Vector Database Simplified
Very Easy to Understand
This is a series where we will discuss important concepts of AI for Product Managers → Subscribe
Imagine you’re a Product Manager at “KiranaKart,” a booming online grocery and essentials platform in India.
You’re digging through your analytics and a particular search query hits you: “healthy breakfast for kids.”
The search volume is high, but the abandonment rate from this query is through the roof. Users search, see the results, and just… leave.
This isn’t just a missed sale. It’s a signal: your platform isn’t meeting a clear user need. Traditional search is failing you, and it’s actively hurting your metrics.
Priyanka, a concerned parent, types “healthy breakfast for kids” into KiranaKart.
What Priyanka means: She’s looking for breakfast options that are nutritious, suitable for children, probably quick to prepare, and perhaps low in sugar. She might be thinking of things like ragi porridge, oats, whole-wheat bread for toast, maybe some fruit, or even poha if the ingredients are right. She’s not necessarily looking for products that literally have “healthy,” “breakfast,” “for,” and “kids” in their names in that exact order.
What traditional, keyword-based search on KiranaKart hears: “Show me items with the exact words ‘healthy,’ ‘breakfast,’ ‘for,’ ‘kids’ in their title or description.”
The Result? A Mess.
Maybe a niche “Kids’ Healthy Breakfast Cereal” pops up (if you’re lucky and it’s tagged perfectly).
But what about a simple packet of “Organic Ragi Flour”? Or “ Oats”? Or “Multigrain Atta”? These are fantastic healthy breakfast options for kids, but Priyanka might never see them because the keywords don’t perfectly align.
She might even see adult “protein bars” if they happen to mention “healthy breakfast” in a generic description, leading to confusion.
Priyanka gets frustrated. KiranaKart seems to have nothing relevant, even though your warehouse is stocked with suitable items. She bounces, maybe to a competitor or, worse, gives up on buying these items online altogether.
This is a classic product problem: your search functionality doesn’t understand user intent or context, leading to poor discovery and high drop-offs.
This is where vector databases become your superpower.
Beyond Keywords: Understanding Meaning with Vector Embeddings
Forget simple keyword matching for a moment. What if you could represent the meaning or essence of every product on KiranaKart, and Priyanka’s search query, as a series of numbers?
These series of numbers are called vector embeddings. [If you don;t understand what embedding is, I would advise you to look at our previous posts where we explained Embeddings]
Think of it like this: sophisticated AI models analyse the text (product descriptions, user queries, even images) and distill its semantic meaning into a list of numbers (a vector).
Priyanka’s query “healthy breakfast for kids” might become a vector like
[0.6, -0.2, 0.9, ..., 0.4]
A product like “Organic Ragi Flour” (often used for healthy kids’ porridge in India) might have a vector:
[0.55, -0.15, 0.85, ..., 0.35]
A “Spicy Adult Snack Mix” would have a very different vector:
[-0.8, 0.5, -0.1, ..., -0.7]
The crucial part: Items with similar meanings will have mathematically similar vectors. They’ll be “close” to each other in this abstract “meaning space.” So, even though “ragi flour” doesn’t contain the word “breakfast” or “kids” in its name, its vector embedding will be close to the vector for “healthy breakfast for kids” because the AI understands the contextual relationship (ragi is a common healthy breakfast for kids in India).
Okay, so you’ve turned all your product descriptions and user queries into these powerful vector embeddings. Now what? You have millions of products. How do you instantly find the product vectors that are “closest” to Priyanka’s query vector?
Comparing one by one is a non-starter — it’s too slow and computationally expensive.
This is the job of a vector database. It’s a specialized database purpose-built to:
Store: Efficiently house vast quantities of these vector embeddings.
Index: Organize these vectors in a way that makes searching super fast (using clever algorithms like HNSW, Faiss, etc. — you don’t need to know the algorithm names, just that they’re smart!).
Retrieve: Perform similarity searches (or “Approximate Nearest Neighbor” — ANN searches) at lightning speed. Given Priyanka’s query vector, it can instantly find the product vectors that are most similar in that “meaning space.”
Back to Priyanka and KiranaKart:
Priyanka types “healthy breakfast for kids.”
KiranaKart’s system (now powered by a vector database) converts this query into its vector embedding.
This query vector is sent to the vector database.
The vector database instantly searches its indexed product vectors and returns the ones that are mathematically closest — e.g., “Organic Ragi Flour,” “Steel Cut Oats,” “Whole-Wheat Bread,” maybe even some specific idli/dosa batters known for being healthy.
KiranaKart displays these truly relevant products to Priyanka.
The Outcome: Priyanka sees a list of items that genuinely match her intent. She finds what she needs, adds items to her cart, and completes her purchase. Your abandonment rate for this query plummets. Your conversion rate climbs.
This isn’t just a technical novelty; it’s a direct solution to core PM challenges:
Solve “Bad Search”: Move from frustrating keyword misses to intuitive, intent-driven discovery. This directly impacts user satisfaction and task completion rates.
Improve Product Recommendations: “Customers who bought this (vector A) also bought these items (vectors B, C, D, which are close to A).” Think “Frequently Bought Together” but on semantic steroids, even for items without direct co-purchase history. Imagine recommending jaggery with ragi flour, even if users haven’t explicitly bought them together yet, because their “health food” and “Indian breakfast” vectors are similar.
Visual Search: “Find kurtas similar to this photo.” (Image vectors!)
Personalized Discovery: “Show me recipes I might like based on ingredients I usually buy.”
Users in India search with incredible diversity — multiple languages, Hinglish, regional terms, and varied intent. A system that understands meaning beyond exact keywords is no longer a luxury; it’s essential for capturing this diverse market.
If your product relies on users finding things, and you’re seeing drop-offs because they can’t, it’s time to look beyond traditional databases. Vector databases offer a powerful way to connect users with what they truly want, even when they don’t say it perfectly. Stop leaving conversions on the table due to a search bar that doesn’t understand context.
Resource
To understand such Tech Concepts and AI/ML Case Studies.
About Me
Hey, I’m Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking.
For more, check out my Live cohort course, PM Interview Mastery Course, Cracking Strategy, and other Resources