Veda vs Elasticsearch

Veda vs Elasticsearch

Elasticsearch added dense_vector + kNN and ELSER on top of a text-search-first engine. Veda is a vector database built from the ground up for the multimodal RAG era. Here's an honest, benchmarked comparison.

The short version

Elasticsearch is a mature, battle-tested platform for text and log search at scale. Veda is the place to ingest any document — PDF, Office, or scans — and search and reason across text, images, audio, and video together, with RAG quality controls built in, and, on equal hardware and a compacted collection, dense-vector search that's faster too.

Dense-vector performance — benchmarked head-to-head

We ran the same 30,000-passage corpus, the same embeddings, and the same queries against Veda and a production Elasticsearch 8.19 cluster. Recall is identical; latency was measured twice — once on a shared, contended host, once on a quiet dedicated one — to isolate the index from the machine it runs on:

Veda (dedicated host)Veda (contended dev host)Elasticsearch 8.19
recall@101.0001.0001.000
search p50 (ef=96)0.53 ms36.6 ms~7 ms
ingest throughput1,837 docs/s~200 docs/s~250–310 docs/s

On comparable hardware, Veda's dense search is roughly 13× faster than Elasticsearch at equal recall — on a compacted collection. That qualifier is load-bearing and the table above does not show it, so here it is plainly.

Veda builds ANN indexes per sealed segment and merges across them at query time. The default bulk-load path seals a segment roughly every 500 records and compacts only opportunistically, so a freshly bulk-loaded 30k-row collection settles at ~38 segments. There, recall falls to 0.265 against Elasticsearch's 1.000 and Veda is the slower of the two (~47 ms vs ~7 ms server-side). Recall degrades monotonically with fragmentation: 1 segment → 1.000, ~4 → 0.853, ~38 → 0.265.

So the honest summary is narrower than the headline: a single-segment Veda HNSW is exact and several times faster than Elasticsearch; a fragmented one is neither. Call POST /v1/collections/{name}/optimize after a bulk load to force-merge, then the numbers in the table are the ones you get. Full methodology, including the case Elasticsearch wins, is in reports/2026-06-21-veda-vs-es-benchmark.md — note that it measured Elasticsearch force-merged to one segment against Veda left fragmented, because Veda had no optimize endpoint on the day it was run.

The 36.6 ms figure on the contended host wasn't an index problem — direct instrumentation showed the HNSW graph visiting only ~1.5% of nodes per query (healthy, not near-exhaustive); the gap was memory-bandwidth contention from other workloads sharing that box. Always benchmark on dedicated hardware — the same code can look dramatically different depending on what else is running on the machine.

Where Veda is categorically ahead

Elasticsearch's architecture makes these hard to bolt on; Veda was designed around them:

  • One vector space for every modality. Text, image, audio, and video share a single 1024-dimensional embedding space — index with text, query with an image, get the audio clip back. Elasticsearch has no native audio/video retrieval; non-text is handled as opaque attachments, not first-class searchable vectors.
  • Any document format, one API call. Drop in a PDF, DOCX, PPTX, XLSX, or a scanned image — Veda runs OCR, extracts tables and images, chunks, embeds, and makes it searchable in one request. Elasticsearch has no built-in document pipeline; you bolt on Tika or Unstructured and feed pre-extracted text back yourself.
  • RAG quality, not just retrieval. Faithfulness/citation scoring, an eval harness, chunker A/B testing, orphan-chunk / coverage detection, a semantic response cache, and embedding-based guardrails ship with the engine. Elastic's Agent Builder is explicitly text-centric and doesn't score faithfulness, quality, or drift.
  • Vector-native ops intelligence. Because Veda owns the embedder, its observability is semantic: embedding-based log categorization (not template matching), semantic incident correlation, spike attribution, and LLM-grounded root-cause analysis — versus Elastic AutoOps's rule-based/statistical approach.
  • Self-hosted, private, LLM-agnostic. Embedder, OCR, reranker, and LLM endpoint are all swappable and self-hostable — the whole stack runs air-gapped, for regulated or sovereignty-bound deployments.
  • 12 tunable ANN/quantization index types (HNSW, DiskANN, IVF-PQ, RaBitQ family, SPANN, SQ4, SQ8) with an explicit recall/latency dial, versus Elasticsearch's more opaque internals.

Where Elasticsearch is still ahead

We'd rather tell you this than have you find out the hard way: Elasticsearch wins decisively on maturity, operational scale, and full-text/log depth — petabyte-scale log analytics, decades of Lucene analyzer richness, Beats/Logstash/APM agents, a huge ecosystem, enterprise support, and compliance certifications. Veda is young. Don't replace a mature Elasticsearch log-analytics deployment with Veda — that's not the target. Veda's lane is multimodal ingestion and RAG, where Elasticsearch fundamentally cannot follow, plus now comparable-or-better dense-vector performance on equal hardware.

Try it yourself

The fastest way to evaluate this is on your own data: bring a folder of mixed PDFs, images, and audio, run them through the document pipeline, and search across all of them from one query box. See the quick start to get running in a few minutes.