Running a literature sweep

Everything except step 4 is direct and free.

Fan out the topic

nisa search \
  --query "retrieval reranking LLM" \
  --query "listwise reranking" \
  --query "cross-encoder reranking efficiency" \
  --max-results 40 --date-from 2023 > sweep.json
nisa semantic-search "how to rerank retrieved passages cheaply" --max-results 20 > sweep-sem.json

Keep the raw JSON — its index_latest_pub_month is your record of what the index held.

Expand from anchors

nisa related 2407.08608 --n 15
nisa related 2312.02724 --n 15

Ids that appear in more than one neighbourhood are usually worth reading.

Verify claims in the text

nisa excerpts 2407.08608 --query "latency" --max 5
nisa excerpts 2407.08608 --query "\"exact match\"" --max 3

If doc_found or full_text_indexed is false, read the PDF: nisa pdf 2407.08608 --out paper.pdf.

Batch the judgment calls

usage-tracked One question per line in qs.txt:

nisa ask-batch --timeout 90 --schema '{"verdict":"string","arxiv_ids":["string"]}' < qs.txt > out.jsonl
jq -c 'select(.ok == false)' out.jsonl        # failures to retry

Sequential by design; one invocation, one rate-limit slot.

Save and export

LIST=$(nisa create-list "reranking-sweep" --emoji "🔎" | jq -r .id)
nisa add-to-list "$LIST" 2407.08608 --note "strong baseline"
nisa list-papers "$LIST" --format bibtex > sweep.bib

The list is in the web app too.

⌘I

Start

Guides

Reference

Running a literature sweep