Agentic Search for the NASA Planetary Data System

An MCP server and reproducible evaluation of multistep natural-language search over the NASA Planetary Data System (PDS) Registry -- the digital data archive for all of NASA's planetary missions, flight and ground-based observations, and laboratory experiments from the 1960s to the present. With more than 1.85 petabytes from 70+ missions, 4,500 datasets, and 700 instruments all peer-reviewed by JPL and other research institutions.

Evaluation study · Results · Dataset · Data engine · Benchmark and reproducibility · MCP server · Development guide

https://github.com/user-attachments/assets/1d6b7035-c07a-4ca3-8bd2-9c19d68c3d5c

Abstract

The NASA Planetary Data System (PDS) serves as a high-quality, hand-curated data source for the whole community of planetary research. However, searching the PDS Registry is difficult, requiring knowledge of exact PDS4 ontology fields and context identifiers to manually construct queries for its public search APIs. The research chain of scientific question -> experimentation -> findings/results is thus bottlenecked by current PDS search capabilities.

We introduce PDS-MCP, an agentic search interface on top of the NASA PDS Registry. From natural language questions, our agent can iteratively resolve PDS entities, inspect relationships, construct valid filters, and retrieve matching products autonomously constructing PDS API queries with no manual human intervention. We evaluate this on our 300-question benchmark, constructed from our scalable data engine, with results achieving 98.02% macro F1 and 95.67% exact result-set match on our full system, compared with 12.29% and 8.33% for single-call MCP and 0.33% on both metrics without PDS access, demonstrating that it can reliably bridge natural-language questions and structured PDS retrieval.

By releasing PDS-MCP publically, we aim to support the researchers of the Planetary Data Science community by making it easier to access NASA PDS data. Our goal is to provide enhanced search capabilities that enable more effective data exploration and improve accessibility for future research endeavors.

Evaluation study

Research question

Can a language-model agent reliably translate natural-language planetary-data requests into complete PDS Registry result sets, and does multistep MCP search outperform the same model with one or zero PDS tool calls?

Benchmark

The benchmark dataset contains 300 independent questions:

Data Engine

The complete benchmark data engine and its dataset documentation are included in this repository.

Ground truth was constructed as follows:

Select compatible PDS context objects → Build a PDS API query → Compile and execute it against the live PDS API → Save search results as target → Derive a natural-language questions as inputs

For each benchmark item,

• input: the natural-language question
• target: the desired search results (from the intermediary PDS query)

Thus, the job of the agent run on the benchmark is to use the NLQ to build a PDS API query (without knowledge of the original one used in the data engine) and get the target search results.

Conditions

Results

Diagnostic analyses

How multistep search works

A request such as:

Find PDS collections from the New Horizons Kuiper Belt Extended Mission 1, collected by the Radio Science Experiment, targeting Arrokoth.

requires the agent to resolve several ontology objects before retrieving data:

1. Resolve the KEM1 investigation identifier.
2. Resolve the New Horizons REX instrument identifier.
3. Resolve Arrokoth and its PDS target identifier.
4. Search collections using the compatible investigation, instrument, and target constraints.
5. Return the complete normalized collection-identifier set.

This mirrors the work a user would otherwise perform manually when discovering PDS4 context identifiers and exact Registry filters.

Benchmark and reproducibility

Readers interested in the benchmark can start with the research/data guide. The underlying pds_research data engine samples compatible PDS entities, builds and validates typed query plans, executes live API requests, and writes the resulting questions and target identifier sets.

Recompute identifier-set metrics with:

pds-research score \
  research/data/main-300-live.jsonl \
  research/results/codex-live-main-300-gpt-5.6-luna.jsonl

See DEVELOPMENT.md for environment setup, server execution, MCP client configuration, and test commands.

MCP server

The FastMCP server in src/pds_mcp_server.py exposes live PDS Registry operations for:

• Investigation, target, instrument-host, and instrument search.
• Context-product traversal.
• Collection search by investigation, target, instrument, and host.
• Detailed product retrieval by PDS identifier.

The server can be used from Claude Desktop, Cursor, Codex, or another MCP-compatible host. Setup instructions are in DEVELOPMENT.md.

Example research queries

• Find Apollo 17 collections produced by the Lunar Surface Experiments Package Heat Flow Experiment and targeting the Moon.
• Find InSight collections produced by the Auxiliary Payload Sensor Subsystem temperature and wind sensor and targeting Mars.
• Find Cassini collections produced by the Imaging Science Subsystem Wide Angle camera and hosted on the Cassini Orbiter.
• Find New Horizons KEM1 radio-science collections targeting Arrokoth.

These examples are drawn from the released benchmark; the complete set is available in main-300-live.jsonl.

License

Code is released under the MIT License.

Support

• PDS Registry API: contact pds-operator@jpl.nasa.gov or open an issue in the PDS API repository.
• This server and study: open an issue in this repository.