Hi folks! 👋 We are currently running a Typesense collection with 3072-dimensional embeddings (float32) and facing significant RAM usage issues as the dataset grows. Current situation: • Schema uses

float[]

with

num_dim: 3072

• Memory consumption is becoming a bottleneck • Considering reducing dimensions to 1024 for better resource efficiency Questions: 1. Migration strategy: What's the recommended approach for changing embedding dimensions from 3072 to 1024 in an existing collection? Can this be done in-place (PQ compression etc) or does it require full re-indexing? 2. Memory optimization: Are there any Typesense-specific configurations or techniques to reduce memory footprint for high-dimensional vectors without changing dimensions? 3. Future considerations: Any plans for built-in vector compression/quantization features in upcoming Typesense versions? Environment: 2 set of embeddings per document (retrieval & classification), 4 GB RAM (~2.8GB used), 70K documents Would appreciate any insights/solution from the community to solve this bottleneck!

1. It requires reindexing 2. We currently don't have a way for compressing vector storage. We don't support binarization or PQ encoding yet. 3. Yes we do have plans to add these but it's not on our immediate roadmap yet.

Gotcha! Thanks for the reply, looking forward 🙂