skills.rest
kiurakku
kiurakkuCOMMUNITY
Explore all skills by this author →
𝕏@thefear078GitHub

rag-review

Diagnose and design retrieval-augmented generation pipelines across chunking, retrieval, and generation stages.

15|3|Updated Jul 9, 2026
Search Tags:#embeddings#rag#llm evaluation#hybrid search#chunking#reranking#retrieval augmented generation
One-click install
npx skills add https://github.com/kiurakku/cursor-kit-for-ai --skill rag-review-kiurakku
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: rag-review
Source: https://github.com/kiurakku/cursor-kit-for-ai/tree/main/plugins/ai/skills/rag-review
Command: npx skills add https://github.com/kiurakku/cursor-kit-for-ai --skill rag-review-kiurakku

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? RAG systems fail in opaque ways: answers cite wrong sources, hallucinate facts, or go stale, and teams blame the LLM when the real bug is in retrieval or chunking. This Skill provides a structured method to bisect the pipeline, measure retrieval separately from generation, and fix the actual failure stage. ## Core Features & Use Cases - Pipeline design guidance: Covers structure-aware chunking, metadata preservation, hybrid dense + BM25 retrieval with reranking, and grounded generation with verifiable citations. - Debugging decision table: Maps observed symptoms (missing answers, mangled chunks, stale results) to the failing stage and its fix zone. - Evaluation framework: Builds a gold question set and measures retrieval recall@k and generation faithfulness independently, with a review checklist for existing systems. - Use Case: Your support chatbot answers questions about refund policy incorrectly. Use this Skill to inspect retrieved chunks, discover retrieval recall is 60%, and fix hybrid search and query rewriting before touching prompts. ## Quick Start Ask the AI to review your RAG pipeline and diagnose why answers cite wrong or missing sources.

Frequently Asked Questions about rag-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug a RAG system that gives wrong answers?▼

Inspect the retrieved chunks before blaming the model. If the right answer is not in the retrieved chunks, it is a retrieval failure; if the right chunks are retrieved but the answer is wrong, it is a generation failure. Measure retrieval recall@k and generation faithfulness separately with a gold question set.

When should I use RAG vs long-context vs fine-tuning?▼

If the corpus fits in the context window (roughly 100-200K tokens) and queries are infrequent, include it directly. Use RAG for current or private data at scale. Use fine-tuning for style, format, or domain behavior, not for facts that change.

What chunking strategy works best for RAG?▼

Chunk by document structure such as headings, paragraphs, and code blocks rather than fixed character counts, typically 200-800 tokens with small overlap. Preserve metadata like source, section path, and date per chunk, and prepend a context line so each chunk is interpretable standalone.

Why does my RAG system miss exact error codes or product names?▼

Pure dense (semantic) retrieval misses exact identifiers like error codes, SKUs, and names. Use hybrid retrieval combining dense embeddings with BM25 keyword search fused via RRF, then rerank the top 20-50 candidates with a cross-encoder.

How do I prevent RAG hallucinations?▼

Instruct the model to answer only from provided chunks, cite chunk ids per claim, and use an explicit "not in the provided context" escape hatch when retrieval is weak. Also delimit chunks as data and test that poisoned documents cannot inject instructions.

How do I evaluate RAG retrieval quality?▼

Build a gold set of 30-100 real questions with the expected answer and source chunk, then measure recall@k and MRR for retrieval independently. If retrieval recall is below roughly 85%, fix retrieval before tuning generation prompts.

Related Skills

Explore More Software Engineering Skills→
hiivmindhiivmind

hiivmind-corpus-refresh

Refresh and sync a local corpus index with upstream documentation sources.

Official
Intermediate
kinnerkarmanishkinnerkarmanish

eval-rag

Evaluate RAG pipelines by scoring retrieval and generation against gold references.

Community
Advanced
VAMFIVAMFI
206

quality-validation

Validate ResearchPacks and Implementation Plans with predefined rubrics and scoring.

Official
Intermediate
mikeparcewskimikeparcewski
8

smaht

Gathers requested context from wicked-brain, wicked-garden, and domain state on demand.

Community
Intermediate