High Vector DB Question 168 of 223

How do you keep embeddings consistent during a full re-embed at scale?

GenAI / LLM · Speak this in 60–90 seconds · Faridabad & Delhi NCR

PICTURE THIS: DATABASE INDEX

Without indexScan every row

With indexJump to keys

CostWrites slower

Simple meaning

Version the embedding model in metadata, dual-write or blue-green a new index, then atomically switch the query path.

WHY — Tokens instead of words?

LLMs use tokens (not full words) because it helps them:

This is a process

question about Vector DB.

Panels listen for order,

trade-offs, and what you would actually do on a GenAI / LLM project - not buzzwords.

Stable token IDs

Each piece maps to a number the network can learn.

Fits the model

Fixed pieces are what transformers expect as input.

STEPS — What happens with tokens?

Before the model can read a sentence, it goes through these steps:

1
Version the embedding model
in metadata, dual-write or blue-green a new index, then atomically switch the query path.
2
Mixing two model versions
in one ANN index corrupts nearest neighbors.
3
CDC from the source
of truth should drive the job, not ad-hoc scripts.
4
Context mix
Attention looks at nearby tokens together.
5
Next token
The model scores what should come next.
6
Decode
IDs turn back into readable text.

EXAMPLE — See it in action

Here's a short line you can speak, broken into clear beats:

Say this line

“Mixing two model versions in one ANN index corrupts nearest neighbors.”

Break into beats

Mixingtwomodelversionsinone

Speaking order

2987408337471632900

Note: Adapt this scaffold to your own project — keep it under 60–90 seconds.

Key takeaway

Version the embedding model in metadata, dual-write or blue-green a new index, then atomically switch the query path. Mixing two model versions in one ANN index corrupts nearest neighbors.