2025-09-29 — Search Got Way Smarter About What You Actually Mean
Ever search for "Python error" and get results from 2015? Or ask about stock prices and get yesterday's close? We taught search to understand when time matters. "Latest news" gets fresh data. "World War 2" gets historical context. Simple concept, surprisingly hard to build.
We rebuilt the entire intelligence layer - search now preserves your exact quotes while fixing typos elsewhere, returns answers in clean markdown instead of rigid JSON, and actually knows when "recent" matters.
Highlights
- Search understands when "recent" matters and when you want historical context
- Quoted phrases stay exactly as you typed them while other words get spell-corrected
- Answers come back faster with better formatting and inline citations
- Technical error searches now include your error code in the context for relevant results
What's New
- Time-aware search that knows "latest news" needs today's data while "World War 2" needs historical sources
- Intelligent quote handling - we preserve
"exact phrases"but fix typos in the rest of your query - Clean markdown answers with inline citations instead of JSON blocks
- Context-aware filtering: news queries check last 24 hours, product queries check last year
- Technical troubleshooting searches (error codes, stack traces) get tagged as time-sensitive automatically
- "Retry without quotes" link when quoted searches fail to find exact matches
What's Better
- "npm ENOENT error" searches now return solutions for current Node versions, not ancient Stack Overflow posts
- "Taylor Swift latest" gets today's news, not her 2006 discography
- "French Revolution causes" gets historical analysis, not modern hot takes
- Stock prices, sports scores, and breaking news prioritize fresh data automatically
- Answers format properly with headers, lists, and readable structure
Technical Improvements
- Replaced rigid JSON parsing with free-text intent analysis (faster, more flexible)
- Time filtering system with 4 modes: Immediate/Current/Recent/Historical
- Query context understanding that adapts to your search domain
- Smarter prompt injection defense with query wrapping
- Better error handling that shows you what went wrong instead of hiding failures
What this means for you: Search stops treating every question the same way and starts understanding what kind of answer you actually need.