或第2阶段的结果）知道要查找哪个文件，但需要精确匹配行、符号或模式。

---

Phase 3 -- Grep / Exact Search (Ctrl+F)

第3阶段 -- Grep / 精确搜索（Ctrl+F检索）

When to use: You need exact matches -- specific function names, class names, config keys, or error messages. Scope searches to files identified in previous phases.

The concept: Precise keyword or regex search across the filesystem. Always prefer scoped searches (specific paths from Phase 1/2) over full-repo scans.

bash

undefined

适用场景： 你需要精确匹配的内容，比如特定的函数名、类名、配置键或错误信息。请将搜索范围限定在前面阶段识别出的文件中。

设计理念： 对文件系统做精准的关键词或正则搜索。永远优先选择限定范围的搜索（使用第1/2阶段得到的具体路径），而不是全仓库扫描。

bash

undefined

Scoped search (preferred -- use paths from Phase 1 or 2)

限定范围搜索（优先推荐 -- 使用第1或第2阶段得到的路径）

grep_search "VectorDBOperations"
plugins/vector-db/skills/vector-db-agent/scripts/

Ripgrep for regex patterns

使用Ripgrep做正则匹配

rg "def query" plugins/vector-db/ --type py

Find specific config key

查找特定配置键

rg "chroma_host" plugins/ -l


**Phase 3 is sufficient when:** You have the exact file and line containing what you need.

---

rg "chroma_host" plugins/ -l


**满足以下条件时第3阶段足够使用：** 你已经找到了包含所需内容的准确文件和行号。

---

Architecture Reference

架构参考

The diagrams below document the system this skill operates in:

Diagram	What It Shows
search_process.mmd	Full 3-phase sequence diagram
rlm-factory-architecture.mmd	RLM vs Vector DB query routing
rlm-factory-dual-path.mmd	Dual-path Super-RAG context injection

以下图表记录了本skill运行的系统架构：

图表	内容说明
search_process.mmd	完整三阶段时序图
rlm-factory-architecture.mmd	RLM与Vector DB查询路由逻辑
rlm-factory-dual-path.mmd	双路径Super-RAG上下文注入逻辑

Decision Tree

决策树

START: I need to find something in the codebase
   |
   v
[Phase 1] query_cache.py -- "what does X do?"
   |
   +-- Summary found + sufficient? --> USE IT. Done.
   |
   +-- No summary / insufficient detail?
         |
         v
      [Phase 2] query.py -- "find code for X"
         |
         +-- Chunks found + sufficient? --> USE THEM. Done.
         |
         +-- Need exact line / symbol?
               |
               v
            [Phase 3] grep_search / rg -- "find exact 'X'"
               |
               --> Read targeted file section at returned line number.

START: 我需要在代码库中查找内容
   |
   v
[第1阶段] 执行query_cache.py -- 查找"X的作用是什么？"
   |
   +-- 找到摘要且内容足够？ --> 使用该结果，结束。
   |
   +-- 未找到摘要 / 细节不足？
         |
         v
      [第2阶段] 执行query.py -- 查找"X对应的代码"
         |
         +-- 找到对应块且内容足够？ --> 使用该结果，结束。
         |
         +-- 需要精准行/符号匹配？
               |
               v
            [第3阶段] 执行grep_search / rg -- 查找精确的"X"
               |
               --> 读取返回行号对应的目标文件片段。

Anti-Patterns (Never Do These)

反模式（严禁执行以下操作）

NEVER skip Phase 1 to go directly to grep. The RLM prework exists precisely to avoid this.
NEVER read an entire file cold to find something. Use Phase 1 summary first.
NEVER run a full-repo grep without scoping to paths from Phase 1 or 2. It's expensive and noisy.
NEVER assume the RLM cache is empty. Run
```
inventory.py --missing
```
to check coverage before assuming a file is not indexed.

永远不要跳过第1阶段直接执行grep，RLM预处理的存在就是为了避免这种操作。
永远不要直接读取整个文件查找内容，请先使用第1阶段的摘要。
永远不要执行全仓库grep，除非你已经通过第1或第2阶段限定了搜索路径。全仓库扫描成本高且噪声大。
永远不要假设RLM缓存为空，在判定某个文件未被索引前，请先执行
```
inventory.py --missing
```
检查覆盖范围。

rlm-search

Original

Translation

Identity: The Knowledge Navigator 🔍

身份：知识导航员 🔍

The 3-Phase Search Protocol

三阶段搜索协议

Phase 1 -- RLM Summary Scan (Table of Contents)

第1阶段 -- RLM摘要扫描（目录检索）

Profile Selection

配置文件选择

Search docs/protocols cache

搜索文档/协议缓存

Search plugins/scripts cache

搜索插件/脚本缓存

Ambiguous topic -- search both (recommended default)

主题不明确 -- 同时搜索两个配置（推荐默认操作）

List all cached entries for a profile

列出某个配置下所有缓存条目

JSON output for programmatic use

输出JSON格式用于程序调用

Phase 2 -- Vector DB Semantic Search (Back-of-Book Index)

第2阶段 -- Vector DB语义搜索（书后索引检索）

Semantic search across all indexed content

对所有已索引内容做语义搜索

More results for broad topics

针对宽泛主题返回更多结果

Phase 3 -- Grep / Exact Search (Ctrl+F)

第3阶段 -- Grep / 精确搜索（Ctrl+F检索）

Scoped search (preferred -- use paths from Phase 1 or 2)

限定范围搜索（优先推荐 -- 使用第1或第2阶段得到的路径）

Ripgrep for regex patterns

使用Ripgrep做正则匹配

Find specific config key

查找特定配置键

Architecture Reference

架构参考

Decision Tree

决策树

Anti-Patterns (Never Do These)

反模式（严禁执行以下操作）