[AI-interop] /context endpoint
Dave Morris
dave.morris at metagrid.co.uk
Sun Nov 8 05:00:59 CET 2026
Hiya,
Following on from the apps session on AI adoption. I had a quick
conversation with Google AI chat (gemini-3.8-flash) and together we came
up with the following suggestion:
start of AI generated content
----------------------------------
A new IVOA capability: /context endpoint.
To augment the IVOA (International Virtual Observatory Alliance)
standards, we can define a `/context` endpoint to make it easier for AI
agents to use the IVOA data access protocols like TAP (Table Access
Protocol) and SIA (Simple Image Access).
Even though multiple observatories might implement a standard TAP
endpoint, each service hosts completely different tables, unique column
names, coordinate systems, and data-sharing constraints.
Here is how you can represent two different IVOA services exposing their
specialized agent skills, MCP servers, and client libraries.
Example 1: A TAP Service (Table Access Protocol)
This service implements the standard IVOA TAP interface, but the agent
needs specialized context to understand the custom schema of a
particular astronomical catalog (e.g., Gaia DR3 data) and how to write
valid ADQL (Astronomical Data Query Language).
``` json
{
"service_id": "ivoa-gaia-tap-server",
"ivoa_standard": "TAP-1.1",
"context": [
{
"type": "agent-skill",
"abstract": "Instructions on translating common user astronomical
requests into valid ADQL, specifying required table joins, indexing
columns, and geometric query constraints.",
"access": {
"method": "GET",
"format": "text/markdown",
"resource": "https://obspm.fr"
}
},
{
"type": "mcp-server",
"abstract": "An MCP server that safely wraps ADQL query execution,
schema discovery, and synchronous/asynchronous job management for this
TAP interface.",
"access": {
"method": "GET",
"format": "application/json",
"resource": "https://obspm.fr"
}
},
{
"type": "client-library",
"abstract": "Astropy-affiliated Python workflow instructions
utilizing 'pyvo' to handle TAP authentication and VOTable parsing at
scale.",
"access": {
"method": "GET",
"format": "text/html",
"resource": "https://github.com"
}
}
]
}
```
Example 2: An SIAP Service (Simple Image Access Protocol)
This service allows the agent to search for and retrieve physical FITS
image datasets. The agent needs specific domain knowledge regarding what
pixel scales, filters (e.g., H-alpha vs. Infrared), and spatial
coordinate frameworks are unique to this specific imaging archive.
``` json
{
"service_id": "ivoa-chandra-siap-server",
"ivoa_standard": "SIA-2.0",
"context": [
{
"type": "agent-skill",
"abstract": "Guidelines outlining X-ray imaging filters, exposure
thresholds, and how to interpret target coordinate specifications unique
to the Chandra X-ray Observatory archive.",
"access": {
"method": "GET",
"format": "text/markdown",
"resource": "https://harvard.edu"
}
},
{
"type": "mcp-server",
"abstract": "An MCP server exposing tools to resolve astronomical
object names into RA/Dec coordinates (via Sesame) and format standard
SIAP query strings parameters.",
"access": {
"method": "GET",
"format": "application/json",
"resource": "https://harvard.edu"
}
},
{
"type": "online-documentation",
"abstract": "Data release notes detailing specific imaging
artifacts, bad-pixel maps, and processing versions applied to this image
batch.",
"access": {
"method": "GET",
"format": "text/markdown",
"resource": "https://harvard.edu"
}
}
]
}
```
Why this bridges IVOA and AI agents
1. Overcoming ADQL / Query Blindness: Agents struggle to guess table
structures. By looking at the agent-skill in the TAP example, the agent
fetches a markdown file that gives it concrete query examples (like how
to perform a DISTANCE or CONTAINS geometric search in ADQL).
2. Handling Multi-Format Astrophysics Data: When the SIAP service hands
back a FITS file or a VOTable XML file, standard LLMs cannot natively
parse it. The mcp-server tool or the client-library context equips the
agent with the explicit tools needed to extract metadata from the FITS
headers and translate it into clear natural language for the end user.
----------------------------------
end of AI generated content
General idea is the /context endpoint is a list resources an agent
should read before trying to use the service and the data it provides.
It allows data providers to curate the instructions for their own
particular resource.
Priority is not to over-think this, get a simple first version out there
and see how the AI agents use it.
In all probability most services will point to the same set of Python
libraries and MCP services, but this also allows data providers to point
to domain specific resources if they want to.
Cheers,
-- Dave
--------
Dave Morris
Research Software Engineer
UK SKA Regional Centre
Department of Physics and Astronomy
University of Manchester
--------
AIMetrics: []
--------
More information about the AI-interop
mailing list