scix-mcp

name

scix-mcp

description

Use the SciX MCP server for scientific literature research across arXiv and the NASA SMD disciplines (astrophysics, planetary science, earth science, heliophysics, and biological & physical sciences as they relate to space science). Provides hybrid semantic+lexical search (INDUS + body BM25 via RRF), citation graph traversal (299M edges, PageRank, communities), entity extraction, and session-aware working sets. Invoke when the user asks about scientific papers, citations, authors, or wants to explore the literature in these domains.

SciX MCP — Scientific Literature Research

A local-intelligence MCP over the full NASA ADS / SciX corpus (32M papers, 299M citation edges, 14.9M paper bodies) spanning arXiv and the NASA SMD disciplines: astrophysics, planetary science, earth science, heliophysics, and biological & physical sciences as they relate to space science. Adds dense semantic retrieval, citation graph analytics, and entity extraction on top of the underlying metadata.

When to Activate

• User asks to find papers on a topic across arXiv or any NASA SMD discipline (astrophysics, planetary science, earth science, heliophysics, biological & physical sciences)
• User wants to trace citations or references of a paper
• User asks about an author's work, research lineage, or impact
• User asks "what's the state of research on X" — multi-hop exploration
• User wants to find gaps in a research area
• User wants to explore a specific paper's body text (methods, equations buried in full-text)
• User mentions bibcodes (ADS identifiers like 2019ApJ...875L...1E)

Tool Overview (15 agent-visible tools)

Category	Tools	When to use
Search	search, concept_search, lit_review	Find papers by query. Default: search (hybrid mode). Use concept_search for formal taxonomy terms (UAT concepts); lit_review to open a seeded review session in one call.
Paper access	get_paper, read_paper	Get full metadata / read sections/body text of a known bibcode.
Citation graph	citation_traverse, citation_similarity, forward_citations	citation_traverse (mode=graph neighbors / chain shortest path), co-citation / coupling similarity, and forward_citations (papers citing a target, annotated by intent or replication relation).
Entity system	entity	Resolve / list-papers / profile named entities (instruments, methods, datasets, objects) via the action enum (resolve / papers / search / profile).
Structure	graph_context, find_gaps, temporal_evolution, facet_counts	Community/topic exploration, gap analysis, field-over-time trends.
Provenance / synthesis	claim_blame, synthesize_findings	Trace a claim to its earliest origin; bin a working set into a section outline.

Legacy names (citation_graph, citation_chain, entity_context, cited_by_intent, find_replications, semantic_search, …) still resolve as deprecated aliases but are not advertised; prefer the consolidated names above.

Typical Workflows

1. Topic survey → citation exploration

search("dark matter halo mass function", limit=20)
  → pick top-cited bibcodes
citation_traverse(bibcode, mode="graph", direction="backward")
  → foundational prior work (references)
citation_traverse(bibcode, mode="graph", direction="forward")
  → recent extensions (citing papers)

2. Body-only discovery (new capability)

Use when a technique/equation is likely buried in the body text (methods section) rather than the abstract. search with mode="hybrid" (default) already includes body BM25 as a 4th RRF signal. No extra step needed — it just surfaces papers the title/abstract search misses.

3. Author impact

search("first_author:'Hogg, David W.'", limit=50)  # via filter
temporal_evolution(author="Hogg, David W.")

4. Entity-driven research

entity(action="resolve", query="JWST NIRSpec")
entity(action="profile", entity_id=<id>)   # full entity profile (formerly entity_context)
entity(action="papers", entity_id=<id>)     # papers tagged with the entity

5. Gap finding

search("exoplanet atmosphere retrieval", limit=30)
  → add top 10 to working set (use `add_to_working_set` in session)
find_gaps(working_set_id)
  → under-cited or under-researched angles

Filter Syntax

search and related tools accept filters:

• year_min / year_max (integers)
• arxiv_class (e.g., "astro-ph.EP", "astro-ph")
• doctype (e.g., "article", "review")
• first_author (string, exact match on normalized name)

Ranking Model (hybrid search)

4 signals fused via Reciprocal Rank Fusion (k=60):

1. INDUS dense embedding (768d, NASA's domain-specific) — semantic similarity
2. text-embedding-3-large (3072d, OpenAI) — optional, skipped if not available
3. Title+abstract BM25 (GIN on papers.tsv) — exact term matching on metadata
4. Body BM25 (GIN expression index on papers.body) — matches full-text body, ~47% coverage

Body matches are ranked by the title/abstract tsvector to keep latency sub-second. For most queries, the hybrid total is <1s.

Performance Characteristics

Operation	Typical latency
Keyword/BM25 search	50-500ms
Hybrid search (all 4 signals)	300-900ms
Citation chain (depth 1-2)	20-100ms
Citation chain (depth 3+)	1-10s
Entity resolution	<50ms
Body BM25 alone	400-800ms
read_paper for full body	100-300ms

Working Set / Session

If the MCP server supports session state, build a working set with add_to_working_set for iterative refinement. find_gaps, citation_similarity, and some summaries operate over the working set. Clear with clear_working_set to start fresh.

Connection Config

For Claude Desktop / Claude Code:

{
  "mcpServers": {
    "scix": {
      "url": "<https://...trycloudflare.com>/mcp/",
      "headers": {"Authorization": "Bearer <token>"}
    }
  }
}

Local stdio (if running the server locally):

{
  "mcpServers": {
    "scix": {
      "command": "/path/to/.venv/bin/python",
      "args": ["-m", "scix.mcp_server"],
      "cwd": "/path/to/scix_experiments"
    }
  }
}

Dos and Don'ts

Do:

• Start with search in default (hybrid) mode — it combines all 4 signals
• Pass a year filter for recent work (e.g., year_min=2023)
• Chain: search → pick bibcode → citation_chain or read_paper for depth
• Use find_gaps on a working set of 10-20 papers for recommendation quality

Don't:

• Make many parallel calls — rate limit is 60/min per token
• Query with just one word (use phrases for hybrid ranking to shine)
• Assume every paper has a body — ~47% coverage. Body BM25 contributes where available, doesn't when not.
• Expect concept_search to work on arbitrary phrases — it maps to the UAT (Unified Astronomy Thesaurus) taxonomy only.

Corpus Scale Facts

• Papers: 32.4M (full ADS metadata)
• Citation edges: 299M
• INDUS embeddings: 32M papers (complete coverage)
• Body text: 14.9M papers (~47% coverage, ~65K chars avg)
• Entity dictionary: ~90K entities (instruments, datasets, methods, objects, missions, software)
• Body GIN index: 37 GB partial expression index on papers.body