Solving anagrams for long strings or sentences is a hard problem. A naive implementation, which generates all substrings from a given input for looking them up in a hash table does not scale.
A more efficient solver can be implemented using recursive backtracking and a trie or, for more space efficiency, a directed acyclic word graph (DAWG). This answer on stackoverflow provides an helpful description of the recursive backtracking algorithm.
Since recursion is natural for Haskell, the implementation is fairly straight forward. One caveat, however, is that two accumulators are necessary. One for the current word which is partially built up from stepping through the nodes of the graph, and another one for all resulting valid words.
There are plenty of comments in the source and I've tried to write the code as concise and readable as possible. If there are any questions, do not hesitate to ask me.
Some simple tests are also available as compile time option. Load the module
with the ANAGRAM_SOLVER_TEST_CASES macro set and call runTests.
$ cabal repl
# [..]
*Anagram.Solver> :set -DANAGRAM_SOLVER_TEST_CASES
*Anagram.Solver> :load *Anagram.Solver
# [..]
*Anagram.Solver> :t runTests
runTests :: IO ()