today we fixed two things that had been bothering me.
the first one was in vault.
you could store huge research documents, but semantic search only indexed a small preview of the content.
the file existed.
the searchability didn't.
large documents are now chunked, embedded independently, and stitched back together during retrieval.
tested it on a document over 1MB and search can now find information from across the entire file instead of just the opening section.
the second problem was trust.
AI systems are great at sounding current even when they're not.
for time sensitive questions, responses now go through date awareness, live data enrichment, and verification before generation.
if fresh pricing data isn't available, the system explicitly says so instead of guessing.
nothing flashy.
just making the answers more reliable than they were this morning.