There are big opportunities precisely because the field seems dead.
(1) The first big story is the dominance of Google. With an advertising-centered model, Google has a reason to degrade result quality. If you get trained to scroll down to find the real results and you found them good, you might avoid touching any of the ads (hard to do because they cover so much of the screen.)
(2) The web is 95% Javascript and 95% Spam -- getting useful results at all requires fairly strict 'censorship' and vast resources if you want to compete on Google's ground. No serious competitor will come in with a different model, nothing will change unless you have a search engine that YOU pay for and not the advertisers.
(3) "Desktop search" is discredited in most peoples minds. Your OS might have added it as a feature back in 1995, but you've kept it turned off because it slows down your computer and never finds what you are looking for. Result quality is an issue, but the #1 perception here is that the indexing process harms the user experience. In the era of multicore, NVMe, etc. can this be changed?
(4) "Website search" is also discredited. Product search commonly works, but search on most web sites is so bad that people are trained to just search on Google. Thus you have very few chances to change people's minds.
(5) There is a big literature (the TREC conference) but there is something profoundly depressing about it. It was one of the first big competitions, but unlike the SAT Solver competition or Imagenet it was not associated with a rapid improvement of technology but rather a painful slog through the mud. If you start reading it at the beginning or in the middle somewhere you will find that 20 or so things that you thought were sure bets to improve relevance don't work. If you read the cliff note's to the first 10 years written by the organizer, you find out that there was an interesting discovery made 5 years in...
(6) The BM25 ranking function which has two tunable parameters. BM25 was a huge advance because it can be tuned to comparable rank documents that are highly variable in size. BM25 is built into Elastic Search, but nobody will give you any advise how to tune those parameters...
(7) Because they don't follow the relevance evaluation protocol in TREC; this is badly flawed, but the data exists, and going from naive tfidf to tuned up BM25 or information theoretic approach (also implemented in Elasticsearch) will put up better numbers AND seem more relevant to end users.
(8) An open-source project to do that evaluation on Lucene got started but never made a project; I have talked with Enterprise Search vendors who were very aware of points 5-7 but did not tune up their search because it was easier to sell customers on having hundreds of "connectors".
(9) The mainstream of TREC (it has broken into many flavors) and IR research has been getting high recall at low precision. Maybe that's because when Gerard Salton was messing around with punched cards at Cornell, 70 abstracts was a lot of documents. Patent searchers and paralegals are interested in deep recall, other people aren't.
(10) A major flaw in the mainstream TREC approach is that they are trying to tune up the wrong function: the ideal relevance score is a probability estimator of how likely the document is to be relevant.
(11) Google and Bing have made noises about personalized search but they don't really do it. They are both stuck at 70% relevance for the first result because of their limits in inferring user intent. The real relevance function has the user's context as an input variable, but sampling by that thins the data points to where it can't be approached as a "big data" problem. "Personalization" works for advertisers who don't know your real intent but are willing to pay for a 5% chance you may click, but not for you where you will feel misunderstood (primed to get irrationally angry) 95% of the time.