Deep Field
Seventy years of NASA research, searchable by just asking. You get an answer with the exact page it came from, and the original diagrams alongside it. I built it partly because it's genuinely useful, and partly because it's the honest way to prove a private system works. Point it at documents everyone can already see.
- 108,448
- reports
- 4.5M
- passages
- 586K
- figures
What the problem was
NASA has published over a hundred thousand technical reports and they're all free to read. The catch is that reading them is the only thing you can do. There's no way to ask a question across all of them.
The documents are also genuinely difficult. The older ones are typewritten pages that were scanned decades later. A lot of the real content is in diagrams and equations, which normal search simply can't see.
I also had a second reason for building it. I sell private AI systems that run inside a company's own network, and the whole objection is that the customer can't hand me their documents to test it on. So I pointed the same system at documents everyone can already see, and left it open for anyone to try.
How I built it
Reading the pages properly
Everything goes through IntraParse, the parsing engine I built. It reads each page the way a person would, noticing where the tables are, where the diagrams are, which bits are equations, then describes each one so it can be searched.
That's the part most tools skip. If you only pull the text out, you lose the thrust curve, the schematic, and the equation, which in an engineering report is usually the point.
Making the pictures searchable
586,000 diagrams and charts are indexed as things in their own right, not as captions attached to paragraphs. Ask about a thrust curve and you get the actual thrust curve back.
It works the other way round too. You can paste a screenshot of a plot or an equation and it'll go and find related material across the whole archive.
Searching like a person, not a search box
A real research question rarely has one answer sitting in one place. So instead of doing a single lookup, it behaves more like a researcher: it runs several searches, reads what comes back, follows whatever looks promising, and only then writes an answer.
Every claim in that answer links back to the exact page it came from, because anyone who knows the subject will check.
How it turned out
- Live and open, with no signup between you and a real question.
- 108,448 reports, 4.5 million passages and 586,000 figures, all searchable by meaning.
- Every answer points back at a real page in a real public report.