Evaluate RAG retrieval quality with Recall@k, Hit@k, MRR, NDCG@k via @mukundakatta/ragmetric-mcp

@mukundakatta/ragmetric-mcp v0.1.1 (serverInfo: ragmetric-mcp/0.1.0) — 5 tools, 1 capability (tools), protocol 2024-11-05 conformant.

Install & run: npx @mukundakatta/ragmetric-mcp or node src/index.js

Tools:

1. recall_at_k — Fraction of relevant docs in top k retrieved. Args: retrieved (string[]), relevant (string[]), k (number).
2. hit_at_k — 1.0 if any relevant doc in top k, else 0.0. Same args.
3. mrr — Mean Reciprocal Rank (1/rank of first relevant). Args: retrieved, relevant.
4. ndcg_at_k — NDCG with binary relevance, log2 discount. Args: retrieved, relevant, k.
5. evaluate_batch — All 4 metrics over a batch of queries. Args: queries (array of {retrieved, relevant, k} objects).

Verified trace (3 runs, 6/6 calls successful):

recall_at_k({retrieved:["doc1","doc2","doc3"], relevant:["doc1","doc3","doc5"], k:3})
→ {recall_at_k: 0.6667}  // 2 of 3 relevant found in top 3

hit_at_k({retrieved:["doc1","doc2","doc3"], relevant:["doc1","doc3"], k:3})
→ {hit_at_k: 1}  // at least one relevant doc found

Performance: p50 init ~1959ms, p50 call ~17ms. All metric computations are pure in-memory, sub-25ms.

Gotchas:

• retrieved and relevant are arrays of string IDs, not objects.
• k is required for recallatk, hitatk, and ndcgatk — no default.
• mrr does NOT take a k argument — it uses the full retrieved list.
• evaluate_batch expects queries (array), not a flat list of args.
• Entry point is src/index.js (not dist/), differs from other @mukundakatta servers.