On-disk RAG evaluation (corpus/ + train.json)

Purpose

Guide agents through NVIDIA RAG Blueprint filesystem benchmarks: preparing corpus/ and train.json, running scripts/eval/evaluate_rag.py, tuning retrieval and generation flags for quality comparisons, interpreting RAGAS JSON outputs, and triaging failures (HTTP/stream errors, empty contexts, collection mismatch, judge API).

For latency, throughput, and load testing, use the rag-perf skill (scripts/rag-perf, docs/performance-benchmarking.md) — not this skill.

When not to use

Do not use this skill for: deploying or repairing services (use rag-blueprint); evaluating APIs without the corpus/ + train.json layout; general ML experimentation unrelated to this evaluator; production monitoring/alerting; or latency/throughput benchmarking (use rag-perf).

Prerequisites

• Repo cloned; run commands from repo root (imports and paths assume this).
• Python 3.11+ and uv; eval deps: uv sync --project scripts/eval.
• Reachable RAG server and ingestor (defaults often localhost:8081 / 8082).
• NVIDIA_API_KEY for RAGAS (see credential hygiene); optional RAG_EVAL_JUDGE_MODEL.
• Dataset roots passed to --dataset-paths each contain corpus/ and train.json.

Instructions

1. Prepare data — Ensure each dataset directory matches the layout and train.json rules in references/dataset-and-conversion.md. When sources arrive as public links (sites or dataset pages), materialize documents under corpus/—prefer PDF for multimodal content so images stay embedded; convert CSV/JSONL/etc. using the patterns there.
2. Run eval — uv run --project scripts/eval python scripts/eval/evaluate_rag.py with --dataset-paths, --host, and --port. See references/benchmark-execution.md for command examples, outputs, and errors. Use references/evaluate-rag-cli.md for flag-level detail.
3. Tune quality — Adjust --top_k / --vdb_top_k, reranker and query-rewriting toggles, and generation overrides (--temperature, --top-p, --max-tokens) as documented in references/benchmark-execution.md when comparing retrieval/generation configs for RAGAS scores.
4. Analyze results — Use references/result-analysis.md for scripts; scan rag_*_evaluation_summary.json for headline RAGAS metrics.
5. Triage errors — Use the error signal table and the Troubleshooting section below.

Examples

Set API key without putting secrets in shell history (preferred patterns): load from a gitignored env file or secrets manager; avoid committing .env; rotate keys if exposed. Details: references/benchmark-execution.md#credential-hygiene-nvidia_api_key.

Minimal eval (key already in environment):

uv sync --project scripts/eval
uv run --project scripts/eval python scripts/eval/evaluate_rag.py \
  --dataset-paths /path/to/my_dataset \
  --host localhost \
  --port 8081

Pretty-print summary JSON:

python3 -m json.tool results/my_dataset/rag_my_dataset_evaluation_summary.json

More examples (skip ingestion, quality sweeps): references/benchmark-execution.md.

Limitations

• Evaluator behavior is fixed to the filesystem contract and evaluate_rag.py; it does not substitute for custom offline judges or non-RAG benchmarks.
• Vector DB / embedding choices follow deployed ingestor and RAG env — not overridden by this CLI alone.
• Scores depend on retrieval quality, judge model availability, and NVIDIA_API_KEY; empty contexts yield partial RAGAS metrics (see references).
• Large procedural detail lives under references/ to keep routing concise; read those files when the user needs step-by-step conversion, full flags, or error tables.

Troubleshooting

Full signal table: references/benchmark-execution.md#common-error-cases-and-signals.

Gotchas

• Run from repo root: paths and imports in scripts/eval/evaluate_rag.py assume this; a wrong directory silently breaks imports.
• --ingestor_server_url: pass http://host:port without /v1—the code appends /v1/ automatically. Including /v1 causes 404s on ingestor calls.
• Vector DB / embedding settings: not set by this CLI; configure via the deployed ingestor and RAG server env vars (e.g. APP_VECTORSTORE_URL, embedding model).
• --model / --llm_endpoint: forwarded verbatim only when explicitly set; omit to keep the server's configured LLM.
• Stale collections: a previous run's ingested data persists unless you use --force_ingestion. Use --collection with a unique name when comparing quality across isolated runs.
• Empty context metrics: if all generated_contexts are empty, RAGAS scores only nv_accuracy and leaves the other two metrics blank—this is not a silent success.

Source of truth

Agent playbook

1. Run eval — uv sync --project scripts/eval then uv run --project scripts/eval python scripts/eval/evaluate_rag.py with required --dataset-paths, --host, and --port (and env NVIDIA_API_KEY). Argument --ingestor_server_url is optional (defaults to http://localhost:8082); pass it only when overriding the ingestor endpoint.
2. Quality tuning — See references/benchmark-execution.md: --top_k/--vdb_top_k, reranker and query-rewriting toggles, --temperature, --top-p, --max-tokens.
3. Data conversion — Follow references/dataset-and-conversion.md.
4. Analyze results — references/result-analysis.md; quick scan: python3 -m json.tool results/<dataset>/rag_<dataset>_evaluation_summary.json.
5. Error triage — references/benchmark-execution.md#common-error-cases-and-signals.

Cross-Client Portability

This skill is written to stay usable across GitHub Copilot, Claude Code, and Codex.

• GitHub Copilot: keep the folder in a Copilot-visible skill path or wrap the workflow in project instructions when folder discovery is unavailable.
• Claude Code: keep the folder in a local skills directory or a compatible plugin source.
• Codex: install or sync the folder into $CODEX_HOME/skills/rag-eval and restart Codex after major changes.

MCP Availability And Fallback

Preferred MCP Server: None required

• Fallback prompt: "Use the On-disk RAG evaluation (corpus/ + train.json) skill without MCP. Rely on its local instructions, bundled resources, standard shell or editor tools, and direct verification. Show the evidence used before concluding."
• Do not claim an MCP operation was used when the active host does not expose it.
• Treat local files, tests, rendered outputs, logs, or screenshots as the fallback evidence path.

Anti-Patterns

• Changing the eval dataset while comparing runs: It destroys the baseline and makes improvements meaningless.
• Confusing latency smoke tests with answer-quality evaluation: Fast responses can still be wrong or ungrounded.
• Claiming gains without showing the baseline, scorer, and prompt or config deltas that changed the outcome.

Verification Protocol

Before claiming "skill applied successfully":

1. Pass/fail: The evaluation plan names the dataset, scorer, and baseline run before comparing variants.
2. Pass/fail: Retrieval and generation quality are separated so failures are attributed to the correct stage.
3. Pass/fail: Reported improvements include reproducible commands, configs, or artifacts that another maintainer can rerun.
4. Pressure-test scenario: Re-evaluate a RAG change where latency improves but groundedness falls on the held-out set.
5. Success metric: Quality claims survive a rerun on the same eval slice with no hidden configuration drift.

Related Skills

• development-workflow: Use it when the eval work needs a scoped implementation plan with explicit quality gates.
• documentation-verification: Use it when the output is an evaluation report or benchmark note that must stay source-backed.
• cloud-design-patterns: Use it when evaluation results drive bigger architecture changes in the RAG stack.