About the Knowledge Hub

All API endpoints are at /api/v1/ and support ?format=text for LLM-friendly plain text. See /llms.txt for a complete list. Examples:

# Search for publications about alpine pollination
curl "https://rmblknowledgehub.org/api/v1/search?q=alpine+pollination&format=text"

# Get publication details
curl "https://rmblknowledgehub.org/api/v1/publications/13?format=text"

# Explore a research neighborhood with primer
curl "https://rmblknowledgehub.org/api/v1/neighborhoods/620?format=text"

# Look up a species
curl "https://rmblknowledgehub.org/api/v1/entities/species/8426?format=text"

# Find related works
curl "https://rmblknowledgehub.org/api/v1/related/publications/13?format=text"

The MCP server gives Claude Desktop (and other MCP-compatible tools) direct access to 8 Knowledge Hub tools: search, publication/dataset/document lookup, entity detail, related works, and neighborhood exploration.

Step 1: Clone the repository and build the MCP server:

git clone https://github.com/ikb-rmbl/RMBL_knowledge_hub.git
cd RMBL_knowledge_hub/mcp
npm install
npm run build

Step 2: Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "rmbl-knowledge-hub": {
      "command": "node",
      "args": ["/path/to/RMBL_knowledge_hub/mcp/dist/index.js"],
      "env": {
        "RMBL_API_URL": "https://rmblknowledgehub.org"
      }
    }
  }
}

Step 3: Restart Claude Desktop. You should see 8 tools available (hammer icon). Try asking:

• “Search for publications about marmot hibernation at RMBL”
• “What is research neighborhood 620 about?”
• “Find works related to publication 13”
• “Look up the species Marmota flaviventer”

Publications are sourced from the RMBL publications database, with additional discovery via OpenAlex and CrossRef. Each record is enriched with metadata from CrossRef (authors, DOIs, abstracts, citation counts) and Unpaywall (open access links). Full text is extracted from PDFs using pdftotext with OCR fallback via Tesseract.

Datasets are discovered from eight repository sources including EDI, DataONE, Dryad, Zenodo, USGS ScienceBase, Pangaea, NCBI, and Figshare. Each dataset is enriched with EML/DataCite metadata including temporal and spatial coverage, creator information, and licensing.

Documents come from the Sustainable Living Library, a collection of community and policy documents relevant to the Gunnison Basin. These include management plans, environmental impact statements, water quality reports, and local planning documents.

Authors are deduplicated across all collections using a two-phase process. First, authors with matching ORCID identifiers are merged. Then, authors sharing the same family name are compared by given name initials, with checks to prevent false merges when middle initials differ (e.g., “R. J. Smith” is kept separate from “R. A. Smith”). Author ordering on publications is repaired from CrossRef metadata to ensure correct first-author attribution.

Entities (species, concepts, protocols, places, and stakeholders) are extracted from publication and document full text using Claude vision models (VLM extraction). Each entity mention is linked to its source item with a confidence score and extraction method. Entities are then deduplicated using embedding-based clustering (Voyage AI voyage-4, 1024 dimensions) with type-specific similarity thresholds.

Species names are validated against the ITIS (Integrated Taxonomic Information System) database. Places are enriched with coordinates from GNIS (Geographic Names Information System) and organized into a parent-child hierarchy.

The resulting knowledge graph has 93,062 entity mentions linking items to entities, plus 115,075 citation references with internal cross-links between publications.

Knowledge Neighborhoods are detected using the Louvain community detection algorithm on the unified knowledge graph. The graph includes all entities and items as nodes, with edges from co-occurrence in publications, co-authorship, and citations. Edge weights are boosted for structural relationships (co-authorship ×5, citations ×3) to ensure that social and citation structure drives community boundaries rather than just shared terminology.

Research primers are generated for the largest neighborhoods using Claude (Opus model) with tiered context assembly: landmark papers (full abstracts + key findings), frontier papers (2020+), breadth papers (single best finding each), and entity context (species, concepts, methods, places). Each primer includes parenthetical citations linked to specific publications in the Hub. Policy-focused neighborhoods receive primers with document citations instead.

Full-text search uses PostgreSQL tsvector with weighted ranking (title > abstract > full text) and stemmed query matching. Search results include highlighted snippets via ts_headline.

Related works are found using four signals: semantic similarity (pgvector cosine distance on Voyage AI embeddings), shared entity mentions (at least 3 shared entities), co-authorship (shared authors across publications), and citation links (from the references_cited table). Signals are merged with a multi-signal bonus for items connected by multiple pathways.

The Knowledge Hub is built with Next.js and Payload CMS on PostgreSQL with pgvector. Graph visualizations use Sigma.js (WebGL). The data pipeline is a set of TypeScript scripts for scraping, enrichment, entity extraction, and graph construction. Vector embeddings are generated by Voyage AI (voyage-4, 1024 dimensions). The site is hosted on Vercel with the database on Neon (serverless PostgreSQL).

The project is open source at github.com/ikb-rmbl/RMBL_knowledge_hub.