You probably don't need a dict; just a list of candidates. And you should clarify that you should sort your list by reversed word length and keep another list o

You probably don't need a dict; just a list of candidates. And you should clarify that you should sort your list by reversed word length and keep another list of indexes into each word. EDIT: You shouldn't sort the list beforehand. You should compile the list of matched candidates after walking through the whole string and pick the longest match. (Or just represent the matched candidates in a list of bools). It's possible that your answer only happens to match the last few characters of your haystack, so you must walk the entire length of the haystack. EDIT2: This solution also makes a bunch of assumptions and you can find a better runtime in certain cases. For example, if you have many more candidates than the length of your haystack, it might be worthwhile to generate all O(2^n) subsequences and store them in a hashmap/bloom filter. Then you can do (almost) O(1) lookups for each word. This would also lend well to a distributed compute approach: you could have k machines sharded in this manner: the first machine computes and stores in your hashmap/bloom filter the first 2^n/k subsequences, the second the second 2^n/k, etc. Then you can either 1) in the case of hashmaps, fan out each request to each machine, succeeding if any machine succeeds or 2) merge the bloom filters on a central machine (assuming the same hash functions and modulus constants) and do O(1) lookups. (Copying hashmaps is O(data stored inside); merging bloom filters is O(bloom filter size).)