sciverse-paper-search

Sciverse Paper Search Skill

Overview

Use this skill when the user needs scientific literature retrieval, paper metadata screening, citation-ready evidence, or a research synthesis grounded in Sciverse search results.

This skill is adapted to the current LazyLLM SciverseSearch implementation. It must only rely on the currently supported tool capabilities:

• sciverse_search.search
• sciverse_search.meta_search
• sciverse_search.meta_catalog
• sciverse_search.get_content

Do not assume Sciverse MCP tools, resource APIs, binary attachment downloads, figure/table downloads, DianShi, or SeqStudio capabilities are available unless the runtime explicitly exposes those tools.

When To Use

Use this skill for:

• Finding scientific papers on a research topic.
• Retrieving citable evidence snippets for a scientific question.
• Screening papers by year, venue, DOI, author, title, or metadata fields.
• Building paper lists for literature reviews.
• Reading fuller text for selected Sciverse results when doc_id is available.
• Producing cited summaries, comparisons, or evidence tables from Sciverse results.

Do not use this skill for:

• Downloading paper images, figures, tables, PDFs, or binary resources.
• Chemical retrosynthesis, molecule/reaction search, or DianShi workflows.
• Protein sequence/structure annotation or SeqStudio workflows.
• Claims that require full-text access when only abstracts or snippets are available.

Available Tool Capabilities

sciverse_search.search

Use this for normal Agent retrieval.

Recommended defaults:

query=<research question or paper topic>
topk=5
search_type="agentic"
include_content=true

Use search_type="agentic" when the user asks a natural-language scientific question and needs evidence passages.

Use search_type="meta" when the user mainly needs paper metadata. You may pass year_from and year_to for year constraints.

Current implementation notes:

• topk is capped at 10.
• Results are normalized to title, url, snippet, source, and extra.
• extra may include doc_id, doi, year, venue, authors, score, chunk_id, page_no, offset, and content.

sciverse_search.meta_search

Use this for advanced metadata search, filtering, pagination, and paper-list tasks.

Important constraints:

• Do not use query together with sort.
• Do not use cursor together with page > 1.
• page_size is capped at 200.
• freshness_boost must be NONE, MILD, or STRONG.

Useful parameters:

query
filters
sort
fields
page
page_size
cursor
freshness_boost
include_content
year_from
year_to

Use year_from and year_to for simple publication-year filtering.

sciverse_search.meta_catalog

Call this before constructing complex filters or sort clauses if you are unsure which fields and operators are supported.

Use:

include_sample_values=false

Set include_sample_values=true only when enum-like sample values are needed.

sciverse_search.get_content

Use this to read fuller text for one search result.

The current implementation:

1. Looks for doc_id in the item or item.extra.doc_id.
2. Calls Sciverse /content with doc_id.
3. Supports chunked reading with offset and limit.
4. Falls back to extra.content, snippet, or URL fetching when /content is unavailable.

Do not claim full text was read unless get_content actually returns fuller content. If the result only contains an abstract or snippet, say that the analysis is based on metadata/snippets.

Workflow

Phase 1: Clarify Search Intent

Classify the user's request:

• Evidence answer: use agentic search.
• Paper list or screening: use meta search.
• Field-specific filtering: call meta catalog first.
• Deep literature review: combine agentic search and meta search.
• Read one selected paper: use get_content on a selected result.

Phase 2: Retrieve Papers

For natural-language evidence questions:

sciverse_search.search(query="<question>", topk=5, search_type="agentic", include_content=true)

For paper screening:

sciverse_search.meta_search(
  query="<topic>",
  fields=["title", "doi", "doc_id", "abstract", "author", "publication_published_year", "publication_venue_name_unified"],
  page_size=25,
  year_from=<optional>,
  year_to=<optional>
)

For precise filters:

1. Call sciverse_search.meta_catalog.
2. Build filters only from supported fields/operators.
3. Call meta_search.

Phase 3: Inspect and Read

For the most relevant results:

1. Extract title, DOI, year, venue, authors, doc_id, and snippet/content.
2. If the user needs deeper analysis, call get_content on selected items.
3. Use offset and limit for chunked reading when needed.

Example:

sciverse_search.get_content(item=<selected_result>, offset=0, limit=2000)

Phase 4: Synthesize With Source Discipline

When answering:

• Separate confirmed full-text evidence from abstract/snippet-only evidence.
• Cite papers using title, year, venue, DOI, and doc_id when available.
• Do not invent bibliographic fields.
• If Sciverse returns limited content, state the limitation.
• For literature reviews, group papers by theme, method, dataset, finding, limitation, and open question.

Output Patterns

Paper Search Results

Use a compact table:

| Paper | Year | Venue | Why relevant | DOI / doc_id |

Evidence Answer

Use:

• Short answer.
• Evidence bullets with paper identifiers.
• Caveats about snippet/full-text availability.
• Suggested next searches if coverage is thin.

Literature Review

Use:

• Search strategy.
• Included papers.
• Thematic synthesis.
• Method and evidence comparison.
• Limitations and open questions.
• Citation table.

Safety and Limitations

• Do not promise resource, attachment, figure, table, or PDF downloads.
• Do not use unsupported Sciverse MCP tool names.
• Do not fabricate citations, DOI values, doc IDs, authors, or venues.
• If authentication fails, tell the user Sciverse API access may need a valid API key or dynamic auth entry.
• If get_content falls back to snippets, clearly label the source as snippet/abstract-based.