Adaptive Rag Engine

Adaptive RAG 引擎 — 从线性检索到自主认知循环。集成胶囊预筛选、智能路由、CRAG纠错、L3校验。当需要搜索记忆/检索信息/回答复杂问题时触发。关键词：RAG、检索、记忆搜索、向量检索、Agentic RAG、CRAG。

Security Scan

The skill's files and instructions are consistent with an on‑system RAG orchestration layer that reads local memory and protocol files and provides helper scripts; nothing in the bundle requests unrelated credentials or hidden network/exfiltration behavior.

Purpose & Capability

Name/description (Adaptive RAG engine) match the provided artifacts: two small Python helper scripts (index builder and relevance evaluator) plus an instruction file describing routing, pre-filtering and evaluation. The skill claims integrations (vector search, bge‑m3, Phoenix memory bridge) but does not embed any external-client code — it acts as an orchestration/protocol layer and expects existing tools to provide vector search and memory storage. This design is coherent, though some high-level capabilities (e.g., bge‑m3 search, Active Memory ↔ Phoenix bridge) are described but not implemented in the included files — they are expected to exist elsewhere.

Instruction Scope

SKILL.md explicitly instructs the agent to read local protocol and memory files: rules/adaptive-rag-protocol.md and memory/topics/.capsule-index.json, and to run scripts under ~/.openclaw/workspace/skills/adaptive-rag-engine/scripts. Reading those files is expected for a RAG/navigation skill, but it is file-system access to user workspace data. Also SKILL.md states it is a "protocol layer" that "does not require explicit invocation" while registry flags show always:false — this is a minor inconsistency in intended activation semantics (not an immediate security problem, but worth noting).

Install Mechanism

No install spec (instruction-only plus two simple scripts) — low install risk. Scripts are small, readable Python files with no downloads or network calls. Minor manifest mismatch: registry metadata lists no required binaries, but the SKILL.md header declares python3 as required; python3 is reasonable for the provided scripts.

Credentials

The skill requests no environment variables or external credentials. The only resources accessed are local files under the user's OpenClaw workspace (~/.openclaw/workspace/memory and rules). That level of filesystem access is proportionate to the stated purpose of building/reading a capsule index and evaluating retrieval quality.

Persistence & Privilege

Skill is not marked always:true and does not request to modify other skills or system settings. Autonomous invocation (disable-model-invocation:false) is platform default; the SKILL.md's claim of acting as a protocol layer without explicit invocation contrasts with the actual flags but does not create extra privileges in the bundle itself.

Assessment

This skill appears to be a local orchestration/protocol layer for retrieval and is internally consistent, but before enabling you should: (1) Inspect the contents of rules/adaptive-rag-protocol.md and the files under ~/.openclaw/workspace/memory/topics to ensure they don't contain sensitive secrets you don't want the agent to read. (2) Note that vector search and memory bridging are not implemented here — the skill expects external components (e.g., a memory_search tool, bge‑m3 backend, Phoenix memory) to be present and may rely on their credentials; verify how those integrations are wired elsewhere. (3) Be aware of the SKILL.md claim that it is a protocol layer that 'does not require explicit invocation' — confirm how your platform will activate the skill (it is not always:on by default). (4) If you plan to run the included scripts, run them in a safe environment or inspect them (they are small and readable) — they only read/write files under your ~/.openclaw workspace and do not perform network I/O. If you need tighter assurance, request the full protocol/routing implementation or provenance for the vector/memory backends the skill expects.

🧠 Adaptive RAG Engine v1.0

不是管道，是认知循环。

核心能力

能力	说明	触发方式
Adaptive Router	判断是否需要检索，简单问题直接答	自动（每次 memory_search 前）
Capsule Pre-filter	42个胶囊标题预匹配，缩小范围 90%	自动（Router 通过后）
Vector Search	bge-m3 向量检索 Top-20	memory_search 工具
LLM Re-rank	语义重排序 Top-20 → Top-5	检索后自动
CRAG Evaluator	质量评估 + 低分补搜	检索后自动
L3 Gatekeeper	输出前与核心洞察校验	生成前自动
Memory Bridge	Active Memory ↔ Phoenix 双向桥接	对话结束时

使用流程

对于 CEO（小鸟文书）

此 Skill 是协议层，不需要显式调用。它通过以下方式生效：

1. 读取协议文件: rules/adaptive-rag-protocol.md — 获取完整决策树
2. 读取胶囊索引: memory/topics/.capsule-index.json — 获取 42 个胶囊元数据
3. 按决策树执行: 每次需要记忆时，走 Router → Pre-filter → Search → Rank → CRAG → Generate → Verify

对于 SubAgent

SubAgent 在执行任务时：

1. 先判断任务类型 → 决定是否需要检索
2. 需要检索时 → 先做胶囊预筛选
3. 检索结果 → 自行判断质量（CRAG 思维）
4. 输出前 → 自检是否与已知信息矛盾

关键文件

文件	用途
rules/adaptive-rag-protocol.md	完整协议（决策树/分类/bridge/CRAG/L3）
memory/topics/.capsule-index.json	42个胶囊的结构化索引
scripts/build-capsule-index.py	重建胶囊索引脚本
scripts/rag-evaluate.py	CRAG 质量评估脚本

快速命令

# 重建胶囊索引
python3 ~/.openclaw/workspace/skills/adaptive-rag-engine/scripts/build-capsule-index.py

# 评估检索质量
python3 ~/.openclaw/workspace/skills/adaptive-rag-engine/scripts/rag-evaluate.py --query "xxx" --results "result1, result2"

与其他 Skill 的关系

• dynamic-rag-capsule — Context 管理层面的胶囊化（对话太长时压缩）
• phoenix-memory — 记忆存储和衰减管理（四层架构）
• 本 Skill — 检索策略和质量控制（怎么搜、搜到后怎么办）

三者关系：phoenix-memory 是仓库，dynamic-rag-capsule 是打包器，adaptive-rag-engine 是导航仪