The Librarian

Build and search lightweight quantized document indexes with TurboVec. Use when you need to create searchable indexes from documents for RAG applications wit...

This skill appears coherent for building/searching TurboVec quantized indexes, but before installing: 1) Confirm where embeddings are sent — the scripts default to http://host.docker.internal:11434 (an Ollama-style local endpoint). If you run the tool, embedding text will be POSTed to whichever API URL you supply; point it to a trusted local service or a trusted remote provider. 2) SKILL.md/help mention an OLLAMA_API env var but the scripts use a default and accept --api — set the --api flag or edit the code if you need a different endpoint. 3) Run the code in an isolated environment (dedicated venv/container) when indexing sensitive documents. 4) Review and vet third-party packages (turbovec, flashrank, rank-bm25) before pip installing them. 5) For high-risk documents (medical, legal, financial) follow the author's own advice and use a higher-accuracy/approved setup (e.g., FAISS) or ensure your embedding provider and runtime are fully trusted.

Type: OpenClaw Skill Name: thelibrarian Version: 1.0.1 The skill bundle implements a legitimate document indexing and search system but contains a critical security vulnerability. Specifically, 'scripts/search.py' uses 'pickle.load()' to deserialize the BM25 index from the disk, which allows for Remote Code Execution (RCE) if the agent is directed to search a malicious or untrusted index directory. Additionally, 'scripts/librarian.sh' utilizes 'LD_PRELOAD' to load the BLAS library; while this is documented as a requirement for the 'turbovec' dependency, it is a high-risk mechanism that could be abused. No evidence of intentional malice, data exfiltration, or harmful prompt injection was found.