Choosing a Real Vector Database

Part 3 introduced three real vector database options. This part covers the actual, practical trade-offs between them — a real, honest decision framework, not a "best overall" ranking that ignores genuinely different real use cases.

Chroma: genuinely simplest to start with

Real strengths: runs locally with zero real infrastructure setup,
  genuinely free, simple real Python API (part 3's example)
Real limits: not built for massive, real production scale on its
  own without additional real operational work; less mature real
  managed-hosting ecosystem than Pinecone

Chroma is the real, correct choice for development, prototyping, and a genuinely smaller-scale real production use case — exactly why this series uses it for part 6's hands-on RAG build. Bright Leaf Coffee's real FAQ (a few dozen documents) never approaches a scale where Chroma's real limits become a genuine, practical concern.

Pinecone: genuinely built for managed, real production scale

Real strengths: fully managed — no real database infrastructure to
  run or maintain yourself; built specifically for real, large-scale
  vector search with strong real performance guarantees
Real limits: a real, ongoing cost based on usage, and a genuine
  external dependency (network calls to a real, hosted service
  rather than an in-process or self-hosted database)

For GreenDesk's real knowledge-base chatbot (part 15), serving potentially thousands of employees across a real, growing internal document set, Pinecone's real managed scale becomes a genuinely more relevant, practical consideration than it is for Bright Leaf Coffee's small FAQ.

pgvector: genuinely useful when Postgres is already in the stack

Real strengths: adds real vector search directly to an existing,
  real PostgreSQL database — no separate database system to run,
  operate, or keep in sync
Real limits: genuinely less specialized and less performant at very
  large real scale than a purpose-built vector database like Pinecone

This is a real, concrete, practical fit for the Django series' own real project — a Django application already running Postgres for its normal, structured data can add vector search via pgvector without introducing an entirely separate, new real database technology into the stack.

A real, practical decision framework

Prototyping, or a genuinely small document set (under a few
  thousand): Chroma
Already running PostgreSQL, and vector search is one feature among
  several real, structured ones: pgvector
Large, growing real document set with genuine production scale and
  reliability requirements, and prefer not to operate the
  infrastructure yourself: Pinecone

The real, concrete choice this series makes, and why

Part 6's RAG build: Chroma, for Bright Leaf Coffee's real FAQ — a
  genuinely small, real document set, prioritizing simple, free,
  local development
Part 15's knowledge-base chatbot: the same real Chroma setup,
  explicitly flagged with a note on when GreenDesk's real, growing
  document set would justify migrating to Pinecone or pgvector

This series deliberately uses one, consistent real tool (Chroma) throughout its hands-on builds, for genuine pedagogical clarity — but every part explicitly names the real, concrete trigger for migrating to a different option, rather than implying Chroma is the universally correct real choice for every possible scale.

Migration is a real, genuine cost worth planning for upfront

Switching vector databases later means re-embedding and re-indexing
  the ENTIRE real document set — a real, non-trivial operation for
  a large, existing knowledge base

This is directly why the choice matters upfront rather than being treated as a reversible detail — a real, deliberate choice made with part 4's framework in mind, before a real system accumulates a large volume of documents, avoids a genuinely costly later migration.

Next: what RAG (Retrieval-Augmented Generation) actually is — the complete, technical explanation, building directly on embeddings and vector databases from these first four parts.