🔭 astrophysics

A Model Context Protocol Server for Astrophysical RAG: Unified Access to HI, Dwarf, Globular Cluster, IntZ, and ALPINE Kinematic Corpora with FAISS Semantic Search

The paper introduces the EPS Research Astro-RAG MCP Server v2.3.0, a unified, cross-platform system that provides machine-readable access to five astrophysical kinematic corpora spanning redshifts 0 to 5.68 through FAISS-accelerated semantic search, REST APIs, and an LLM-native interface.

Original authors: David C. Flynn

Published 2026-08-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: David C. Flynn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a massive, cosmic library stretching from our own backyard to the very edge of time. In this library, astronomers have been collecting books about different types of stars and galaxies for decades. But here's the catch: the books are written in different languages, stored in different rooms, and organized by different librarians. One room has books about slow-spinning galaxies nearby, another has books about tiny, messy dwarf galaxies, and a third holds ancient books about galaxies from when the universe was just a toddler. If you wanted to find a specific type of galaxy that spins fast, you'd have to run between these rooms, translate the catalogs, and manually check every single book. It's a lot of work, and it's easy to miss the connections between the different sections.

To make things easier, scientists are starting to use "smart assistants" (called Large Language Models) that can talk to us in plain English. But these assistants need a special bridge to talk to the library's data. This bridge is called the Model Context Protocol (MCP), which acts like a universal translator, letting the smart assistant ask questions and get precise answers without needing to know the complex code behind the scenes. The big question is: Can we build one of these bridges that connects all these different galaxy books at once, so a smart assistant can find patterns across the entire history of the universe?

David C. Flynn has built exactly that bridge. He created a new digital tool called the EPS Research Astro-RAG MCP Server v2.3.0. Think of this server as a super-smart, all-knowing librarian who has organized 2,064 different "galaxy books" into a single, easy-to-search pile. These books cover everything from our local neighborhood (where the redshift, or "cosmic distance," is 0) all the way out to the farthest reaches of the early universe (redshift 5.68).

The library contains five specific collections:

The "Unified HI Rotation Curve" collection: 438 galaxies that spin like merry-go-rounds right here in our local universe.
The "Dwarf/Irregular" collection: 129 small, messy galaxies that don't have a neat shape.
The "Milky Way Globular Cluster" collection: 174 tight groups of ancient stars orbiting our own galaxy.
The "IntZ" collection: 1,292 galaxies from the middle era of the universe (when it was about 0.4 to 2.7 times its current age).
The "High-z ALPINE" collection: 31 very distant, ancient galaxies from the universe's teenage years (redshift 4.26 to 5.68).

The magic of this new server is how it lets you search. Before, if you wanted to find a galaxy, you had to know its exact ID number or its specific scientific name. If you didn't know the name, you were stuck. But this server uses a clever trick called FAISS semantic search. Imagine you have a giant pile of index cards, and instead of writing "Galaxy A" on the card, you write a description like "a small, messy galaxy with low surface brightness." The server turns your question into a mathematical "fingerprint" and finds the cards that have the most similar fingerprints.

So, instead of typing a complex computer command, you can just ask the server, "Find me dwarf irregular galaxies that are low mass," or "Show me metal-poor globular clusters in the outer halo." The server understands the meaning of your words, not just the spelling. It then uses its pre-made "fingerprints" (created using a model called MiniLM-L6-v2) to instantly pull up the top matches from its collection of 2,064 objects.

The paper shows that this system works incredibly well. When the author tested it with questions like "dwarf irregular low mass," the server correctly returned specific galaxies like CVnIdwA and DDO 210. When asked for "metal poor outer halo cluster," it found the right star clusters. It even works across different eras of the universe; you can ask it to find galaxies similar to a nearby one, and it can find matches in the distant, ancient collections.

The server is built to be used by both humans and AI assistants. It has a website where you can click and search, a standard computer interface for other programs to talk to it, and a special "MCP" door that lets AI assistants (like the ones you might chat with) ask questions directly. The author emphasizes that this isn't just a list of data; it's a unified system where the data is cleaned up and organized so that a galaxy from the nearby universe and a galaxy from the distant universe can be compared side-by-side without confusion.

The paper is very clear about what this tool is not. It is not a replacement for the original, live databases where astronomers go to get brand-new, fresh data. It is also not a tool that measures the physical distance between galaxies in space; it only measures how similar their descriptions are. If you need to know the exact speed of a galaxy, you use the server's "filter" tool to get the hard numbers. If you want to discover a new type of galaxy based on how it "feels" in a description, you use the semantic search.

The author has made sure everything is open and reproducible. The code, the data, and the special "fingerprints" used for the search are all available for anyone to download and check. The server is currently running on a public platform called HuggingFace Spaces, meaning anyone with an internet connection can try it out. This represents a first step in a new way of doing astronomy: instead of being a detective who has to hunt through dusty archives, researchers (and their AI helpers) can now have a conversation with the entire history of the universe's kinematic data, asking questions in plain English and getting structured, reliable answers in return.

Technical Summary: EPS Research Astro-RAG MCP Server v2.3.0

Problem Statement
The proliferation of multi-epoch astrophysical survey data has created a significant retrieval bottleneck. While individual corpora (e.g., local HI surveys, high-redshift kinematic targets) are well-documented within their specific communities, they remain difficult to query programmatically across different epochs, surveys, or object classes without extensive custom tooling. Researchers face heterogeneous data formats, inconsistent field naming conventions, and fragmented access infrastructures. Furthermore, while Large Language Models (LLMs) offer potential for natural-language data access, their utility is currently limited by a lack of structured, deterministic, and low-latency retrieval infrastructure capable of serving corpus data directly to AI assistants.

Methodology
The author presents the EPS Research Astro-RAG MCP Server v2.3.0, a cross-platform implementation of the Model Context Protocol (MCP) designed to unify access to five distinct astrophysical kinematic corpora. The system aggregates 2,064 objects spanning redshifts from $z=0$ to $z=5.68$ , including:

Unified HI Rotation Curve Corpus (438 galaxies).
Dwarf/Irregular HI Corpus (129 galaxies).
Milky Way Globular Cluster Corpus (174 clusters).
IntZ Kinematic Corpus (1,292 galaxies, $z \approx 0.4–2.7$ ).
High-z ALPINE Kinematic Corpus (31 galaxies, $z = 4.26–5.68$ ).

Architecture and Stack
The server is implemented in Python using FastMCP and FastAPI, exposing three concurrent interfaces over a single process:

MCP Endpoint: An LLM-native tool API using Server-Sent Events (SSE) and Streamable HTTP, allowing AI assistants (e.g., Claude, Copilot) to invoke typed tools directly.
REST API: Standard HTTP endpoints with OpenAPI documentation for programmatic access.
Browser UI: A single-page HTML interface for point-and-click querying and result export.

Data is loaded from JSONL files into an in-process cache at startup. To enable semantic search, the system utilizes FAISS (Facebook AI Similarity Search) with pre-built vector indexes. These indexes are constructed offline using the all-MiniLM-L6-v2 model (from sentence-transformers), generating 384-dimensional dense embeddings. Crucially, embeddings are derived from curated summary strings of key metadata fields (e.g., morphology, distance, kinematic class) rather than raw JSON records, ensuring the vector space captures scientifically meaningful content. The system uses faiss.IndexFlatL2 for exact nearest-neighbor search, guaranteeing deterministic results.

Key Contributions and Features

Unified Schema: All five corpora share a common JSON schema with native-typed fields and consistent naming conventions, enabling cross-corpus queries.
MCP Toolset: The server exposes seven specific tools:
- list corpora, list objects, get object, search metadata, filter objects, get corpus schema, and semantic search.
- These tools allow LLMs to perform complex tasks, such as filtering by numeric ranges (e.g., "rotation velocity > 150 km/s") or semantic discovery (e.g., "dwarf irregular low surface brightness") without the user needing to know specific field names or query syntax.
FAISS-Accelerated Semantic Search: Version 2.3.0 introduces natural-language similarity search across all corpora. This allows for discovery queries that return semantically ranked object lists without requiring knowledge of the underlying database schema.
Reproducibility and Deployment: The server is publicly deployed on HuggingFace Spaces and is fully reproducible. All components (server software, five corpora, and FAISS indexes) are versioned and assigned Zenodo DOIs. A local launcher supports deployment on Windows and Ubuntu.

Results and Validation
The paper demonstrates the system's efficacy through representative use cases:

Rotation Curve Retrieval: The system successfully retrieves full kinematic records (inclination, position angle, Vrot profiles) for specific objects (e.g., NGC3198) with sub-millisecond latency from the cache.
Metadata Filtering: Numeric range filters (e.g., distance between 1–10 Mpc) execute in under 100 ms, returning precise subsets of the data.
Semantic Exploration: The semantic search tool successfully retrieves physically coherent sets of objects based on natural language queries. For instance, querying "dwarf irregular low mass" correctly returns CVnIdwA, DDO 210, and IC 1613. Similarly, "metal poor outer halo cluster" returns ESO-SC06, 2MS-GC02a, and ESO 93-8.
Cross-Epoch Consistency: The system enables comparisons between local ( $z=0$ ) and high-redshift ( $z \approx 5$ ) kinematics using a unified interface, a task previously hindered by disparate data formats.

Significance and Claims
The author claims this work represents the first MCP-compliant astrophysical data server to provide unified kinematic access across the full redshift range of $z=0$ to $z=5.68$ with semantic search capabilities. The platform is positioned as a complementary tool to existing infrastructure like astroquery (which targets live database queries) and the Virtual Observatory (which uses ADQL/TAP). While astroquery is suited for fresh, targeted catalog queries, the Astro-RAG MCP server is optimized for structured retrieval and semantic exploration of pre-assembled, schema-unified corpora.

The paper emphasizes that this architecture lowers the barrier to corpus exploration for researchers unfamiliar with specific database schemas, enabling AI assistants to compose complex queries (e.g., combining filters and metadata searches) without human intervention in the syntax. The system is released under the MIT License, ensuring open access and reproducibility.

Limitations
The author notes that FAISS indexes based on L2 distance capture semantic similarity in natural language descriptions but do not encode physical distance in parameter space; two galaxies may be semantically similar in embedding space but differ significantly in physical properties like inclination. Consequently, numeric similarity queries are better served by the filter objects tool, while semantic search is best utilized for discovery and reconnaissance. Additionally, the IntZ corpus currently contains records with missing identifiers that return numeric index IDs, a known issue slated for correction in future versions.

Technical Summary: EPS Research Astro-RAG MCP Server v2.3.0

More like this