AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

paperclip

Searches and reads biomedical papers, FDA/PMDA/EMA documents, clinical trials, and protein records with the GXL Paperclip CLI and Python SDK. Supports source-scoped search, full-text grep, metadata SQL, map/reduce extraction, figure analysis, optional repositories and claim verification, and line-pinned citations. Use when a task names GXL paperclip, asks to install or authenticate it, or requests literature retrieval and evidence extraction through Paperclip.

Original source
K-Dense-AI
License
MIT
Pinned revision
5f2bc0341354fd5590aef76a533ada1e1e5b4e62

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

Ask your assistant

Use Agentic Memory skills to read k-dense-ai--paperclip at revision 5f2bc0341354fd5590aef76a533ada1e1e5b4e62, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 7

references/search-and-retrieval.md

Search, retrieval, and query craft

Reviewed on 2026-09-30 against the 0.7.92 CLI/SDK, the core reference, protein reference, and web docs. Retrieval examples are illustrative: no paid or authenticated searches were executed for this review. Load credentials as described in installation.md.

Choose a source explicitly

| Source | Scope | |---|---| | pmc | PMC full-text papers | | arxiv, biorxiv, medrxiv | Corresponding preprint corpora | | papers | The four paper corpora combined | | abstracts | Title/abstract-grain retrieval across scholarly corpora; not a full-text promise | | fda | Regulatory search; use a regional source/path to disambiguate | | fda/jp, fda/eu, /fda/us | PMDA, EMA/EPAR, and US FDA | | trials | Trial registries; trials/us, trials/eu, trials/jp, trials/cn narrow geography | | proteins | Unified UniProt/PDB/ChEMBL records; uniprot, pdb, chembl are aliases in the client | | clipboard | The user's documents; clipboard/FOLDER scopes to a folder | | geo | GEO Series (documented CLI source; not listed in the current v1 JSON search enum) | | patents | Patent records; load paperclip skill patents first |

The abstracts CLI source and the public /api/v1/search default have different published coverage wording. Choose a source explicitly and report the selected interface; do not label all abstract results as OpenAlex or assume coverage equals PMC. Counts in upstream documents disagree and change, so use a dated catalogue or query result when a corpus count matters.

paperclip search -s pmc,biorxiv "protein design" -n 5
paperclip search "pembrolizumab" /fda/us -n 5
paperclip search "breast cancer" /trials/us -n 5

Match the task to paper evidence, trial registration, regulatory material, or structured protein records. Infer that choice from a clear user request; clarify only when the intended output remains ambiguous. Regional paths /trials/... and /clinicaltrials/... are aliases in current docs.

Search options

| Option | Contract | |---|---| | -n/--limit N | Requested hit count, not a corpus census or a pagination cursor | | -e/--exact | Exact-phrase search | | --year YYYY | Publication-year filter; validate returned metadata | | --since WINDOW | Recency strings such as 30d, 6m, 1y; support depends on source | | --sort relevance\|date | Ordering | | --author, --journal | Metadata narrowing; source coverage varies | | --ranking hybrid\|bm25\|vector\|analogical | Retrieval mode | | --min-embedding-similarity FLOAT | Floor on the vector leg | | --min-bm25-score FLOAT | Floor on the lexical leg | | --corpus | Full-corpus discovery even when scoped to a repo |

The SDK and vendor core reference disagree on arXiv --since support; use --year for arXiv, medRxiv, and abstracts unless current server help confirms the needed recency behavior. Do not assume an arbitrary date string or a filter unsupported by a source is applied.

The relevance floors do have CLI equivalents. For hybrid mode, each applies only to its own retrieval leg: use both to bound both legs, or pair one with its matching ranking. Floors do not expand the requested candidate set. They are unsupported with regex/analogical modes; a vector floor is incompatible with date sorting. These restrictions are documented in the 0.7.92 SDK.

paperclip search -s pmc --ranking hybrid \
  --min-embedding-similarity 0.7 --min-bm25-score 5 \
  "protein design" -n 5

hybrid is a good discovery default; bm25 fits exact terminology and gene symbols, vector fits varying vocabulary, and analogical fits cross-domain method transfer. Write one or two sentences about the method/problem for analogical search, or use an identified reference paper's abstract:

paperclip search -s arxiv --ranking analogical \
  "I need to approximate an expensive leave-one-out computation cheaply by exploiting low-rank structure in my parameter space" -n 10

Upstream describes stable search prefixes through -n 100 for unchanged query, filters, and index; requests above 100 can expand the candidate pool and reorder earlier results. This is not pagination or a cross-date reproducibility guarantee. Save exact queries, source filters, returned IDs, and date.

Structured results and cohort preservation

Search previews and generated summaries are triage, not paper quotations. In 0.7.92 SDK, result.papers, result.count, and result.result_id expose structured results and hydrate missing hit data from saved results when possible. A failed hydration can leave incomplete data: check requested versus returned count and reported truncation. See python-sdk.md.

paperclip results --list
paperclip results s_ID --save search.csv
paperclip results s_ID --export-bundle cohort/
paperclip results s_ID --sample 20 --seed 42

CSV exports and bundle formats depend on result kind; do not assume every map export has the search metadata header. Cohort paging retrieves the saved cohort, not additional corpus search hits. The CLI supports --save-as NAME aliases on result-producing commands; their scope is session-bound, so record the durable result ID for another session.

paperclip filter --from s_ID --require 3 "in vivo delivery with quantitative outcomes"

filter is an LLM pass that overwrites the cohort in place. If too few survive, rerun a broader search for a fresh ID; a second filter cannot recover discarded rows. For deterministic quality filters, consult current refine --help and preserve both input and output cohort IDs.

Exact metadata lookup

paperclip lookup doi 10.1073/pnas.2307796121
paperclip lookup pmc PMC7194329
paperclip lookup pmid 32943797
paperclip lookup arxiv 2403.03507
paperclip lookup author "James Zou" -n 10

Current docs list doi, title, author, abstract, source, date/month_year, pmc, pmid, journal, publisher, type, keywords, category, license, year, volume, issue, issn; the core reference also lists arxiv. Many metadata fields are PMC-specific. Do not turn an empty field match into a claim that the paper does not exist. The public HTTP lookup's narrower typed contract is documented separately in python-sdk.md.

Full-text grep and scan

paperclip grep -l "SLC30A8" /papers/
paperclip grep -n -C 3 "lipid nanoparticle" /papers/PMC10945750/content.lines
paperclip grep -i -e "off-target" -e "offtarget" /papers/PMC10945750/content.lines
paperclip grep --from s_ID "IC50"
paperclip scan -i -C 3 /papers/PMC10945750/content.lines "IC50" "EC50" "dose"

-n means line numbers; -m N bounds matches. Corpus scans are parallel and time-bounded, so a limit or timeout can change returned membership/order. --exhaustive allows more scan time, not guaranteed completeness. The current guide says corpus grep -c counts documents exactly only through 500 bitmap candidates, then approximately; file grep -c counts matching lines. Do not report a corpus estimate as an exact prevalence count.

Boolean search and grep use different semantics: search --bool --ranking bm25 uses analyzed phrase operands; grep --bool uses regex predicates over whole documents. Do not interchange them in a review protocol without validating the intended matching behavior.

Metadata SQL

paperclip sql "SELECT source, COUNT(*) AS n FROM documents GROUP BY source"
paperclip sql "SELECT pub_year AS year, COUNT(*) AS n FROM documents WHERE title ILIKE '%CRISPR%' GROUP BY pub_year ORDER BY year DESC LIMIT 10"

The current core reference documents SELECT on documents, 15-second server timeout and a 200-row cap. Its columns are id, title, doi, authors, source, abstract_text, pub_date, pub_year, journal_title, article_type, pmid, keywords, and categories. Source-specific missing fields may be NULL. Use explicit aliases and inspect actual returned columns.

This metadata SQL route does not search paper bodies. abstract_text ILIKE '%X%' misses a term found only in Methods or supplementary material. The web docs expose lower-level document_id, content_blocks, and figures in an export context; do not assume those names or joins are interchangeable with ordinary sql. Current docs differ on some column naming, so check server help before introducing a new query rather than claiming automatic normalization.

Protein SQL

paperclip skill proteins
paperclip sql -s proteins "SELECT COUNT(*) FROM uniprot_v.proteins"
paperclip sql -s proteins "SELECT * FROM pdb_v.structures_by_accession WHERE accession='P00533' LIMIT 10"
paperclip cat /proteins/P04637/meta.json

The current domain reference exposes uniprot_v.proteins, uniprot_v.features, pdb_v.structures_by_accession, chembl_v.bioactivities_by_accession, and chembl_v.drugs_by_accession. Check that reference for exact column names, enum values, join cardinality, and accession mapping before writing a biological query. An accession-based join is not a guarantee of one row per protein or of directly comparable bioactivity assays.