Of all the data structures I’ve encountered, the prefix tree (also known as a trie) still fascinates me the most because of its simplicity, elegance, and practical applications.
In this tutorial, we’ll implement a trie in Python from scratch. We’ll test that our code works using Python’s
unittest library. Let’s get started!
Overview: What Is a Prefix Tree?
The prefix tree is one of the easiest data structures to understand both visually and in terms of the code required to implement it. But what is a prefix tree, and why might we want to create one?
First, let’s run through a little exercise. Have you ever wondered how search engines like Google are able to quickly auto-fill your search box with suggestions that start with whatever you’ve typed? Take this as an example:
How would you go about implementing this behavior, all other complex considerations aside? The (very) naive approach is to take the text that the user has typed so far—like
app—and check if any words in our database start with that substring, using a linear search. That would maybe work for search engines with a relatively small database. But Google deals with billions of queries, so that would hardly be efficient. It gets even more inefficient the longer the substring becomes.
The efficient answer to this problem is a neat little data structure known as a prefix tree. It’s just a tree (not necessarily a binary tree) that serves a special purpose.
Example Prefix Tree
Let’s say we’re building such a database of words, but we want to be clever with how we do it so that our search doesn’t take forever. Suppose we want to record the words
big in this catalog.
Instead of storing these four words as-is, what we’ll do is create a tree. This tree will consist of branching prefix nodes that, when followed in the right sequence, will lead us to a complete word.
It helps to look at a picture of this and break it down. The corresponding prefix tree for these words would look like this:
Each node in a prefix tree represents a string, with the root node always being the empty string (
''). That string may be a complete word that someone entered into the prefix tree—like
bat—or it may be a prefix that leads to a word, such as the
Each branch coming out of a node represents the addition of a character (in yellow) to the end of that node’s string. For example, if our current node is
app and the word we’re trying to insert is
apple, we’d simply tack on an
l to get the next node:
appl. From there, we’d append an
e to get the node representing our complete word:
Trace the path from the root of the trie to the word
apple. Make a mental note of how each branch coming out of a node is like a key-value pair in a
dict. The key is the character that we add to the end of the current prefix. The corresponding value is the node that the branch leads to!
This branching pattern allows us to reduce our search space to something much more efficient than just a linear search of all words. If a user has entered the substring
appl so far, then we won’t ever consider the branching paths for
ape, let alone all the branching paths that start with
Tracing a path from the root of a trie to a particular node produces either a prefix for a word that we know (e.g., the
apple) or the word itself (e.g.,
apple). This distinction is important because our dictionary doesn’t actually contain the word
app just yet; that node is merely a prefix. You’ll see why this is important later on, but for now, just keep that in mind.
Trie Methods and Operations
A prefix tree has three main operations:
- Inserting a word into the tree.
- Searching for a node that matches a given string exactly.
- Returning a list of all nodes (if any) that begin with a given string prefix.
Let’s put this into the context of a search engine. A company like Google might take an enormous list of words, insert all of them into a trie, and then use the trie to get a list of all words that begin with a certain prefix. That prefix is the partial text the user has entered into the search bar. (Of course, things are a little more complicated than that because of search term popularity and your past searches, but that’s not important for our purposes here.)
Or perhaps you’re creating your own autocomplete widget in, say, React. You’d create the trie beforehand and subscribe to your input field’s
onkeyup event. As the user enters text, you adjust the list of options that you show to them by using your trie.
Building a Prefix Tree in Python
So now that we understand what a prefix tree looks like and how it can be used, how can we represent it in code? You’ll be happy to know that it’s actually really simple, especially if you’re comfortable with recursion.
First, like all trees, a prefix tree is going to consist of nodes. Each node will keep track of three pieces of data. I’ll cover the two important ones here and bring up the third one when we get to it:
- A string. We saw above that each node keeps track of the “prefix” that has accumulated along a specific path from the root. A node may contain a word that was inserted, like
bat, or it may contain a prefix, like
b-. So, each node will certainly need to have a member to store this string. I’ll call that member
- Children. The nodes of a prefix tree, like those of many other trees, branch out, leading to child nodes. Thus, each node of a trie will have zero or more children. The most natural way to represent this relationship is with a map data structure (
dictin Python): Each node maps a character to a child
Let’s build the
class TrieNode: def __init__(self, text = ''): self.text = text self.children = dict()
Pretty simple, right? Again, note that there’s one additional piece of data we’ll need to add here later, but you don’t have to worry about it right now.
Next, the prefix tree itself consists of one or more of these
TrieNodes, so I’ll add another class named
class TrieNode: def __init__(self, text = ''): self.text = text self.children = dict() class PrefixTree: def __init__(self): self.root = TrieNode()
Awesome! Finally, our
PrefixTree needs the following operations, which we’ll fill in step by step:
That last one is useful for testing; it isn’t required.
1. Inserting Words Into a Trie
Let’s consider how we’d build a prefix tree. We’ll always start with a root node that has an empty string as its
text and an empty dictionary as its
children. Then, we want to insert the words we looked at earlier:
big. As a reminder, this is what the trie looks like once we finish inserting all of those words:
Detailed Explanation: How to Build a Prefix Tree
How would we go about building this structure from the ground up? I encourage you to grab a pen and paper to work through this by hand, starting with an empty trie and inserting one word at a time (in no particular order, as that doesn’t change things).
Each time we insert a word into our tree, we start at the root, which is the empty string. We loop over the word that we’ve been given one letter at a time:
l, and finally
In each iteration, if our current
TrieNode doesn’t have an entry in its
children map for that letter, we create such an entry. In other words, we map the current letter (e.g.,
l) to a new child node (
appl). However, if the
letter: TrieNode mapping does exist, we don’t have to do anything. For example, if we first insert
ape and then insert
apple, we won’t have to insert any new nodes for the shared substring,
ap-. However, as we branch out of that common ancestor, we will need to create new nodes.
In either case, at the end of the current iteration, we move on by setting the current node to be the child node: either the one that existed before or the new one that we just created. We repeat this process until we’ve iterated through the entire word.
Recall that each node also has to keep track of the string that’s been generated so far. For example, if we’re inserting the word
apple, then we’ll need to create nodes with the following
apple. In Python, doing this is simply a matter of slicing the string with
i is the current index in the word that we’re inserting. Thus, if we’re inserting
prefix = word[0:3] = 'app'.
The Code: Inserting a Word Into a Trie
Here’s the full* Python implementation of inserting nodes into a trie:
def insert(self, word): current = self.root for i, char in enumerate(word): if char not in current.children: prefix = word[0:i+1] current.children[char] = TrieNode(prefix) current = current.children[char]
*There’s one line missing that I’ll mention in the next section. It’s not going to complicate things at all.
2. Checking if a Word Exists in a Trie
This operation proceeds in a manner similar to insert, except we’re not creating new nodes.
We’ll loop over the word like we did before, one character at a time. At any point, if we can’t find the current letter in the current
TrieNode’s map of children, then the word we’re looking for was never inserted in the first place. Thus, we immediately return
Otherwise, if we reach the last character of the string without having returned
None, then that must mean that the current node is the word we were searching for*. So in that case, we return the current node.
Exercise: Using the diagram from earlier, try to find the word
appreciate, going down the trie one node at a time. What do you return, and at what point in the loop? Answer: Everything is fine for the common
app-substring. But when we look at the letter
r, we find that the current node,
app, doesn't have that letter in its dictionary; it only has the letter
l. Thus, we return
Here’s the code:
def find(self, word): ''' Returns the TrieNode representing the given word if it exists and None otherwise. ''' current = self.root for char in word: if char not in current.children: return None current = current.children[char] return current
*Remember that little caveat I kept bringing up before? This code is mostly correct except for the
return current line at the end. Let’s discuss why this needs to change.
Fixing Our Code: Learning from Mistakes
I kept this hidden from you on purpose so the learning experience would be more memorable. If you figured out what’s wrong with the last line of the code and why, then great! If not, I’ll help you understand.
Let’s say we insert the word
apple into our trie, with nothing else, and now invoke
Our algorithm in its current state won’t return
None as it should. It sees that our trie has a node with the string
app and considers that a match. But notice that we never actually inserted
app into the tree as a word. It only exists in the trie as a prefix node leading up to the inserted word
apple. So clearly, there are two classes of nodes that we must distinguish between: ones that are prefixes, and ones that are “real” words that were inserted.
Fortunately, the fix here is super simple. We need to make three changes:
- Add a boolean flag in the
TrieNodeclass. This flag will be set to
Truefor “word nodes.”
- Set that flag to
Truefor any words we
insertand keep it
Falsefor all others by default.
- Change the last line of our
findfunction to account for this flag.
Let’s modify the
class TrieNode: def __init__(self, text = ''): self.text = text self.children = dict() self.is_word = False # New code
And then the
def insert(self, word): current = self.root for i, char in enumerate(word): if char not in current.children: prefix = word[0:i+1] current.children[char] = TrieNode(prefix) current = current.children[char] current.is_word = True # New code
And finally, to verify that the given word exists in our trie, and that the node we found isn’t just a prefix node, all we need to do is check the current node’s flag. If the flag is
True, then it’s a word, and we found a match; in that case, we return the node. Otherwise, it must be a prefix node, in which case we return
Here’s the final code for
def find(self, word): ''' Returns the TrieNode representing the given word if it exists and None otherwise. ''' current = self.root for char in word: if char not in current.children: return None current = current.children[char] # New code, None returned implicitly if this is False if current.is_word: return current
And that’s it!
Let’s go back to our earlier example of inserting just the word
apple into a trie. What happens if we later try to insert the English word
app (as in “application”) into the trie as a word? Well, the code will go down the trie, reach the
app prefix node, and switch its flag to
True so that we know it’s now officially an inserted word and not just a prefix node:
Technically, it’s both—a prefix leading up to
apple and a word in and of itself. Of course, that’s totally fine! And what matters from a code validity standpoint is that its flag has been set to
That’s another one down, with two more to go. We’re almost done!
3. Return a List of All Nodes Starting with a Given Prefix
The code for finding partial matches in a trie is also really simple. Here’s the algorithm spelled out in English:
- Given a string prefix, first find the prefix node that matches it (if one exists).
- Then, if we found a prefix node, simply iterate through all of its subtrees recursively and add all
is_wordstrings to a list.
The first step is simple—we just did something very similar above for finding exact matches. This time, though, instead of returning
None when a match isn’t found, we’ll return an empty list.
def starts_with(self, prefix): ''' Returns a list of all words beginning with the given prefix, or an empty list if no words begin with that prefix. ''' words = list() current = self.root for char in prefix: if char not in current.children: # Could also just return words since it's empty by default return list() current = current.children[char] # Step 2 will go here
Once we’ve found the prefix node (if it exists), we’ll utilize a helper method for the recursion (step two). It will take two arguments: the current node (
node) and a cumulative list that we’ll add nodes to (
words). If the current node is a word, we add it to the list. Then, for each child of that node, we make a recursive call to the same helper function, passing in the child as the new
Here’s the recursive helper function:
def __child_words_for(self, node, words): ''' Private helper function. Cycles through all children of node recursively, adding them to words if they constitute whole words (as opposed to merely prefixes). ''' if node.is_word: words.append(node.text) for letter in node.children: self.__child_words_for(node.children[letter], words)
And here’s the completed code for
def starts_with(self, prefix): ''' Returns a list of all words beginning with the given prefix, or an empty list if no words begin with that prefix. ''' words = list() current = self.root for char in prefix: if char not in current.children: # Could also just return words since it's empty by default return list() current = current.children[char] # Step 2 self.__child_words_for(current, words) return words
Note: Here, I’m returning a list of strings. In reality, it’s probably better to return a list of
TrieNodes, but it’s much easier to test our code if we return strings.
That’s all we need! Feel free to run through this algorithm by hand to better understand how it works.
4. (Optional) Size of a Prefix Tree
This one depends on your definition of “size.” Is it the number of words that were inserted into the tree, or is it the total number of nodes in the tree? I’ll use the latter definition for consistency with how “size” is defined for trees in general.
I’ll just show the code this time—hopefully you’re comfortable with recursion and tries by now:
def size(self, current = None): ''' Returns the size of this prefix tree, defined as the total number of nodes in the tree. ''' # By default, get the size of the whole trie, starting at the root if not current: current = self.root count = 1 for letter in current.children: count += self.size(current.children[letter]) return count
current has a default value of
None. This allows the user to simply invoke
size() without passing in any arguments for the most common use case: the size of the entire tree. Optionally, if the user wants to check the size of a subtree, they are welcome to do so by passing in the appropriate
Testing Our Code
We can add this to the end of our script to manually test our code:
if __name__ == '__main__': trie = PrefixTree() trie.insert('apple') trie.insert('app') trie.insert('aposematic') trie.insert('appreciate') trie.insert('book') trie.insert('bad') trie.insert('bear') trie.insert('bat') print(trie.starts_with('app'))
But this is tedious and frankly not a very rigorous way of testing our code. So instead, I’ll create a separate file named
test_trie.py and use the Python
unittest library. I’ll import the
PrefixTree class I created in
Below are eight tests covering edge cases. This is where our
size method really comes in handy.
import unittest from trie import PrefixTree class TrieTest(unittest.TestCase): def setUp(self): self.trie = PrefixTree() def test_trie_size(self): self.trie.insert('apple') self.assertEqual(self.trie.size(), 6) def test_prefix_not_found_as_whole_word(self): self.trie.insert('apple') self.trie.insert('appreciate') self.assertEqual(self.trie.find('app'), None) def test_prefix_is_also_whole_word(self): self.trie.insert('apple') self.trie.insert('appreciate') self.trie.insert('app') # 10: [app], [appr], [appre], [apprec], [appreci], [apprecia] # [appreciat], [appreciate], [appl], and [apple] self.assertEqual(self.trie.size(self.trie.find('app')), 10) self.assertEqual(self.trie.find('app').is_word, True) def test_starts_with(self): self.trie.insert('apple') self.trie.insert('appreciate') self.trie.insert('aposematic') self.trie.insert('apoplectic') self.trie.insert('appendix') self.assertEqual(self.trie.starts_with('app'), ['apple', 'appreciate', 'appendix']) def test_starts_with_self(self): self.trie.insert('app') self.assertEqual(self.trie.starts_with('app'), ['app']) def test_bigger_size(self): self.trie.insert('bad') self.trie.insert('bat') self.trie.insert('cat') self.trie.insert('cage') self.assertEqual(self.trie.size(), 10) def test_starts_with_empty_and_no_words(self): self.assertEqual(self.trie.starts_with(''), ) def test_starts_with_empty_returns_all_words(self): self.trie.insert('bad') self.trie.insert('bat') self.trie.insert('cat') self.trie.insert('cage') self.assertEqual(self.trie.starts_with(''), ['bad', 'bat', 'cat', 'cage']) if __name__ == '__main__': unittest.main()
unittest, all method names must begin with
testin order for the library to detect and run them. This is per the unittest docs.
setUp method is something that’s common in unit testing. It’s a method that gets called before each individual test is run. That way, we don’t have to manually initialize a new
PrefixTree in every single method; we can let the framework do that for us.
And here’s the output:
All tests ran correctly—awesome!
And We’re Done!
Prefix trees are rarely ever taught in CS curriculum, with more emphasis placed on associative data structures, linked lists, and trees. But it’s good to know how to make one by hand if you want to create your own autocomplete dropdown widget from scratch, for example, or if you’re looking to add another data structure to your toolkit.
I hope you found this tutorial helpful! You can find the full code for this post in its GitHub repo.