Past couple of days I had this idea in mind of implementing a autocomplete that uses fuzzy matching. For fuzzy matching of a partial string with all the strings in dictionary it uses the Levenshtein distance. Now finding the Levenshtein distance of given string with each string in dictionary is very inefficient. So I used the idea described here. Converted it to java and modified it to fit the needs.
It is available in 2 modes. First is the Http server mode in which you run the standalone server and call it directly from your ajax script.It returns data in json format with decreasing scores. For this you need to configure it by providing the file name containing the words for suggestion. To use it run the server and then use this url to get the results http://server-ip/autoc?word=word_for_autocomplete&tf=max_typos_allowed
Specify the word_for_autocomplete and the max_typos_allowed parameter.
The second is you can embed it directly in your application by first creating trie and then calling the getResult() method in CHandler class. See the source code for better understanding. A sample main method is given which uses the second method.
You can find the source code for the project here.
If you like the idea and want to improve it or port it to other languages fork me github.