Saw a viral post on Twitter comparing "Prashant" and "Croissant" and it got me thinking - how exactly does search work? What logic is behind it?
Luckily, I follow Arpit who had already made a video breaking this down (I've linked that below). After watching it, I figured: why not try building something similar over the weekend? Surprisingly, it only took 2 hours because I already had a basic understanding of search algorithms.
To really get into this, I had to go deeper — here's what I found:
Lexical Search
Lexical search is your classic, text-based search. It matches documents based on exact word matches.
How it works:
- Words are put into an inverted index — a data structure mapping words to the documents they appear in.
word 1 -> [doc1, doc2, doc3]
word 2 -> [doc1, doc4, doc5]
Real-world example:
"optimization" -> ['git.pdf','codeofconduct.md','readme.md']
"CICD" -> ['azurefundamentals.pdf','readme.md']
Limitations:
- Doesn't understand meaning or sound of words.
- One-to-one word mapping only:
"PullRequest" != "PullRaquest"
Common issues:
- Spelling variations:
Hrushi vs Hrishi vs Rishi vs Rushi - Typos/misspellings:
hrushikesh vs hrushkesh vs hrushiksh - Synonyms:
car vs vehicle,automobile vs SUV - Abbreviations:
AI vs Artificial Intelligence
Fuzzy Search
Fuzzy search handles typos or small spelling differences. It matches terms even when they aren't exact.
Example:
Query: Prashant
Result matches (within edit distance <= 4):
- merchant
- elephant
- present
- variant
Edit Distance is how many changes you need to convert one word to another.
Phonetic Search
Phonetic search matches based on how a word sounds, not how it's spelled.
It uses algorithms like:
- Soundex
- Metaphone
- NYSIIS
How it works:
word ------------> phonetic key
"coffee" --------> KF (metaphone)
Example:
mikaela -> MKL (Soundex)
micaela -> MKL (Soundex)
kristen -> KRSTN (Metaphone)
cristen -> KRSTN (Metaphone)
Strengths:
- Great for matching spelling variations that sound alike.
- Fast & doesn't need training data.
- Lookup is indexed and efficient (O(n) encoding).
Concerns:
- Can lead to false positives — different words that sound similar can match.
Example:
"file" => Soundex: F400 | Metaphone: FL
"phile" => Soundex: P400 | Metaphone: FL
Semantic Search
Semantic search goes beyond words — it captures the meaning using NLP and embeddings.
How it works:
- Each word/sentence is converted into a vector.
"laptop" -> [0.25, 0.742, 0.1232, 0.5453]
- Perform k-NN search to find the closest vectors (similar meaning).
Pre-reqs:
- Trained language model (like BERT, Word2Vec).
- Corpus for training (e.g., news articles).
Strengths:
- Understands context.
- Matches concepts not present in the corpus.
For example, your data doesn't have "laptop" but has "notebook" — it'll still match!
Concerns:
- Needs training and resources.
- Vector comparisons are resource-intensive.
- Approximation (ANN) may be needed for large datasets.
Back to "Prashant vs Croissant"
None of the phonetic or semantic algorithms would match them — because:
prashant => Soundex: P625 | Metaphone: PRXNT | NYSIIS: PRASAD
croissant => Soundex: C625 | Metaphone: KRSNT | NYSIIS: CRASAD
They're phonetically and semantically unrelated.
So how did that tweet match them?
Solution: Manual Synonym Mapping
Define a custom synonym map.
1. When query is "prashant"
2. Expand to "prashant OR croissant"
3. Fire the expanded query
4. Elasticsearch returns docs matching either
Simple and effective!
ElasticSearch Demo
I've created a full demo showing different types of search using AWS ElasticSearch. Check it out on GitHub.
References: