Vector search and RAG with Elasticsearch: dense_vector, kNN, hybrid search and RRF

1. Vectors and kNN #

Searching by meaning.

Mapping and query #

PUT docs
{ "mappings": { "properties": {
    "title":   { "type": "text", "analyzer": "french" },
    "content": { "type": "text", "analyzer": "french" },
    "content_vector": { "type": "dense_vector", "dims": 384, "index": true, "similarity": "cosine" } } } }

GET docs/_search
{ "knn": { "field": "content_vector", "query_vector": [0.12, -0.03, "..."], "k": 10, "num_candidates": 100 },
  "_source": [ "title" ] }

Embeddings are computed by a model (in your application, or in Elasticsearch via the inference API and the semantic_text type, which handles chunking and vectorization automatically).

2. Hybrid search #

The best of both.

RRF #

GET docs/_search
{
  "retriever": { "rrf": {
    "retrievers": [
      { "standard": { "query": { "match": { "content": "réinitialiser le mot de passe" } } } },
      { "knn": { "field": "content_vector", "query_vector": [0.12, -0.03, "..."], "k": 20, "num_candidates": 100 } }
    ],
    "rank_window_size": 50
  } },
  "size": 5
}

• BM25 excels on exact terms (error codes, product names); vector search on rephrasings.
• Reciprocal Rank Fusion combines rankings without having to calibrate incomparable scores.
• For RAG, the 3 to 5 best passages are sent to the LLM with the question, citing their sources.

Good to know: Some features depend on the version and the license level (Basic, Platinum, Enterprise): check them in the documentation and on the Elastic subscriptions page before using them in production.