README.md

metadata

title: Ask CGIAR Research
emoji: 🌾
colorFrom: green
colorTo: yellow
sdk: docker
app_port: 7860
pinned: true
license: apache-2.0
short_description: MCP server to search and cite CGIAR research publications
tags:
  - mcp
  - mcp-server
  - agriculture
  - food-security
  - retrieval
  - cgiar

Ask CGIAR Research

A public Model Context Protocol server that lets AI assistants search 29,556 openly licensed CGIAR research publications (CC-BY, CC-BY-SA, CC0; English, Spanish, French, Portuguese) and answer agrifood questions with citations.

Server URL: https://cgiar-ask-cgiar-research.hf.space/mcp (streamable HTTP, no sign-in)

Code: the main copy is the Hugging Face Space, which runs the service. github.com/cgiar/ask-cgiar-research is a mirror for browsing, issues and installing the plugin.

Tools

Tool	What it does
search_cgiar_publications	Hybrid search (BM25 keyword + multilingual BAAI/bge-m3 embeddings, reciprocal rank fusion) over ~1.0M passages, with country, region, year, language and full-text filters
get_cgiar_document	Metadata, abstract and passages around a hit, or the whole document

Connect

Claude Code, plugin with search guidance (recommended):

claude plugin marketplace add cgiar/ask-cgiar-research
claude plugin install ask-cgiar-research@cgiar

Claude Code, server only:

claude mcp add --transport http cgiar-research https://cgiar-ask-cgiar-research.hf.space/mcp

Claude.ai / Claude Desktop:

1. Go to Customize → Connectors → Add custom connector, paste the server URL, and choose no sign-in.
2. Optionally upload plugin/ask-cgiar-research-skill.zip under Customize → Skills.

Other MCP clients: add the URL as a remote streamable-HTTP MCP server.

Data

The index is built from the CGIAR agrifood research corpus (GARDIAN-indexed CGIAR publications), restricted to documents whose license allows redistribution: CC-BY 27,771, CC-BY-SA 1,630 and CC0 155. Of these, 24,779 are full text and 4,777 are abstracts. The text was extracted from PDFs, and tables and figure text were removed. Each result carries its document's license and source URL; reuse must follow that license.

Answers generated with this tool are not official CGIAR positions. Coverage is uneven: when nothing is found, that is not evidence that no CGIAR research exists on the topic.

Privacy and usage data

For reporting, each tool call is logged with:

• the time and the tool;
• the search query after automated removal of personal information (names, emails, phone numbers, URLs, ID numbers, precise coordinates);
• the filters, the documents returned and the response time;
• the client application name;
• a session id that rotates daily.

IP addresses and raw queries are not stored. Logs are kept in a private CGIAR dataset for 24 months, then deleted automatically. Automated removal is imperfect, so please don't put personal details in questions.

Repository layout

app.py            MCP server (tools, landing page, /health)
search_core.py    hybrid search over SQLite FTS5 + dense vectors
pii.py            query scrubbing before logging
usage_log.py      scrubbed usage logs -> private HF dataset, 24-month retention
static/           landing page
build/            scripts that build the index from the corpus
plugin/           Claude Code plugin + claude.ai skill

Configuration (Space variables and secrets):

Name	Purpose
HF_TOKEN	secret, fine-grained: read INDEX_REPO, write LOG_REPO
INDEX_REPO	default CGIAR/ask-cgiar-research-index
LOG_REPO	e.g. CGIAR/ask-cgiar-research-logs (private)
RATE_PER_MIN	default 60

Contact

Questions and feedback: Jawoo Koo, via the IFPRI contact form.

Credits

Owned and operated by CGIAR. Developed under the GAIA project, funded by the Gates Foundation and UK International Development (FCDO), with the support of the CGIAR Digital Transformation Accelerator program. Code licensed under Apache-2.0.