RAG That Actually Works 5 Claude Skills for Shipping Retrieval Systems

Every RAG tutorial ends where the real work startsYou chunked the PDFs, embedded them, got a working demo. Then you pointed it at the real corpus.Now it says "I don't have that information" about a document sitting right in front of you. It states numbers that appear in no source. It can't find your own error codes. And it got worse after you added more documents.Every guide teaches you to build RAG. Almost none tell you what to do when it doesn't work.This pack is that missing half.What's differentMost skill packs are markdown files with opinions in them. This one ships 10 Python scripts that actually run — against your real PDFs, your real index, your real EXPLAIN plan.All of them run offline. No API key needed to test. You can prove this pack is worth it in ten minutes, before spending a cent on tokens.The five skillsrag-ingestion — Audits your document folder for what nobody checks: scanned PDFs that extract to whitespace and get indexed as empty vectors (silent, no error, fails weeks later), headers and footers polluting every chunk, and near-duplicate documents that are really the same policy in three versions. Plus a chunking previewer that flags splits which break tables and truncate silently.rag-retrieval — Runs a BM25 baseline over your chunks and reports recall@k and MRR. If BM25 finds the answer and your system doesn't, your embedding model isn't the problem — and you just saved a week. Plus an index auditor for defects that cap quality no matter what you tune.rag-eval — Scores runs and gates CI on regressions, with hallucination detection that needs no model call. Every number and date in the answer is checked against the retrieved context. Invented figures get flagged instantly, for free. Handles "18 weeks" vs "eighteen weeks" correctly, so you don't drown in false alarms.rag-debugger — Takes one failing query and walks six checkpoints to name the exact stage that lost the answer:✓ 0. in the corpus found as policy.pdf#00000 ✗ 1. survived the filter excluded on `effective_from`Plus a matcher for 30 catalogued failures — from pasted stack traces to plain-English complaints.supabase-pgvector-rag — Complete migration, hybrid search RPCs, RLS policies, typed TypeScript client. Plus a preflight that catches the pgvector fault producing no error at all: an index built vector_cosine_ops isn't used by a query using <->. Results stay correct, the scan goes sequential. Fine on 500 dev rows, an outage at 500,000.They work as a systemThe debugger doesn't fix anything — it tells you which of the other four to open. The eval scripts read the format the ingestion pipeline writes. The Supabase schema passes its own preflight check.What you get 5 installable Claude Skills (.skill files) 10 Python scripts, all offline, no API key Complete Supabase schema + typed TS client 17 in-depth reference docs, CI workflow, judge prompts Works in Claude Code, Claude Desktop, Cursor, Codex CLI — anything supporting the open Agent Skills standard.

Found on Devadex — the discovery index for independent software the big search engines bury. More from gumroad.

Report this listing