Castorini Pipeline

Castorini Pipeline

End-to-end pipeline orchestration across rank_llm, ragnarok, nuggetizer, and umbrela.

Use this skill to reason about handoffs between repositories, not as a substitute for repo-local verification. After each stage, inspect the actual artifacts before advancing.

rank_llm usually belongs before ragnarok: use it to retrieve and rerank candidate passages, then feed those contexts into ragnarok for answer generation and downstream nuggetizer evaluation.

Pipeline Stages

[0. Retrieve + Rerank]     rank-llm rerank --dataset ...
         │ JSONL / TREC rerank artifacts
         ▼
[1. Generate Answers]      ragnarok generate --dataset ...
         │ JSONL (cited answers)
         ▼
[2. Create Nuggets]        nuggetizer create --input-file ...
         │ JSONL (scored nuggets)
         ▼
[3. Assign Nuggets]        nuggetizer assign --contexts ... --nuggets ...
         │ JSONL (assigned nuggets)
         ▼
[4. Calculate Metrics]     nuggetizer metrics --input-file ...
         │ JSONL (per-query scores)
         ▼
[5. Judge Relevance]       umbrela evaluate --qrel ... --result-file ...
         │ Modified qrels + nDCG@10
         ▼
[Results]

Reference Files

• references/pipeline-walkthrough.md — Complete end-to-end example with commands
• references/stage-handoffs.md — JSONL format compatibility between stages

Stage Dependencies

Note: Stages 2-4 (nuggetizer) and Stage 5 (umbrela) are independent evaluation paths — they measure different things:

• Nuggetizer path (stages 2-4): Measures answer completeness against extracted nuggets
• Umbrela path (stage 5): Measures retrieval quality against human relevance judgments

Gotchas

• Format alignment: ragnarok output uses topic_id/topic; nuggetizer expects qid/query. The fields are compatible — nuggetizer normalizes both.
• rank_llm comes first: use it when the workflow starts from retrieval and reranking. It usually feeds ragnarok or umbrela, not the other way around.
• Nugget pool: For nuggetizer create, the "pool" is the candidate passages, not the generated answers. Use the original retrieval input, not ragnarok's answer output.
• Assign contexts: For answer evaluation, use ragnarok's answer output as the contexts file (--input-kind answers). For retrieval evaluation, use the retrieval result as contexts (--input-kind retrieval).
• Write policies: Use --resume on long-running stages to allow restart without reprocessing.
• Model consistency: Document which model was used at each stage for reproducibility.
• pyserini dependency: rank_llm retrieval workflows, ragnarok dataset mode, and umbrela evaluate all rely on pyserini-compatible setups.
• Evaluation paths diverge: Nuggetizer stages evaluate answer completeness, while umbrela evaluates retrieval relevance. Do not merge those metrics into one score without making the distinction explicit.