AgenticWire MCP vs RAG benchmark

Purpose
-------
This is a small, deterministic evidence-access comparison for the AgenticWire article:
https://www.agenticwire.news/article/mcp-vs-rag

It compares three local conditions:

1. Retrieval-only: lexical token overlap over a small document corpus.
2. MCP read-tool: a JSON-RPC-shaped tools/call adapter over structured records.
3. Hybrid: a hand-authored router that sends tool-shaped queries to the read tool and other queries
   to retrieval.

Dataset and protocol
--------------------
- 12 labeled queries
- 8 local records
- 1,000 repetitions per query and method for local latency
- Queries cover semantic retrieval, structured fields, cross-record reads, and negative/no-match cases.
- The test does not call a language model, embedding service, vector database, remote MCP server, or
  network transport.

Metrics
-------
- Evidence recall: all gold record IDs were returned.
- Structured exactness: structured and cross-record queries returned the required records through the
  read-tool path.
- Negative safety: no-match queries returned an empty result.
- Failure count and protocol-error count.
- Median local execution time across the repeated runs.

The machine-generated values are in results.json. They are a bounded local routing result, not a
universal ranking of MCP, RAG, or language-model answer quality.

Source date: 2026-08-16