How do you keep vector embeddings in sync with your MongoDB data? (feedback on a planned operator feature)

[hors]

Hi everyone 👋

We're exploring a new feature for the PSMDB operator and want real-world input before we build it.

The problem we think exists: if you're building semantic search or RAG on top of MongoDB, every document needs a vector embedding, and that embedding must be regenerated every time the source text changes. Today, that means writing and operating your own sync pipeline, change streams, batching, retries, backfills, and model calls. MongoDB recently shipped this as a managed feature ("Automated Embedding" in Atlas), but it's tied to their cloud and their models.

What we're considering: the operator keeps embeddings in sync for you, using any OpenAI-compatible embeddings endpoint you point it at, self-hosted vLLM, Ollama, or a managed API. Your data never has to leave your cluster.

You'd apply something like this:

apiVersion: psmdb.percona.com/v1
kind: PerconaServerMongoDBEmbedding
metadata:
  name: product-catalog-embeddings
spec:
  clusterName: my-cluster
  inference:
    endpoint: http://vllm.ai-stack.svc:8000/v1
    model: intfloat/e5-mistral-7b-instruct
    credentialsSecret: vllm-api-key
  source:
    database: shop
    collection: products
    fields: [title, description]
  target:
    vectorField: embedding
    dimensions: 4096
  initialSync: true

The operator would run a small worker that backfills existing documents, then tails change streams and re-embeds only when the configured fields actually change. Vector search on those fields is a separate track we're also working on. This feature is about keeping the vectors themselves fresh.

What it would NOT do: deploy or manage vLLM/GPUs for you (you bring your own endpoint), chunk large documents (v1 is one document = one vector), or send your data anywhere you didn't configure.

We'd love to hear:

1. Do you run (or plan) semantic/vector search against data in MongoDB today? What does your embedding pipeline look like?
2. Where does your embedding model run, OpenAI/managed API, self-hosted (vLLM, Ollama, TEI), or nowhere yet?
3. Is "data must not leave our cluster" a hard requirement in your org, or just nice to have?
4. If you evaluated Atlas's automated embedding, what made you adopt it or pass on it?
5. What's missing from the YAML above that would make this a no-go for you?

Honest "we'd never use this because…" answers are the most valuable ones. Thanks