Vector Database — Module 14.1

How This Lesson Fits the Module & Volume

Retrieval needs somewhere to store vectors, IDs, and metadata. A vector database (or vector-capable store) provides ANN indexes, filters, and CRUD for embeddings at scale. Module 14.2 will deepen with FAISS and friends; here we establish the product concept and ops responsibilities.

Learning Objectives

By the end of this lesson, students should be able to:

Define a vector database as storage + ANN search over embeddings.
Contrast libraries (FAISS) vs managed vector DBs vs pgvector-style extensions.
List core operations: upsert, delete, search, filter, backup.
Explain why payload/metadata is first-class for RAG tenancy and citations.
Identify scaling concerns: RAM, replicas, index rebuilds.
Preview vector search as the query API over that store.

Definition

A vector database is a system optimized to store high-dimensional vectors with identifiers and metadata, and to answer nearest-neighbor queries efficiently (often via approximate indexes), optionally with filtered search.

Deployment Shapes

Shape	Example	Tradeoff
Library / embedded	FAISS in-process	Fast; you own durability & HA
DB extension	pgvector	SQL joins; ANN maturity varies
Managed vector DB	Cloud services	Ops ease; vendor & cost model

Must store

Vector + id
Chunk text or pointer
Filterable metadata

Must support

k-NN / ANN search
Upsert & delete
Metric choice

Nice to have

Hybrid sparse+dense
Multi-tenant isolation
Snapshots / CDC

Code: In-Memory Store Sketch

import numpy as np
from dataclasses import dataclass

@dataclass
class Row:
    id: str
    vector: np.ndarray
    text: str
    metadata: dict

class TinyVectorDB:
    def __init__(self):
        self.rows: list[Row] = []

    def upsert(self, row: Row):
        self.rows = [r for r in self.rows if r.id != row.id] + [row]

    def search(self, q: np.ndarray, k: int = 5, where: dict | None = None):
        cand = self.rows
        if where:
            cand = [r for r in cand if all(r.metadata.get(k_) == v for k_, v in where.items())]
        if not cand:
            return []
        M = np.stack([r.vector for r in cand])
        scores = M @ q  # assumes L2-normalized
        idx = np.argsort(-scores)[:k]
        return [(cand[i], float(scores[i])) for i in idx]

# Production: replace brute force with ANN (HNSW/IVF) — see Indexing / FAISS.

Strengths

Purpose-built for similarity
Filters + vectors in one query
Scales beyond numpy loops

Tradeoffs

Another system to operate
Index params affect recall
Re-embed storms on model change

Common Misconception

“A vector DB replaces our document store.” Usually you still keep canonical documents in object storage or a CMS; the vector DB indexes derived chunks. Treat vectors as a search projection, not the system of record.

Knowledge Check

Short Answer: What query type are vector DBs optimized for? Answer: Nearest-neighbor / similarity search over vectors.
True/False: Metadata is optional noise in RAG indexes. Answer: False—filters, tenancy, and citations depend on it.
Multiple Choice: FAISS is best described as: (a) a library, (b) a social network, (c) a CSS framework. Answer: (a).
Short Answer: Name one vector DB deployment shape. Answer: Embedded library, DB extension, or managed service.
True/False: Upsert/delete matter for fresh knowledge bases. Answer: True.
Multiple Choice: System of record for docs should usually be: (a) only ANN graph, (b) canonical doc store + vector projection, (c) browser cookies. Answer: (b).
Short Answer: Why does model change stress a vector DB? Answer: All vectors must be re-embedded and re-indexed.
True/False: Brute-force search scales forever. Answer: False—use ANN at scale.
Multiple Choice: Next lecture: (a) Vector Search, (b) Vol. 1 only, (c) printers. Answer: (a).
Short Answer: What does filtered search combine? Answer: Metadata predicates with vector similarity.

Key Takeaways

Vector DBs store embeddings + metadata for ANN retrieval.
Choose library vs managed vs SQL-extension based on ops needs.
Keep canonical documents elsewhere; index is a projection.
Next: Vector Search.

Trainer’s Guide

Lab: Implement TinyVectorDB; then discuss which production features are missing for a 10M-chunk corpus.

Prompt: When is pgvector enough vs a dedicated vector service?

Recap: Vector databases host the searchable projection of your knowledge base. Continue with Vector Search.