Guides

Building Hybrid keyword and vector search with Algolia an...

This guide provides a structured approach to implementing an AI-powered search system, focusing on hybrid keyword-vector search, reranking, and UI integration. Steps address latency, indexing, and ranking challenges specific to AI search workflows.

4-6 hours5 steps

Set up search infrastructure

Provision an Elasticsearch cluster with vector support or configure Pinecone with OpenAI embeddings. Install required plugins like elasticsearch-knn for vector search capabilities.

setup_elasticsearch.sh

docker run -e "discovery.type=single-node" -p 9200:9200 docker.elastic.co/elasticsearch/elasticsearch:8.7.0

⚠ Common Pitfalls

•Missing vector plugin causes index failures
•Incompatible Elasticsearch version with vector extensions

Create hybrid search index

Design an index mapping that combines keyword fields (text) with vector fields (embeddings). Use a multi-field approach to enable both term matching and semantic similarity searches.

hybrid_index.json

{
  "mappings": {
    "properties": {
      "content": { "type": "text" },
      "embedding": { "type": "dense_vector", "dims": 1536 }
    }
  }
}

⚠ Common Pitfalls

•Incorrect vector dimensions cause indexing errors
•Missing keyword fields limit traditional search capabilities

Implement hybrid search query

Construct a search query that combines match_all with a knn_search for vector similarity. Use function_score to weight results by both keyword relevance and vector distance.

hybrid_query.json

{
  "query": {
    "function_score": {
      "query": { "match_all": {} },
      "knn": { "field": "embedding", "k": 10 },
      "score_mode": "max"
    }
  }
}

⚠ Common Pitfalls

•Improper score_mode leads to unexpected ranking
•KNN search without filtering causes excessive computation

Add reranking with business logic

Use a separate API (e.g., Cohere Rerank) to reorder results based on custom criteria after initial vector search. Pass original query and results to the reranker for context-aware sorting.

rerank_results.py

import cohere
co = cohere.Client('YOUR_API_KEY')
response = co.rerank(
  query='search term',
  documents=[result['text'] for result in initial_results],
  top_n=5
)

⚠ Common Pitfalls

•Reranking API rate limits during high traffic
•Missing document context causes poor reranking decisions

Build search UI components

Create a search bar with auto-suggestions, result cards showing confidence scores, and filters for different content types. Implement loading states for asynchronous search requests.

SearchBox.jsx

function SearchBox({ onSearch }) {
  return (
    <input 
      type="text"
      placeholder="Search..."
      onChange={(e) => onSearch(e.target.value)}
    />
  );
}

⚠ Common Pitfalls

•Ignoring loading states causes UI freezes
•Missing confidence indicators leads to unclear result quality

What you built

This implementation addresses core AI search challenges through hybrid indexing, reranking, and structured UI components. Validate each stage with real user queries and monitor performance metrics for continuous optimization.