APOLLO 11 / AN EXPERIMENT IN GROUNDED ANSWERS
Ask the
mission archive.
A question is only as useful as the evidence behind its answer. Explore Apollo 11 through mission reports, transcripts and technical records — with the sources in reach.
What would you like to know?
Explore the landing, crew decisions and mission science through the original documents.
ENGINEERING NOTES
Measured, with limits.
This is a work in progress. The cloud backend combines keyword and vector retrieval in Qdrant and streams GPT-5.6 Luna answers with links to the original documents. No dedicated reranker is used.
The archive contains 265 canonical PDFs and 931 standalone images. Search currently covers eligible extracted text; standalone images have not been embedded and full-corpus OCR has not been performed.
| Page hit | BM25 | Vector | Hybrid |
|---|---|---|---|
| Top 5 | 45.8% | 31.2% | 45.8% |
| Top 10 | 60.4% | 41.7% | 62.5% |
| Top 20 | 68.8% | 60.4% | 83.3% |
Supporting-page hits are location proxies, not answer accuracy. These are the original local results; the first Qdrant comparison reproduced these page-hit rates. The split has some topic overlap with held-out questions.
What the answer review found
Ten model-assisted reviews produced 5 passes, 2 partial answers and 3 failures. Failures included confusing computed and executed velocity changes, choosing a later docking time, and accepting a fictional second-landing premise.
Those failures remain in the original test record. More precise context, numeric-evidence checks and stronger abstention instructions still need a fresh claim-level evaluation before this assistant can be called reliable.