Bm25 Rerank Memory

本地 BM25 + Embedding + Rerank 混合记忆检索。检索 /root/workspace/Remember 下的 Markdown 记忆文件，支持关键词+语义+重排。每次记忆搜索用它替代 mem0。

Security Scan

The skill's stated purpose (local BM25+embedding+rerank over a local 'Remember' folder) is plausible, but the runtime instructions reference external embedding/reranker APIs without declaring credentials, suggest a risky pip install flag, and depend on a local /opt script and /root files that are not bundled — these mismatches warrant caution.

Purpose & Capability

The skill claims to perform local BM25+embedding+rerank retrieval over /root/workspace/Remember, which matches the provided instructions to call a local script. However, the SKILL.md also relies on SiliconFlow / BAAI models for embeddings/reranking (networked services) but declares no credentials or network endpoints; that is an unexplained capability mismatch.

Instruction Scope

Runtime instructions tell the agent to execute /opt/memory/bm25_rerank_search.py and read/search markdown files under /root/workspace/Remember. Because the skill is instruction-only and provides no code, it assumes those files and scripts exist and will be executed/read — this gives the skill the ability to surface arbitrary local file contents and to run arbitrary code in /opt if present.

Install Mechanism

There is no formal install spec, but the SKILL.md recommends running `pip install rank-bm25 --break-system-packages`. That flag is invasive (can circumvent OS package manager protections) and is risky to run system-wide; even though rank-bm25 is reasonable for BM25, recommending --break-system-packages is disproportionate and potentially harmful.

Credentials

The instructions depend on external embedding/reranker services (SiliconFlow / BAAI models) but the skill declares no required environment variables or credentials. Either credentials are expected to be configured out-of-band (not declared) or the skill will fail/attempt unknown network access — both are problematic. It also references system paths (/opt and /root) without declaring config path requirements.

Persistence & Privilege

The skill does not request always:true or other elevated platform privileges. It is user-invocable and can be invoked autonomously per platform default, which is normal. The main risk is execution of a local script and reading /root files (described above), not elevated registration.

What to consider before installing

This skill expects and will execute a local script (/opt/memory/bm25_rerank_search.py) and read files under /root/workspace/Remember — verify those files/scripts are trustworthy before using. Confirm how the SiliconFlow/BAAI embedding and reranker APIs are authenticated (there are no env vars declared); if API keys are needed, ensure they are supplied securely and not exposed. Do not run the suggested `pip install ... --break-system-packages`; instead install dependencies in a virtualenv or container. If you cannot validate the /opt script or the origin of any network calls, run this in an isolated environment or avoid installing/using it.

BM25 Rerank Memory Search

调用本地检索脚本 /opt/memory/bm25_rerank_search.py，返回精确的上下文片段。

依赖

• BM25: rank_bm25
• 向量+重排: SiliconFlow API (BAAI/bge-m3 + BAAI/bge-reranker-v2-m3)
• 安装: pip install rank-bm25 --break-system-packages

使用方式

搜索记忆

python3 /opt/memory/bm25_rerank_search.py search <查询词>

重建索引（记忆文件变更后需执行）

python3 /opt/memory/bm25_rerank_search.py reindex

检索逻辑

1. BM25 关键词匹配 → 候选集 A
2. BGE-M3 向量相似度 → 候选集 B
3. A ∪ B 合并
4. BGE-Reranker-v2-m3 重排 → 最终结果

输出格式

每条结果包含: [score] path:chunk_id + 内容片段 (最多显示前200字符)