Clinical Trial Literature Search / 临床试验文献检索专家
Search medical-domain scholarly literature with full data-source coverage (OpenAlex + Europe PMC/MeSH + NCBI PubMed/MEDLINE + bioRxiv/medRxiv preprints + arXiv methodology breadth), normalize, merge, and de-duplicate the results, produce a qualitative CSM safety-literature subset, and assist in providing open-access (OA) full-text PDF downloads. Integrated Docling PDF structured parsing. / 全数据源覆盖检索医学领域学术文献(OpenAlex + Europe PMC/MeSH + NCBI PubMed/MEDLINE + bioRxiv/medRxiv 预印本 + arXiv 方法学广度),归一化合并去重,可产出 CSM 安全性定性子集,协助提供 OA 文献 PDF 下载。集成 Docling PDF 结构化解析。
Data Egress Guardrail (CT family standard · F33)
This skill may call external endpoints (Coze services, shared third-party search endpoints, public APIs or LLMs). Internal confidential data must not leave the local domain. Mandatory rules:
- Never outbound: patient-level records, unpublished protocols / CSRs / internal documents, personal identifiers or sponsor-confidential content — as query terms, attachments or prompt payload.
- Allowed minimum: public search terms and non-content metadata only (machine hash, skill version, entry point, language, request id). Results from public registries (ct.gov / CDE / FAERS / OpenAlex / PubMed, etc.) are processed locally.
- Local first: computation, de-identification, aggregation and reporting stay on this machine; confirm the payload carries no confidential field before any cloud call.
- Visible disclosure: workbench search forms must state the outbound endpoint and this boundary in visible text, not only in a hover tooltip.
- Fail loud: stop and tell the user if any rule cannot be guaranteed; never degrade silently.
Machine-checked by ct-base/scripts/spec_lint.py rule F33.
Language
- English guide → README.md · Chinese guide → README_zh-CN.md
- Bilingual auto-switch: the answer language follows the user's question language (English question → English answer, Chinese question → Chinese answer).
Purpose
Retrieve published scholarly literature (peer-reviewed papers, systematic reviews, conference abstracts, preprints) about a drug / disease / method, normalize heterogeneous records from multiple public bibliographic sources into one de-duplicated evidence base, and surface the evidence landscape plus a CSM (cumulative safety monitoring) qualitative subset. Supports trial-planning background, protocol / CSR introductions, and published-safety literature checks.
Positioning within the ct- library
Four complementary public-intel skills: ct-registry (registered trials) · ct-literature (published evidence) · ct-safety (FAERS signals) · ct-pipeline (aggregate intel brief).
Boundaries: the safety lane here is qualitative published-safety literature — never a substitute for ct-safety's FAERS disproportionality. Unsure which skill? Route via ct-advisor (§15). Full four-skill matrix + boundary detail → references/capabilities.md.
Pipeline position: ct-registry landscape hypotheses seed a search here; the assembled evidence base then feeds ct-pipeline (intel evidence dimension) and supplies background for ct-protocol / ct-csr, while the safety subset corroborates ct-safety qualitatively.
Data Sources
Default lane: OpenAlex (primary) + Europe PMC + bioRxiv/medRxiv preprints. Opt-in: arXiv, Cochrane CDSR (via the Europe PMC journal filter), Semantic Scholar, PROSPERO. PubMed lane: an independent direct-to-NCBI channel covering both PubMed/MEDLINE and Cochrane CDSR. Off by default because Europe PMC already covers both; automatic fallback is on — when the Europe PMC leg fails (exception or zero hits) the PubMed lane takes over, restoring PubMed and Cochrane coverage. Lane state and the fallback reason are recorded in the evidence log. Local imports (no network): a local Zotero collection, and files the user exported from Chinese databases (CNKI / WanFang / SinoMed). These are imports, not retrieval — they join the same dedupe / verification / PRISMA pipeline. Clinical guideline corpus: a version-pinned, pointer-only local corpus, built once and then read offline; full texts are never shipped inside the skill, and a failed fetch degrades to a "not retrieved" pointer rather than a fabricated citation. Recency caveat: the publication-date window is pushed down into every source query and clipped again after merging; the online channel applies the same window but can still truncate the newest items at the result cap, so a hard recency guarantee requires the local engine. Per-source access, role and status table → references/capabilities.md.
Capabilities
- Search & filters: topic / drug / disease search; review-type, publication-year and month-level recency filters; relevance ranking; document-type inclusion and exclusion.
- Evidence assembly: multi-source merge, dedupe, citation ranking, MeSH / concept / funder normalization, PRISMA funnel.
- Safety subset: qualitative CSM tagging over published safety literature.
- Structured concepts: PICO-style condition / population / intervention / comparator / outcome axes, MeSH-normalized, with any unmapped term kept literal rather than invented.
- Citation-network expansion from one seed paper: start from a DOI, a PMID or a free-text title, and expand one hop backward (its references) and forward (works citing it); the seed itself is returned and tagged as such.
- Local-source enrichment: paragraph-level provenance anchors over already-downloaded PDFs, Zotero import, and Chinese-database import with cross-database CJK dedupe.
- Batch retrieval: a list of topics in one spec file, run sequentially with aggregated results, a format check, a call-quota gate and checkpoint resume; the quota gate estimates the run first and aborts before any network call when the estimate exceeds the daily safe threshold unless the user explicitly confirms.
- Deliverables & add-ons: HTML + Excel reports; on demand citations (BibTeX / RIS / styled), open-access PDF download, Obsidian notes, Zotero export.
- Anti-hallucination: citation verification, DOI audit, and an evidence provenance log.
- Honest source accounting: every source is logged as ok / empty / error / skipped, the run returns a coverage verdict (healthy / partial / empty / critical gap), disabled sources are recorded as skipped rather than zeroed, and a primary-source failure is flagged as a critical gap with an explicit warning against extrapolation; a narrowed query that returns too few unique works triggers one broader rescue run (single-shot, non-fatal) instead of a thin result set.
- PDF structured parsing: table- and hierarchy-preserving conversion (Docling primary, PyMuPDF / PdfPlumber fallback) for table-dense protocol, CSR and open-access full-text PDFs before any downstream reuse.
- Meta-analysis topic evaluation: early fatal-defect gates plus six-dimension scoring, offline and decoupled from retrieval.
Exact command syntax and the full parameter table → references/sop.md; the complete capability / parameter matrix → references/capabilities.md.
Outputs
Standard deliverables = HTML + Excel only — present lit_report.html; lit_report.xlsx keeps the complete filterable result. On-demand add-ons: citations, open-access PDF download, Obsidian notes, Zotero exports. After a PDF download, point users at the output folder rather than opening or listing individual files.
Full output catalogue, sheet layout, server-side supplement chain and concurrency notes → references/capabilities.md.
Published Application
The workbench is published as a Python HTTP service at https://ct-literature.app.workbuddy.host/ (appId wbapp_WuOPnJ4caJ32GTquNlR7vc, deploy bundle workbench/publish-dist/, workspace-independent). Re-publishing is always an overwrite of the same appId, never a new app, and the returned shareLink must be asserted against the expected domain — a mis-bound bundle path silently serves another app.
Full registry table and re-publishing protocol → references/published_app.md — load it before any deploy or re-publish operation.
⚠️ Safety
- Safe preview by default: the pipeline only generates and displays a plan; network requests run only after the user explicitly confirms execution.
- Reads public publications only — no confidential research or subject data is ever an input.
- The safety lane is qualitative — never feed it into FAERS disproportionality; it only corroborates
ct-safetyqualitatively. - Output is for reference / background only, not a regulatory submission.
Dialogue — Triage-First (mandatory)
Every natural-language turn MUST start at Step 0 triage — never skipped, never jump straight to execution (§6.2 + references/search_menu.md). Four buckets: Simple (topic + at least two parameters → preview) · Complex (menu → keyword gate → preview → run) · Vague (grill-me, at most two rounds, then re-triage) · Middle (answer directly). Red lines: never skip triage; never decide for the user (the Complex menu keeps the "explain the differences" entry); the keyword-system confirmation gate is mandatory whenever search terms are parseable. Cross-turn continuity (high-risk, mode A): the query spec lives only in the thread — after every run append a "current search spec" echo line carrying topic, review type, year, safety, sources, maximum results and verification state (never omit fields); follow-ups override only the changed fields and are re-merged by default; remote calls are stateless, so continuity is local-only. Full triage tables, echo / merge rules and the execution flow → references/dialogue.md — load it when executing a dialogue round.
AI-Generated Asset Governance (ct-base §22)
This skill is bound by
ct-base/BASE.md§22 (AI-generated asset governance). It produces search / synthesis output, not SDTM / ADaM / TLF / CSR assets; three obligations apply and all three are implemented: the RAG three-column to-verify checklist (question / source location / basis, derived from the real run ledger, zero fabrication), shadow-mode promotion thresholds (guardrail-only for this skill's LLM normalisation / translation / labelling layer), and layered verification (deterministic layer versus LLM layer disclosed separately, never blended into one headline accuracy). Red line: search / synthesis conclusions must never omit provenance, and uncovered items must be listed explicitly rather than assuming the search was complete (§22.16 + §17.1). Full obligation table → references/gov22_obligations_zh.md.
Where to Look
- Capabilities, parameters, per-source tables, output catalogue → references/capabilities.md
- Command examples and the parameter table → references/sop.md
- Batch spec format and quota / resume rules → references/batch_spec.md
- Triage tables, echo / merge rules → references/dialogue.md · keyword gate → references/search_menu.md
- Multi-database import contract → references/multi-db-search.md
- Meta-analysis topic-evaluation rubric → references/meta_topic_rubric.md
- Error handling → references/errors.md · maintenance and change notes → references/maintenance.md · published application and re-publishing protocol → references/published_app.md
- Downloads
- 1.6kAll time
- Security audit
- Review
- Last updated
- 6d ago
- Current version
- v1.7.10
- License
- MIT-0
