Searching the corpus

Direct commands answer from the indexes in about a second and are free. Chat-backed commands create a model turn — slower, usage-tracked. Compose direct commands first; reach for the agent when you need judgment.

Which command

You want	Run
Papers about a topic	`nisa search` with 3–5 phrasings (below)
Papers like a description	`nisa semantic-search "…"`
Everything about one paper	`nisa paper <id>` → `nisa excerpts <id> --query "…"` → `nisa related <id>` → `nisa cite <id>`
”From paper X, what else should I read?”	`nisa related <id> --n 15`
Who builds on X	`nisa search "<X's coined term>"`, then `nisa excerpts <hit> --query "<term>"` to verify
Compare A and B at claim level	`nisa paper A`, `nisa paper B`, `nisa excerpts` on both
Rank, pick, decide among N	`nisa ask --schema '…'` — chat-backed
A long synthesis	`nisa exec` — chat-backed, multi-turn

Several phrasings, one call

nisa search \
  --query "mechanistic interpretability diffusion" \
  --query "concept unlearning diffusion" \
  --query "SAE diffusion" \
  --max-results 25 \
  | jq '.papers | map(select(.year >= 2024))'

Parallel fan-out, deduplicated by arxiv_id, re-sorted. Caps: 8 queries for search, 4 for semantic-search; the server returns at most 50 results per call — paginate with --offset (clamped at 500; 200 for semantic-search). In list mode you lose which query found which paper.

Read the ranking honestly

search — recency is under-ranked

BM25 blended with citation counts: a recent, low-citation paper ranks far below an older, well-cited one. An empty recency window is a ranking artifact, not proof of absence. Every response carries index_latest_pub_month — check it before concluding a paper doesn’t exist.

excerpts — zero hits ≠ absence

Tokenized OR matching (quoted phrases exact). Check matched_terms, missing_terms, doc_found, full_text_indexed before treating zero excerpts as absence.

Two indexes, different freshness

Full-text (search, excerpts) and embeddings (semantic-search) are refreshed on different schedules and can differ. There’s no published cadence — use index_latest_pub_month.

Filters and formats

search: --date-from 2024-06 --date-to 2025 (month precision; a bare year in --date-to means December), --author "Vaswani". semantic-search: --year-min, --year-max. --format json|bibtex|csv on search, semantic-search, related, list-papers. --envelope wraps direct-command JSON as {meta, data} for scripts. CLI reference

Start

Guides

Reference

Which command

Several phrasings, one call

Read the ranking honestly

Filters and formats

​Which command

​Several phrasings, one call

​Read the ranking honestly

​Filters and formats

Which command

Several phrasings, one call

Read the ranking honestly

Filters and formats