Module 05: Embeddings and Vector Databases

Duration: 4 hours | Difficulty: Intermediate | Prerequisites: Modules 01–03

Learning Objectives

By the end of this module, you will be able to:

Explain what embeddings are and why they're foundational to modern AI
Generate text embeddings using OpenAI and open-source models
Implement similarity search using cosine similarity and dot product
Set up and query ChromaDB, Pinecone, Weaviate, and Qdrant
Choose the right indexing strategy (HNSW, IVF) for your use case

Embeddings turn text into vectors (lists of floats). Similar meaning → vectors point in similar directions, even when the words differ. Picture a meaning map: “king” and “queen” are neighbors; “king” and “cucumber” are not.

Text A	Text B	Embedding story
“Refund my invoice”	“I need my money back”	Close vectors
“Refund my invoice”	“Deploy Kubernetes”	Far apart

Fun Fact: You can’t read individual dimensions like English—each dimension is a learned feature mash-up.

Why Embeddings Matter

Use case	Role of embeddings
Semantic search	Query vector near doc vectors
RAG	Retrieve chunks by meaning
Clustering / dedupe	Group similar items
Recommendations	“More like this”

Try This! Write three paraphrases of one policy sentence—your goal later is for all three to hit the same chunk in retrieval.

2. Generating Embeddings

OpenAI Embeddings

Hosted models: send strings, get vectors. Great when you don’t want GPU ops on your laptop.

Batch Embeddings

Batching amortizes HTTP overhead—one call, many rows—and usually saves money vs naive loops.

Open-Source Embeddings

sentence-transformers runs locally; good for dev, air-gapped, or high-volume offline indexing.

Model flavor	When it fits
OpenAI `text-embedding-3-small`	Fast API, solid default
OpenAI `text-embedding-3-large`	Max quality, higher cost
`all-MiniLM-L6-v2`	Free, small, great for prototypes

Key Example: Create a vector from a sentence with the OpenAI embeddings endpoint.

python

[object Object], openai ,[object Object], OpenAI

client = OpenAI(api_key=,[object Object],)

,[object Object], ,[object Object],(,[object Object],) -> ,[object Object],[,[object Object],]:
    r = client.embeddings.create(,[object Object],=text, model=model)
    ,[object Object], r.data[,[object Object],].embedding


vec = get_embedding(,[object Object],)
,[object Object],(,[object Object],, ,[object Object],(vec))
,[object Object],(,[object Object],, vec[:,[object Object],])

Embedding Model Comparison

Model	Dims	Where it runs	Cost
`text-embedding-3-small`	1536	API	Low per 1M tokens
`text-embedding-3-large`	3072	API	Higher
`all-MiniLM-L6-v2`	384	Local CPU/GPU	Free (compute only)

Key Takeaway

Pick one embedding model per index—mixing models without re-embedding breaks geometry.

3. Similarity Search

Cosine Similarity

Measures angle between vectors—popular because length matters less than direction for many text models.

Metric	Intuition
Cosine similarity	Same direction = high score
Dot product	Cares about magnitude too
Euclidean distance	Straight-line distance in space

Plain-English algorithm: (1) embed query, (2) embed corpus (or precompute), (3) score every doc, (4) return top-k.

Concept: “Nearest neighbors” in embedding space ≈ “most related paragraphs” in human space—usually.

4. Vector Databases

They store vectors + metadata and run approximate nearest neighbor (ANN) search fast.

Engine	Vibe
Chroma	Local-friendly, fast to prototype
Pinecone	Managed, scales without ops drama
Weaviate	Graph + vector hybrid features
Qdrant	OSS, rich filtering, nice Docker story

Key Example: Chroma in-memory collection + query—pattern matches most tutorials.

python

[object Object], chromadb

client = chromadb.Client()
col = client.create_collection(,[object Object],, metadata={,[object Object],: ,[object Object],})

docs = [
    ,[object Object],,
    ,[object Object],,
    ,[object Object],,
]
col.add(documents=docs, ids=[,[object Object], ,[object Object], i ,[object Object], ,[object Object],(,[object Object],(docs))])

hits = col.query(query_texts=[,[object Object],], n_results=,[object Object],)
,[object Object], doc, dist ,[object Object], ,[object Object],(hits[,[object Object],][,[object Object],], hits[,[object Object],][,[object Object],]):
    ,[object Object],(,[object Object],(,[object Object], - dist, ,[object Object],), doc)

Metadata Filtering

Add metadatas=[{...}] per row, then where={...} on query—think SQL predicates on top of semantic search.

Vendor	Standout
Pinecone	Managed scale, straightforward ops story
Weaviate	Hybrid + modules ecosystem
Qdrant	Strong OSS filtering, Docker-friendly
Chroma	Fastest “hello world” locally

Try This! Tag chunks with doc_version now; future you will thank past you when PDFs update.

Concept: Metadata filters are cheap precision—use them before you bump embedding dimensions.

5. Indexing Strategies

HNSW (Hierarchical Navigable Small World)

Graph-like index: “local hops + occasional long jumps.” Default in many DBs—great general-purpose ANN.

Knob	Tradeoff
More links / higher `ef`	Better recall, slower build/search

IVF (Inverted File)

Cluster vectors, search only promising clusters—great at huge scale; needs training / parameter tuning.

Choosing the Right Strategy

Strategy	Best for	Build	Query	Memory
Flat brute force	Tiny sets	instant	slow	low
HNSW	Most prod RAG sizes	medium	fast	higher
IVF (+ PQ variants)	Massive corpora	higher	fast	tunable

Fun Fact: ANN means approximate—you trade perfect recall for speed. Always benchmark on your queries.

Practice Exercises

Level	Idea
Beginner	20 FAQs in Chroma; compare keyword vs semantic hits
Intermediate	Same corpus with two embedding models; compare misfires
Advanced	Benchmark Chroma vs Qdrant at 100k vectors (latency + recall)

Mini-Project: Personal Knowledge Base

CLI: add path/to.md, search "question". Steps: load text → chunk → embed with one model → store in persistent Chroma → query with citations in the printed output.

Key Takeaways