Scienceclaw Retrieval

scienceclaw-retrieval

Retrieve scientific information from databases, literature repositories, and knowledge bases using structured search strategies, relevance ranking, and citation chaining.

When to Use

• Finding relevant papers on a specific research topic or question
• Querying scientific databases (PubMed, arXiv, Semantic Scholar, CrossRef, OpenAlex)
• Cross-referencing findings across multiple sources and databases
• Building comprehensive bibliographies for a research project or review
• Conducting systematic literature searches with reproducible methodology
• Tracking citation networks to discover related or derivative work
• Locating datasets, code repositories, or supplementary materials linked to publications

When NOT to Use

• Answering specific scientific questions -- use scienceclaw-qa
• Summarizing papers or synthesizing findings -- use scienceclaw-summarization
• Running data analysis or computations on retrieved data -- use code-execution skill
• Extracting structured information from paper text -- use scienceclaw-ie
• Verifying claims or checking calculations -- use scienceclaw-verification

Multi-Database Search Strategies

Parallel Search: For broad discovery, query PubMed, arXiv, Semantic Scholar, and OpenAlex simultaneously, collect results with DOIs and metadata, deduplicate by DOI, apply relevance ranking, then filter by date/type/discipline.

Sequential Refinement: For targeted retrieval, start with a broad query to gauge the landscape, analyze initial results for recurring keywords and author clusters, refine with Boolean operators and filters, snowball via citation chaining on top hits, and stop at saturation (new queries return mostly known results).

Systematic Review Search: Define PICO/PEO framework, construct Boolean queries with synonyms and controlled vocabulary, document every query/database/date/count for reproducibility, include grey literature (preprints, proceedings, registries), screen via title/abstract then full-text phases, and track numbers through a PRISMA flow diagram.

Database-Specific Query Syntax

• PubMed: MeSH terms via [MeSH Terms], field tags [tiab]/[au]/[dp], capitalized Boolean operators, [pt] for publication type. Example: "machine learning"[tiab] AND "drug discovery"[tiab] AND "2023"[dp]
• arXiv: Field prefixes ti:/au:/abs:/cat:, Boolean AND/OR/ANDNOT, category codes (cs.AI, q-bio.BM), trailing wildcards. Example: ti:"neural network" AND cat:cs.LG AND au:bengio
• Semantic Scholar: API parameters query/year/fieldsOfStudy/venue, field filtering, pagination via offset/limit, direct lookup by DOI or arXiv ID
• CrossRef: /works?query= endpoint, filters like from-pub-date:2023,type:journal-article, field queries query.title=/query.author=, sort by relevance/published/is-referenced-by-count
• OpenAlex: Entity endpoints /works//authors//sources, filters with commas (AND) or pipe (OR), search= for full-text, group_by= for aggregation, open access filtering via open_access.is_oa:true

Relevance Ranking

Combine multiple scoring signals: textual similarity between query and title/abstract (primary), citation count with recency weighting, publication date, venue quality (impact factor or acceptance rate), author authority (h-index in subfield), and reference overlap with known relevant papers. For active fields, apply time-decayed citation scoring: adjusted_score = citation_count / (current_year - publication_year + 1). Support user-guided re-ranking by marking papers as highly relevant, somewhat relevant, or not relevant, then refine queries using terms from top-marked papers.

Citation Chaining

• Forward chaining (cited-by): From a seed paper, find all papers that cite it via Semantic Scholar or OpenAlex, filter by date/venue/topic, repeat for new relevant hits (limit depth to 2-3 hops)
• Backward chaining (references): Extract the seed paper's reference list, score references by co-occurrence frequency across your relevant set, identify foundational works
• Co-citation analysis: Gather citation neighborhoods of 3-5 seed papers, find papers appearing in multiple neighborhoods as conceptually related candidates
• Bibliographic coupling: Find papers sharing high reference overlap with seed papers, indicating they address similar research questions

Deduplication

• DOI-based: Primary key for deduplication; prefer the record with richest metadata when merging
• Fuzzy title matching: For records without DOIs, normalize titles (lowercase, strip punctuation/articles), apply Jaccard > 0.85 or edit distance ratio > 0.90, verify by checking author overlap and publication year
• Preprint-publication linking: Match arXiv preprints to journal versions via DOI metadata or title matching, prefer published version but retain preprint if it has additional content (appendices, code), flag substantial differences between versions

Integration with Specialized Skills

• PubMed: biomedical and life sciences, MeSH controlled vocabulary, structured abstracts, clinical trial metadata
• arXiv: physics, math, CS, quantitative biology preprints, open-access full-text PDFs, new submission monitoring
• Semantic Scholar: cross-disciplinary search, citation graph features, TLDR summaries, influential citation filtering
• CrossRef: DOI resolution, comprehensive metadata, funding and license data, reference lists, citation ambiguity resolution
• OpenAlex: large-scale bibliometrics, trend discovery, open-access links via Unpaywall, concept tagging for topic filtering

Output Format

Single Query Result

Query: [Search query text]
Database(s): [Databases searched]
Total Results: [Count]
After Deduplication: [Count]

Top Results:
  1. [Title] | [Authors] | [Year] | [Venue]
     DOI: [DOI] | Citations: [Count]
     Relevance: [Score] | Abstract: [First 200 chars...]

Systematic Search Report

Search Strategy Report
======================
Research Question: [PICO-formatted question]
Date Executed: [Date]

Database Searches:
  - PubMed: [Query] -> [N results]
  - arXiv: [Query] -> [N results]
  - Semantic Scholar: [Query] -> [N results]

Total Retrieved: [N]
After Deduplication: [N]
After Title/Abstract Screening: [N]
Final Included: [N]

Included Papers:
  [Numbered list with full bibliographic details]

Zero-Hallucination Rule

ALL factual claims, citations, database results, and scientific data presented to the user MUST come from actual tool results (API calls, code execution, web search) in this conversation. NEVER fabricate or "fill in" details from training data. If a tool returns no results or partial data, report exactly what happened.