RAG Architect
RAG Architect is a skill for designing, tuning, and evaluating production RAG pipelines. It provides three deterministic tools to analyze a corpus, design a pipeline from requirements, and evaluate retrieval quality, ensuring decisions are data-driven and every design ends with an evaluation run.
✨ What it does
- Runs chunking_optimizer.py to analyze the corpus and recommend a chunking strategy with sample chunks.
- Generates a complete pipeline design (chunking, embedding, vector DB, retrieval, reranking, evaluation) from a requireme
- Evaluates retrieval quality with retrieval_evaluator.py, reporting precision@k, recall@k, MRR, NDCG@k, and poor-performa
- Enforces a verification loop: designs must meet evaluation targets, and changes are made one variable at a time.
- Enforces pricing discipline: model names and prices are presented as tiers with placeholders to verify against live pric
🎯 When to use it
- When asked to design a RAG pipeline for a document corpus.
- When needing to choose a chunking strategy or embedding model.
- When selecting a vector database for a RAG system.
- When evaluating retrieval quality against ground truth (precision@k, recall@k, NDCG).
🚀 How to use
Trigger by asking to design or evaluate a RAG system, e.g., 'design a RAG pipeline for our docs' or 'evaluate my retriever against ground truth'. Provide the path to your documents and, for design, a requirements JSON with keys like document_types, document_count, queries_per_day, latency_requirement, budget_monthly, accuracy_priority, cost_priority, maintenance_complexity. For evaluation, provide queries.json and ground_truth.json. Run the tools in order: chunking_optimizer.py, rag_pipeline_designer.py, retrieval_evaluator.py.
Example prompts:
Design a RAG system for our technical docs. Requirements are in requirements.json.
Evaluate my retriever using queries.json and ground_truth.json on the docs in /path/to/docs.
📄 Output: The skill produces chunking.json, design.json (with architecture diagram and config templates), and eval.json (retrieval metrics and failure examples).
📦 Add this skill to Claude Code
# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills
# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering/skills/rag-architect .claude/skills/rag-architectSkill source: engineering/skills/rag-architect/SKILL.md
⚠️ Good to know
Model names and prices are not current facts; users must verify against live pricing pages. The skill is not for general LLM cost tuning or agent loops over retrieval.
❓ FAQ
What does the skill do if retrieval quality is below target?
It instructs you to inspect the poor-example lists, change one variable (chunking, embedding tier, reranking, or hybrid retrieval), and re-run the evaluation until targets are met.
How does the skill handle embedding model selection?
It recommends a tier (fast/self-hosted, balanced open, quality API, code) with current-generation examples, but requires you to verify model availability and pricing against the provider's live page.
What inputs are required for the pipeline designer?
A requirements JSON with all required keys: document_types, document_count, avg_document_size, queries_per_day, query_patterns, latency_requirement, budget_monthly, accuracy_priority, cost_priority, and maintenance_complexity.
🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.