Stuff (extra citations from the library)

paperful stuff suggests extra citations you already hold for each claim in a draft. It splits the extract into sentences (or outline sections), runs the same retriever as paperful rag search, and writes a pack of hits. There is no Ask narrative, no invented BibTeX, and no library write.

Needs [rag].enabled and a built index. Does not need [llm]. Fail closed if the index is missing or empty.

This is not a rewrite and not claim verification. Similarity is not proof the source supports the sentence. Extra cites belong in Zotero before \cite.

CLI

--item KEY and --file PATH are mutually exclusive. --file accepts the same text formats as critique (pdf, txt, md, html, rtf, docx, odt, pptx; legacy .doc if a converter is on PATH).

uv run paperful stuff --file ./draft.docx -C "My Collection" --k 3
uv run paperful stuff --file ./outline.docx --lane outline -C "My Collection"
uv run paperful stuff --item KEY -C "My Collection"
uv run paperful stuff --file ./draft.md --dry-run

--k is extra sources per claim (default [stuff].k = 3, capped by [rag].top_k). -C / --library / --year-from / --year-to / --type scope the index the same way as rag search. When --item is set, that item is never listed as a hit for itself.

--lane selects the retrieval unit (not an LLM extra). One lane per run.

Lane	Default	Unit	Skips
manuscript	yes	sentences after dropping the bibliography window	fragments under min_sentence_words (default 8)
outline	--lane outline or [stuff].lane	one row per section (outline_slices)	canonical References / Acknowledgements; empty queries

Outline query is {heading} {body}, truncated. Thin named sections still search (the heading is a useful query). Same heading ledger shape as critique --lane outline.

--dry-run prints counts and does not write state/stuff/. --format json prints paperful.agent.json.v1. Not in paperful all. paperful jobs → completeness.

Pack

Packs land in state/stuff/<id>/<stamp>/ as pack.json (paperful.stuff.pack.v1) and pack.md. <id> is the library item key or file-<filename-stem>.

Each row is a quote plus locator plus ≤ k hits (item key, title, pages, score, passage snippet). Skips record short, biblio, over_cap, or empty. Outline runs also write an outline ledger. [stuff].max_units (default 200) caps how many claims are searched.

pack.md leads with an honesty banner: hits are in-library similarity, not proof of support; this pack does not invent citation keys.

Config

[stuff]
k = 3
min_sentence_words = 8
min_query_words = 4
max_units = 200
lane = "manuscript"   # manuscript | outline

See config.md.

What this is not

• Not ask (no grounded paragraph).

• Not refs gap (cited-in-PDF works you do not own).

• Not precedent (no OpenAlex / public-metadata-only rows).

• Not auto-insert into the manuscript. --apply and stance labels stay later.

Workbench

Advanced Stuff (next to Suggest edits; CLI is still paperful stuff): pick a library item or upload a draft, choose lane (manuscript / outline) and k, download pack.md / pack.json. Needs [rag] and a built index. Does not need [llm]. Never writes Zotero. Uploads stay under state/critiques/uploads/ like Critique.

Read-only MCP tool stuff (item or file, lane, k, collection) returns the same envelope as --format json.