PC: Created page with "How I used embeddings, similarity search, and retrieval pipelines to build context-aware AI that actually remembers things 500px Every time someone says “AI models forget context”, I grin. Because that’s only true if you haven’t yet played with vector databases. In my experience, building context-aware AI isn’t just about prompt engineering — it’s about memory management. In this article, I..."

2025-11-28T23:14:06Z

Created page with "How I used embeddings, similarity search, and retrieval pipelines to build context-aware AI that actually remembers things 500px Every time someone says “AI models forget context”, I grin. Because that’s only true if you haven’t yet played with vector databases. In my experience, building context-aware AI isn’t just about prompt engineering — it’s about memory management. In this article, I..."

New page

How I used embeddings, similarity search, and retrieval pipelines to build context-aware AI that actually remembers things

[[file:Building_Smarter_AI_Systems_with_Vector_Databases.jpg|500px]]

Every time someone says “AI models forget context”, I grin. Because that’s only true if you haven’t yet played with vector databases. In my experience, building context-aware AI isn’t just about prompt engineering — it’s about memory management.
In this article, I’ll walk you through how I built a production-grade retrieval-augmented generation (RAG) pipeline using vector embeddings, similarity search, and OpenAI’s API.
 If you’ve ever wanted your AI system to remember documents, conversations, or domain-specific knowledge — this one’s for you.

1. What Are Vector Databases, Really?

At the core of any retrieval-based AI is one simple idea: turn text into numbers that capture meaning, and then compare those numbers to find related content.

Each document, paragraph, or even sentence can be transformed into a vector (a list of floating-point numbers) using an embedding model.
<pre>
from openai import OpenAI
import numpy as np

client = OpenAI(api_key="your_sk")

text = "Python is a high-level programming language."

# Get embedding vector
response = client.embeddings.create(
model="text-embedding-3-small",
input=text
)

vector = np.array(response.data[0].embedding)
print(len(vector), "dimensions")
</pre>

Most modern embeddings (like OpenAI’s text-embedding-3-small) have 1,536 dimensions—enough to capture complex semantics.
When you store these vectors in a specialized database (like Pinecone, Weaviate, or FAISS), you can query them for semantic similarity. That means you’re not just matching keywords, but ideas.

2. Setting Up a Vector Database with FAISS

For experimentation, I love using FAISS, a library developed by Facebook AI Research. It’s fast, local, and perfect for prototypes.
<pre>
import faiss
import numpy as np

# Let's create some fake embeddings
data = np.random.random((100, 1536)).astype('float32')

# Build FAISS index
index = faiss.IndexFlatL2(1536)
index.add(data)

# Query vector
query = np.random.random((1, 1536)).astype('float32')

# Find 5 closest vectors
distances, indices = index.search(query, 5)

print(indices)
</pre>

Each query compares the distance between vectors. Smaller distance = higher similarity. That’s the foundation of semantic search.

3. Chunking Documents for Semantic Memory

Before we can store anything, we need to chunk our documents. This is one of those underrated tasks that can make or break your retrieval accuracy.
Chunking is about breaking large texts into meaningful sections — big enough to hold context, small enough to stay precise.
<pre>
def chunk_text(text, chunk_size=500, overlap=50):
chunks = []
start = 0
while start < len(text):
end = start + chunk_size
chunk = text[start:end]
chunks.append(chunk)
start += chunk_size - overlap
return chunks

document = "AI systems are designed to simulate human intelligence..." * 10
chunks = chunk_text(document)
print(len(chunks))
</pre>

Overlapping chunks ensure continuity of meaning — critical for maintaining context when the model retrieves relevant sections.

4. Creating and Storing Embeddings for Each Chunk

Once we have our chunks, we generate embeddings for each and store them in our vector database.
<pre>
from openai import OpenAI
import pandas as pd

client = OpenAI(api_key="your_sk")

def create_embeddings(chunks):
data = []
for chunk in chunks:
embedding = client.embeddings.create(
model="text-embedding-3-small",
input=chunk
).data[0].embedding
data.append(embedding)
return np.array(data)

embeddings = create_embeddings(chunks)
df = pd.DataFrame({"chunk": chunks, "embedding": list(embeddings)})
df.head()
</pre>

We’ll use this DataFrame to connect chunks with their corresponding embeddings. In production, this data would go straight into Pinecone or Weaviate.

5. Performing Semantic Search Over Stored Knowledge

Now comes the fun part: querying our AI’s memory.
<pre>
query = "How can AI systems retain long-term knowledge?"
query_embedding = client.embeddings.create(
model="text-embedding-3-small",
input=query
).data[0].embedding

# Compute cosine similarity
from numpy import dot
from numpy.linalg import norm

def cosine_similarity(a, b):
return dot(a, b) / (norm(a) * norm(b))

df["similarity"] = df["embedding"].apply(lambda x: cosine_similarity(query_embedding, x))
top_chunks = df.sort_values("similarity", ascending=False).head(3)
print(top_chunks["chunk"].values)
</pre>

Now the AI retrieves the most semantically relevant text instead of just keyword matches — essentially “remembering” what’s important.

6. Integrating Retrieval with GPT for Contextual Answers

Here’s where it all comes together. We combine retrieved context with the user query, and pass it to GPT for a grounded, accurate answer.
<pre>
context = "\n\n".join(top_chunks["chunk"].values)
prompt = f"""
You are an expert AI assistant.
Use the following context to answer the question accurately.

Context:
{context}

Question: {query}
"""

response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": "You are a knowledgeable assistant."},
{"role": "user", "content": prompt}
]
)

print(response.choices[0].message.content)
</pre>

This is the backbone of RAG (Retrieval-Augmented Generation) — a pattern that underpins many of today’s intelligent chatbots, knowledge assistants, and internal AI tools.

7. Building a Simple Gradio Interface

Once you’ve got retrieval and generation nailed, the next step is a user interface. Gradio makes this ridiculously easy.
<pre>
import gradio as gr

def answer_question(query):
query_embedding = client.embeddings.create(
model="text-embedding-3-small",
input=query
).data[0].embedding
df["similarity"] = df["embedding"].apply(lambda x: cosine_similarity(query_embedding, x))
context = "\n\n".join(df.sort_values("similarity", ascending=False).head(3)["chunk"].values)
prompt = f"Answer this using context:\n\n{context}\n\nQuestion: {query}"

response = client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": prompt}]
)
return response.choices[0].message.content

gr.Interface(fn=answer_question, inputs="text", outputs="text", title="AI Memory Assistant").launch()
</pre>

Now you have a fully interactive, memory-aware chatbot powered by your own knowledge base.

8. Scaling It Up: Vector Databases in Production
When you outgrow FAISS, move to managed services like:
* Pinecone (super fast and serverless)
* Weaviate (supports hybrid search and metadata filters)
* Milvus (open-source powerhouse)
* Chroma (great for local prototypes)
With these, you can index millions of embeddings, support metadata-based search, and plug in your RAG system to handle real workloads.
Quote: “A good memory doesn’t just recall facts — it recalls relevance.” That’s exactly what vector databases give your AI.

9. Lessons Learned from Building Memory-Driven AI
* Chunking strategy is everything. Overlap matters more than you think.
* Embedding quality determines recall accuracy. Garbage in, garbage out.
* Context window limits are real. Be strategic with what you send to the model.
* Store metadata. Source, author, date — it’ll save you later.
* Iterate fast. The beauty of Python is that you can prototype entire pipelines in hours.

Final Thoughts
After integrating vector search into my AI systems, my models started behaving like they actually understood history. They stopped hallucinating and began responding with grounded, contextual answers.
This isn’t just a trick — it’s the evolution of intelligent systems.
If you want to build AI that feels alive, give it memory. And as you just saw, all it takes is a few Python scripts, some embeddings, and a vector database.

Read the full article here: https://medium.com/@abromohsin504/building-smarter-ai-systems-with-vector-databases-a2a9fe113c33

Building Smarter AI Systems with Vector Databases - Revision history