Skip to content

Evaluation Overview

Evaluation is a core input to RAGhelm’s release-control layer. Evals measure retrieval, generation, regression, latency, and production-readiness evidence, but eval summaries alone are not the product object. Per ADR-002, release decisions should be expressed through ReadinessScorecards linked to RAGRunManifests.

What the eval suite provides

  • Golden dataset loading and validation
  • Retrieval metrics: Recall@k, MRR, NDCG@k
  • Generation scoring: faithfulness, relevance, completeness, overall
  • Regression checks against baselines
  • Production-mode hooks for Pinecone retrieval and LLM judging
  • Machine-readable JSON artifacts under data/eval_results/

Local vs production evals

Mode Purpose Evidence status
local Fast deterministic development feedback Useful for development; not production proof
production Hosted Pinecone retrieval plus production judge behavior Eligible as input to scorecards/manifests when config and artifacts are valid

Public claims and badges must fail closed unless backed by production scorecard and manifest evidence per ADR-010.

Key components

Component Purpose
golden_dataset.py Load and validate evaluation datasets
runner.py Orchestrate eval runs
metrics.py Compute Recall@k, MRR, NDCG@k
scorer.py Score generation quality
regression.py Detect regression against baseline
benchmark.py Latency benchmarking helpers
production_adapter.py Production-mode retrieval/generation/judge integration

Golden dataset sources

The default 100-example dataset uses open-license RPG/system-reference content. It is a development and demonstration dataset, not customer production data.

Source License Examples
Cairn RPG CC-BY-SA 4.0 20
SCP Foundation CC-BY-SA 3.0 20
Fate Core / related SRDs CC-BY 3.0 21
Dungeon World CC-BY 3.0 20
D&D 5.1 SRD CC-BY 4.0 20