# Corpus Retrieval Lab

> Benchmark current research-query and direct lexical baselines against experimental vector, BM25, typed concept-graph PPR, and RRF source selection without replacing production retrieval.

- Skill: `jmagly-ai-writing-guide/corpus-retrieval-lab` (Agent Skill)
- Install (CLI): `npx skillmds@latest add jmagly-ai-writing-guide/corpus-retrieval-lab`
- Raw SKILL.md: https://api.skillmd.com/api/skills/jmagly-ai-writing-guide/corpus-retrieval-lab/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: jmagly (https://skillmd.com/u/jmagly-ai-writing-guide)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/jmagly-ai-writing-guide/corpus-retrieval-lab

---


# Corpus Retrieval Lab

Run the local experimental benchmark:

```bash
aiwg corpus retrieval-lab --queries <jsonl> --concepts <json> --json
```

Use reviewed narrow implementation questions with expected REF evidence. Preserve the current `research-query` behavior regardless of the result; the report may only recommend a later explicit adoption decision.

The lab reports baseline and hybrid Hit@k, MRR, latency, failures, confidence/dispersion, faithfulness, and concept-scheme drift. Use `--expected-scheme-hash` when comparing runs across time.

See `docs/guides/corpus-retrieval-lab.md` for fixture schemas and interpretation.