ivyxivyx
← All skills
IVYXSTUDIO · SKILL
R

RAG Knowledge Base Setup

rag-knowledge-base-setup · v1.0.0

ivyx

Use when someone wants a knowledge base (bilgi tabanı), RAG or document search over files in the project (docs, notes, manuals) so an assistant can answer questions from them; builds, checks and queries it.

Who is asked

Risk medium: you are asked when it loads, and that yes covers reading its files in the same chat.

Tools it may call

While this skill is active, a call to any other tool is refused. These still meet every policy rule; the list narrows, it approves nothing.

knowledge.base.listknowledge.base.upsertknowledge.base.ingestknowledge.queryknowledge.chunk.comparevectors.embedder.status

What it tells the agent

The published SKILL.md, as written.

---
name: rag-knowledge-base-setup
description: Use when someone wants a knowledge base (bilgi tabanı), RAG or document search over files in the project (docs, notes, manuals) so an assistant can answer questions from them; builds, checks and queries it.
allowed-tools: knowledge.base.list knowledge.base.upsert knowledge.base.ingest knowledge.query knowledge.chunk.compare vectors.embedder.status
metadata:
  risk: medium
---

Building a knowledge base

The point is an assistant that answers from the base. So you do not read the documents yourself to answer: every answer you give comes from knowledge.query, and you build the base before you query it.

1. Look at what exists

  • knowledge.base.list: the bases already defined and what their indexes hold. Extend an existing base rather than defining a second one over the same files.
  • vectors.embedder.status: which embedder builds the index. When source is hash, tell the person, in your own words, what its nextStep says: this embedder matches shared words, not meaning, and how to switch to a local model. Then go on; the base still works for words the question and the documents share.

2. Define the base

knowledge.base.upsert with:

  • id: lower-case letters, digits and dashes (company-docs);
  • include: globs, not a list of files (docs/**/*.md); ** crosses folders and * does not; exclude for what should stay out;
  • description: what the files are, read from their names and folders. Never name a topic you have not seen in a path.

When the documents are long or have many headings, call knowledge.chunk.compare with one representative file and pass the strategy whose chunks keep a heading with its text as strategy.

3. Check, then build

  1. knowledge.base.ingest with the base and dryRun: true. It lists the paths the globs resolve to. If a file the person meant is missing, or one they did not mean is there, fix include or exclude first.
  2. knowledge.base.ingest with the base. Read files and chunks from the answer and say them.

4. Ask it

knowledge.query with the base and the person's question, in their words.

  • Answer only from hits, and name each hit's path.
  • When no hit holds the answer, say the base does not answer it and which files you expected to; do not answer from memory.
  • Report the mode the answer came back in, and with the hash embedder say that a reworded question can miss.

End with the base id, its files and chunks, the embedder, and the answer with its source.