3 patterns for multimodal RAG. Here's how they differ and when each one breaks down. Most RAG systems add multimodal support by converting everything to text first. Is your system natively multimodal, or just a conversion pipeline? The architecture choice shapes what you can query and what you lose. Shared vector space - Cross-modal search without format conversion - Requires large multimodal training datasets - Semantic drift is a real risk if training data is narrow Single grounded modality - Works with any existing text search setup - Spatial relationships in images don't survive conversion - Retrieval quality depends on captioning/transcription accuracy Separate retrieval pipelines - Best per-modality retrieval accuracy - Most complex to rank across modalities - Highest compute cost, independent search per modality Pick your pattern, clone the repo, and build it.

Exact kNN on 1M vectors: ~10 QPS HNSW on the same dataset: ~1000 QPS This is because HNSW builds a layered graph: shortcuts at the top, all vectors at the bottom. A query enters at the top layer, finds the closest node, uses it as the entry point for the layer below, and repeats until it reaches a tight local search at layer 0. It also let you tune your indexing speed vs query speed vs recall, so you can get the index that suits your needs.

30x faster than Prometheus on gauge averages sounds like a typo. It's not. Elasticsearch 9.4 rebuilt TSDS and ES|QL into a fully columnar metrics engine. Vectorized time series aggregations, zero-copy decoding, doc value skippers replacing inverted indices. The result at 1.4M time series: queries that took Prometheus 30+ seconds finish in under 2.

Elastic is heading to KubeCon Amsterdam 2026 🇳🇱 Visit booth #980 to see how we’re building OpenTelemetry to simplify ingestion, reduce pipeline complexity, and deliver consistent observability across metrics, logs, and traces with a focus on agentic observability. Come see the demos and meet the team: go.es.io/4uGtF3N