Neat. I do think there is a lot of room to refine the search engine interface. If the dumpster fire that's Google it's pretty clear that natural language search is just not a good approach. It's extremely frustrating when the search engine is second guessing you.
Your "tags" look like search terms, like key n-grams. I'm guessing because when I extract search terms from a document, I get stuff of a similar "vibe". Maybe easier if I show than try to describe, this is from a unit test I have that runs keyword extraction code on a few documents. I think these are from some page about SSH clients
unix_source, msi, ssh_authentication_agent, gitweb, pscp, cryptographic_checksums, windows_source, ssh, puttytel, command-line_secure, telnet, windows_html, most_up-to-date_version, psftp, zip, scp, standalone, version_of_putty, scp_client, dsa, putty, plink, rsa_and_dsa, sftp, binaries, unix, up-to-date_version, checksums, rsa, individual_executables, 64-bit_arm, 64-bit_version, source_archive, .zip_archive, 32-bit_arm, latest_release, checksum, standalone_binaries, cryptographic, windows_on_arm, ftp, zipped, versions_of_putty, alternative_binary_files, unix_source_archive, download_putty, sftp_client, latest_released_version, ssh_and_telnet_client, ssh_and_telnet
I've been playing with the idea of doing some sort of Naive Bayes-categorization of the general topic of a web page and using those to offer a broad filtering on my search results. May be a lot of work since it relies on a degree of manual curation, but seems be doable once you have a decent model cooked up. I'd like, when you search for Plato, to be able to offer the alternative of refining the search to the philosopher or the computer system.