Corpus & licences

Generation is retrieval-grounded: a request is answered against relevant passages from a curated GMAT corpus, so the model writes against real syntax rather than from memory. The corpus and a prebuilt index ship with the package, so you never need a GMAT install to generate.

What's in it

The corpus has five tiers. Four are extracted from a GMAT R2026a distribution; one is first-party.

Tier	Source	Chunks
help	GMAT help documentation (reference, tutorials, how-to and chapter pages)	1427
sample	the stock sample mission scripts	586
catalogue	the resource/command field catalogue (types, fields, enums)	102
domain-note	hand-written modeling-semantics and gotcha notes	20
gmf	the sample GmatFunction (.gmf) files	9

That is 2144 chunks in total. Each chunk records its tier and origin, so a result's retrieval trace attributes every grounding passage back to a help page, a sample-script section, a GmatFunction, a catalogue type, or a domain note.

The User's Guide PDF (the same source as the help HTML), internal spec PDFs, and Python-API notes are deliberately excluded.

Ship vs. rebuild

Maintainers extract the chunked corpus text at build time and the package ships both the text (corpus.jsonl) and a prebuilt FAISS index (index.faiss) for the default embedder, BAAI/bge-small-en-v1.5 (384-dimensional). At runtime the index loads directly — no GMAT, no network. The index is rebuilt on first use only as a fallback, when you supply a non-default embedder or the corpus has changed; a content hash in the manifest keys that fallback cache so a stale rebuild is invalidated automatically. Shipping the prebuilt index also makes retrieval deterministic, which the evaluation suite relies on.

Licences

Redistribution is the load-bearing question for any GMAT-derived corpus, and it is clean:

• GMAT-derived tiers (help, sample, catalogue, gmf) come from GMAT, which NASA's Goddard Space Flight Center distributes under the Apache License 2.0. That licence explicitly covers documentation source and permits redistribution with attribution. The attribution ships in THIRD-PARTY-NOTICES alongside the project's MIT LICENSE. Only derived text and embeddings are redistributed — no GMAT source code or binaries.
• The domain-note tier is original first-party content authored for this project, covered by the project's MIT licence.
• The embedding model, BAAI/bge-small-en-v1.5, is MIT-licensed. The model itself is not redistributed; only the embeddings it produced for the corpus are.

Apache-2.0 content is one-way compatible inside an MIT project given attribution, so the whole corpus redistributes without a contractual block.