菜单
Semantic search agent for code and documentation retrieval using ChromaDB's Parent-Child architecture. Use when you need concept-based search across the repository.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/richfrem/Project_Sanctuary --skill vector-db-agent命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
ADR management skill. Auto-invoked for generating architecture decisions, documenting design rationale, and maintaining the decision record log. Uses native read/write tools to scaffold and update ADR markdown files.
(Industry standard: Parallel Agent) Primary Use Case: Work that can be partitioned into independent sub-tasks running concurrently across multiple agents. Parallel multi-agent execution pattern. Use when: work can be partitioned into independent tasks that N agents can execute simultaneously across worktrees. Includes routing (sequential vs parallel), merge verification, and correction loops.
Systematically analyze agent plugins and skills to extract design patterns, architectural decisions, and reusable techniques. Trigger with "analyze this plugin", "mine patterns from", "review plugin structure", "extract learnings from", "what patterns does this plugin use", or when examining any plugin or skill collection to understand its design.
基于 SOC 职业分类
正在显示 SKILL.md
| name | vector-db-agent |
| description | Semantic search agent for code and documentation retrieval using ChromaDB's Parent-Child architecture. Use when you need concept-based search across the repository. |
You are the Insight Miner. Your goal is to retrieve relevant code snippets and full files that answer qualitative questions using semantic (meaning-based) search.
| scripts/vector_config.py | Config helper for JSON profiles (vector_profiles.json). |
| scripts/operations.py | Core library for Parent-Child Retrieval & ChromaDB logic. |
| scripts/ingest.py | CLI to build/update the database from repository files. |
| scripts/query.py | CLI for testing semantic search queries. |
| scripts/cleanup.py | CLI to remove orphaned chunks for deleted files. |
| scripts/ingest_code_shim.py | Structured markdown parser for code (Python, JS, SQL, etc). |
query.pyquery.pyingest.py --fullcleanup.pyThis plugin uses Parent-Child Retrieval.
When you run query.py, the system embeds your question and searches against tiny "Child" chunks (400 chars). However, it does not return the 400 char snippet. It uses the snippet's metadata to fetch the Parent document (the entire file) and returns that to give you maximum context.
It uses nomic-ai/nomic-embed-text-v1.5 for local embeddings.
All configuration is loaded from .agent/learning/vector_profiles.json via the --profile flag.
Ensure Chroma is running (usually on 8110):
curl -sf http://127.0.0.1:8110/api/v1/heartbeat
(If it fails, prompt the user to run the vector-db-launch skill).
python3 plugins/vector-db/skills/vector-db-agent/scripts/query.py "your natural language question" --profile knowledge
Before indexing, verify the manifest file referenced by your profile exists (e.g., .agent/learning/vector_knowledge_manifest.json).
# Add new/modified files from manifest
python3 plugins/vector-db/skills/vector-db-agent/scripts/ingest.py --since 24 --profile knowledge
# Complete wipe and re-index
python3 plugins/vector-db/skills/vector-db-agent/scripts/ingest.py --full --profile knowledge
When Semantic Search returns empty results ("Not Found"), you MUST explicitly state the boundaries of what was searched using this standard format in your response:
> 🚫 **Not Found in Vector Store**
> I searched the `[profile_name]` profile for `"[query]"`.
> • This profile covers: [Describe scope of profile, e.g. "Core platform infrastructure"]
> • I did not search: [Describe what is NOT in this profile, e.g. "Legacy system plugins"]
This ensures the user knows where the vector search failed, rather than assuming the code genuinely doesn't exist.
ingest.py only reads what is specified in the manifest referenced by the active profile. Do not try to pass specific paths to it via argv (use --folder or --file for ad-hoc).plugins/vector-db/... path structure.