In our previous blog, we learned about Semantic Search.
Semantic Search helps AI find information based on meaning, instead of looking only for exact keywords.
But now an important question comes:
How does AI decide which information is most similar to our question?
The answer is Similarity Search.
Similarity Search is one of the important techniques used behind semantic search and many RAG systems.
In this blog, let's understand it with a simple real-world example.
What Is Similarity Search?
Similarity Search is the process of finding the most similar information from a collection of stored information.
In AI applications, similarity is often calculated by comparing embeddings.
Simple definition:
Similarity Search finds the information that is closest to a given query based on its representation.
Let's Take a Real-World Example
Imagine a college has a student handbook.
The handbook contains different information:
- Attendance Rules
- Exam Rules
- Hostel Rules
- Library Rules
- Fee Rules
These sections are converted into smaller chunks and then into embeddings.
Now a student asks:
"How much attendance do I need for my final exam?"
The system needs to find which information is closest to this question.
It may compare the question with different stored chunks:
User Question
↓
"How much attendance do I need
for my final exam?"
↓
Compare with stored information
↓
Attendance Rule → Very Relevant
Exam Rule → Relevant
Hostel Rule → Not Relevant
Library Rule → Not Relevant
Fee Rule → Not Relevant
The attendance rule is the best match.
This process of finding the clo
sest matches is called Similarity Search.
But How Does AI Know What Is Similar?
This is where Embeddings come in.
We already learned that embeddings convert text into numerical representations.
For example:
"How much attendance do I need?"
↓
Embedding
↓
[0.21, 0.67, 0.45, ...]
The attendance rule also has an embedding:
"Students must maintain 75% attendance..."
↓
Embedding
↓
[0.22, 0.65, 0.48, ...]
The system compares these representations.
If they are close according to the chosen similarity method, the information is considered more similar.
What Is a Similarity Score?
When two embeddings are compared, the system can produce a similarity score.
The score helps the system understand how closely two items match according to the selected similarity method.
For example:
Question ↔ Attendance Rule
↓
High Similarity
Question ↔ Hostel Rule
↓
Low Similarity
So the system can select the most relevant information.
Remember:
Higher similarity generally means the information is more closely related.
The exact score and its range depend on the similarity method being used.
Similarity Does Not Mean Exact Words
This is very important.
Consider these two sentences:
Sentence 1:
"How much attendance is required for the final exam?"
Sentence 2:
"Students need a minimum of 75% attendance to appear for the examination."
The words are different.
But the meaning is closely related.
Embeddings help represent that meaning, and similarity search can help identify that relationship.
That's why AI systems can find relevant information even when the user's words don't exactly match the document.
One Common Method: Cosine Similarity
One commonly used method for comparing embeddings is Cosine Similarity.
You don't need to learn the mathematics right now.
The basic idea is:
It compares the direction of two vectors.
Think about two arrows.
If they point in a similar direction, they are more similar.
Vector A ↗
/
↗ Vector B
If they point in very different directions, they are less similar.
Vector A ↗
↘ Vector B
So cosine similarity is one way an AI system can measure how similar two embeddings are.
Where Does Similarity Search Fit in RAG?
Now let's connect this with the RAG concepts we have already learned.
A simplified RAG flow looks like this:
Similarity Search helps answer this question:
"Which stored chunks are closest to the user's question?"
What Does the Vector Database Do?
A Vector Database stores embeddings and allows the system to search through them efficiently.
For example:
Vector Database
Attendance Chunk → Embedding
Exam Chunk → Embedding
Hostel Chunk → Embedding
Library Chunk → Embedding
Fee Chunk → Embedding
When the student asks a question:
User Question
↓
Query Embedding
↓
Similarity Search
↓
Vector Database
↓
Most Relevant Chunks
The relevant chunks can then be passed to the LLM.
Semantic Search vs Similarity Search
This is where many beginners get confused.
They are closely connected, but we can explain them differently.
Semantic Search
Semantic Search is the search approach.
Its goal is:
Find information based on meaning.
Similarity Search
Similarity Search is the process of comparing representations to find the closest matches.
Its goal is:
Find which stored items are most similar to the query.
So, in many AI systems:
Semantic Search
↓
Search based on meaning
↓
Embeddings
↓
Similarity Search
↓
Relevant Information
Important: These terms can overlap in real-world AI discussions. We are separating them here only to make the concepts easier for beginners to understand.
A Simple Everyday Example
Imagine you are searching for a product online.
You type:
"Comfortable shoes for daily walking"
The search system may find products described as:
"Lightweight sneakers for everyday walking."
The words are not exactly the same.
But the meaning is closely related.
The system can use embeddings and similarity-based techniques to find relevant results.
This is the basic idea behind many modern AI-powered search systems.
Similarity Search vs Keyword Search
Let's compare them.
Keyword Search
User:
"attendance"
The system mainly looks for matching words such as:
attendance
Similarity-Based Search
User:
"How much attendance do I need for my exam?"
The system can find information such as:
"Students must maintain a minimum of 75% attendance to appear for the final examination."
Even though the wording is different, the meaning is closely related.
Why Is Similarity Search Important?
Similarity Search is useful because people don't always ask questions using the exact words present in a document.
For example:
Document:
"Minimum attendance requirement is 75%."
User:
"Can I write my final exam if my attendance is 70%?"
The wording is different.
But the question is related to the attendance requirement.
Similarity-based retrieval can help find the relevant attendance information.
Similarity Search in One Simple Flow
Let's remember the entire idea with our college example.
Student asks a question
↓
Question is converted into embedding
↓
Compared with stored embeddings
↓
Similarity is calculated
↓
Most relevant information is selected
↓
Relevant chunk is retrieved
↓
LLM uses the information
↓
Final answer
One Important Point
Similarity Search does not generate the final answer.
Its main job is to find relevant information.
The LLM then uses that information to generate a human-readable answer.
So remember:
Similarity Search
↓
Finds relevant information
LLM
↓
Generates the answer
How Similarity Search Connects With What We Learned
So far, our concepts are building one after another:
Embeddings
↓
Vector Database
↓
Semantic Search
↓
Similarity Search
↓
Relevant Information
↓
Retrieval
↓
Context
↓
LLM
↓
Answer
Each concept has a different role.
That's why understanding similarity search helps us understand how RAG retrieves the right information.
Simple Recap
What is Similarity Search?
It finds the most similar or relevant information from stored data.
What does it compare?
Usually, embeddings represented as vectors.
What is a similarity score?
A value used to indicate how closely two representations match according to a chosen similarity method.
What is Cosine Similarity?
One common method for comparing the similarity of vectors.
Does Similarity Search generate an answer?
No.
It finds relevant information. The LLM generates the final answer.
Conclusion
Similarity Search is an important part of many modern AI systems.
It helps the system find information that is closely related to the user's question.
In RAG, similarity search helps identify relevant chunks from a vector database. Those chunks can then be provided to the LLM as context.
The simple idea to remember is:
Similarity Search = Find the closest relevant information.
And the overall process is:
Question
↓
Embedding
↓
Similarity Search
↓
Relevant Information
↓
LLM
↓
Answer
Now we understand how the system finds relevant information.
But another important question comes next:
After finding the relevant information, how does RAG actually retrieve it and give it to the LLM?
That takes us to our next topic:
No comments:
Post a Comment