Rag

Skill Neznakometz/StackForge/storage/skills/rag

Thin core + domain sets for a fast, stack-aware project start. Cross-agent (Claude Code, Cursor, Codex, Gemini) via the SKILL.md standard. MIT.

Install

npx -y skills add Neznakometz/StackForge --skill rag

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Retrieval-Augmented Generation — chunking, embeddings, vector storage, search and reranking. Apply when building search over knowledge/documents, Q&A over a corpus, semantic search.

SKILL.md

1.9 KB, as published. Nobody here has run it

RAG (retrieval-augmented generation)

When RAG is needed / not needed

Needed: a large knowledge/docs corpus that doesn't fit in context and changes; you need fresh/private data with citation.
NOT needed: the data fits in context; the task is about code (a code index is better, see the long-context skill); a one-off question (direct search is simpler).

Pipeline

Chunking. Cut by meaning (sections/paragraphs), not by fixed characters blindly; a reasonable chunk size + overlap so you don't tear a thought apart. Keep metadata (source, section) for citation and filters.
Embeddings. One embedding model per corpus (query and documents — by the same model). Version it: changing the model → reindexing.
Storage. pgvector (if Postgres already exists) or Qdrant. Index — HNSW by default (good recall without tuning lists, parameters m=16, ef_construction=200); IVFFlat — only for very large, rarely changing data for the sake of build speed/memory. Don't create an index on an empty table (k-means centroids on empty data → a broken index) — load the data first.
Search. Query understanding (intent/filters/tenant) → optional lexical prefilter → ANN top 50–200 → reranking by exact distance + metadata → top-k into context.
Hybrid. Semantic + full-text (BM25/FTS) via Reciprocal Rank Fusion (RRF) — noticeably better on terms/exact matches.
Generation. Into the answer — only the retrieved chunks, each statement with a link to the source; don't mix the model's memory in as fact.

Multitenancy/security

Scope the search by tenant (a filter in the query/metadata) — don't hand back someone else's chunks.
Don't index secrets/PII.