I’m unhappy with Elasticsearch semantic search quality

Sorry you’re dealing with that — semantic search quality in Elasticsearch can be good, but it’s easy for it to underperform if the embeddings, indexing strategy, or query flow aren’t tuned.

A few common causes:

• Embedding model mismatch: the model may not fit your domain or query style.
• Chunking issues: documents may be split too large or too small.
• No hybrid search: semantic-only often misses exact keyword matches.
• Wrong similarity / field setup: vector field mapping, normalization, or kNN parameters may be off.
• Poor query formulation: the query text may not match how the embeddings were trained.
• Insufficient re-ranking: top results from kNN may need a second-stage ranker.

Typical ways to improve:

1. Use hybrid retrieval

   • Combine BM25 keyword search + vector search.
   • This often gives a big boost over semantic-only.

2. Improve chunking

   • Aim for semantically coherent chunks, not arbitrary token splits.
   • Keep chunks small enough that each one has a clear topic.

3. Try a better embedding model

   • Especially if your content is technical, legal, product-specific, or multilingual.

4. Add re-ranking

   • Use a cross-encoder or LLM-based reranker on the top 20–100 candidates.

5. Tune kNN settings

   • Increase candidate pool, adjust num_candidates, and test recall vs latency.

6. Evaluate with a small labeled set

   • Create 20–100 representative queries and judge top results.
   • This helps identify whether the problem is retrieval, chunking, or ranking.

If you want, I can help you diagnose it systematically. For example, I can walk you through:

• a debug checklist
• a hybrid search setup
• better chunking strategies
• or how to evaluate Elasticsearch semantic search quality

If you share:

• your Elasticsearch version,
• the embedding model,
• your mapping,
• and an example query + bad results,

I can suggest concrete fixes.