bio-search

paper-wiki bio-search

Overview

bio-search is the biomedical-literature surface for paper-wiki. It delegates the actual search to paperclip via the paperclip MCP server (mcp__paperclip__* tools when registered) and optionally hands hits to /paperwiki:wiki-ingest so they land in your wiki the same way arXiv papers do.

This SKILL is opt-in: paperclip is a third-party service with its own auth and tier model. paper-wiki's other SKILLs (digest, analyze, wiki-*) work without it. Only invoke bio-search when the user is specifically asking for biomedical-domain literature.

When to Use

• The user types /paperwiki:bio-search <query>.
• The user says "search bioRxiv", "search medRxiv", "search PubMed", "find biomedical papers on X".
• The user asks for clinical / translational research that arXiv generally does not cover (pathology, oncology, cardiology, genomics, drug discovery, etc.).
• The user follows up an analyze/digest thread with "what do bioRxiv preprints say about this?" or similar.

Do not use for:

• arXiv / general CS work — that routes to /paperwiki:digest with the daily-arxiv recipe.
• Single-paper deep dives — that routes to /paperwiki:analyze.
• Searches inside the user's own wiki — that routes to /paperwiki:wiki-query.

Process

1. Check that paperclip MCP is available. Run ${CLAUDE_PLUGIN_ROOT}/.venv/bin/python -m paperwiki.runners.diagnostics and inspect mcp_servers. If paperclip is missing, stop and show the user the registration steps from docs/paperclip-setup.md. Do not run claude mcp add yourself; the user opts in.

   As an offline fallback, mention the bundled recipes/biomedical-weekly.yaml recipe — it shells out to the paperclip CLI and works without the MCP layer if the user prefers.

2. Resolve the query. Accept free-text. If the user gives a vague topic ("pathology"), ask one clarifying follow-up before burning a paperclip call: "do you want preprints, peer-reviewed PMC, or both?", "any time window?", etc. Default to bioRxiv + medRxiv + PMC over the last 14 days unless told otherwise.

3. Call paperclip via MCP. Use the mcp__paperclip__search tool (or whatever name the registered MCP server exposes). Surface the raw paperclip results — title, authors, source, abstract, link — in a numbered list. Do not paraphrase; users want to scan the actual abstracts.

4. Offer follow-ups.

   • "Analyze paper #N in depth" → invoke /paperwiki:analyze paperclip:bio_<id> (or arxiv:<id> if the hit has an arXiv id under external_ids).
   • "Save these as wiki sources" → for each chosen hit, invoke paperwiki.runners.fetch_pdf <vault> <canonical-id> (when Phase 8 ships) or MarkdownWikiBackend.upsert_paper directly so the source lands under Wiki/sources/. Then chain to /paperwiki:wiki-ingest.
   • "Refine the search" → loop back to step 2 with the new query.

5. Summarize at the end. Tell the user what was added to the wiki (canonical ids + concept names if any), and suggest the next step: /paperwiki:wiki-lint to spot dangling sources, or /paperwiki:wiki-query to ask follow-up questions.

Common Rationalizations

Excuse	Why it's wrong
"I'll auto-register paperclip MCP for the user."	Auth is sensitive and the user may be on a metered plan. Hand over the registration command, never run it.
"I'll paraphrase the abstracts to save space."	Users want to scan the actual paperclip text — paraphrasing introduces drift and hides what the search engine actually returned.
"If paperclip times out, I'll fall back to my training data."	Out-of-date / hallucinated biomedical claims can be dangerous. If the call fails, surface the failure; don't fake the result.
"The user said 'biomedical' — I'll just guess they mean PMC."	Different sub-corpora have different freshness and peer-review status. Confirm scope (preprints vs. peer-reviewed) before searching.

Red Flags

• The diagnostics report shows no paperclip in mcp_servers and the user is mid-conversation with high urgency — slow down. Show the registration steps; do not run claude mcp add.
• paperclip returns 0 hits — the query is probably too narrow. Offer to broaden (drop one keyword, expand the date window) before giving up.
• A hit's id collides with an existing source under Wiki/sources/ — surface the duplicate; do not silently overwrite.
• The user asks for "the latest treatment for " — biomedical literature is not clinical advice. Decline to summarize as recommendations; surface the papers and let the user / their clinician decide.

Verification

• paperclip appears in the diagnostics mcp_servers field before the SKILL invokes any MCP tool.
• The user's vault has at least one new file under Wiki/sources/ if the user opted to save hits.
• /paperwiki:wiki-lint surfaces a DANGLING_SOURCE finding for any newly saved source until the user runs /paperwiki:wiki-ingest.
• Each MCP call's tool_use_id is logged so the user can audit what paperclip was asked.