What should I use for semantic search: embeddings API or reranker API?

For semantic search, the usual choice is:

Use the Embeddings API for retrieval

Embeddings are best for:

• converting documents and queries into vectors
• finding semantically similar items with nearest-neighbor search
• scaling to large corpora efficiently

Typical flow:

1. Embed all documents
2. Store vectors in a vector index
3. Embed the user query
4. Retrieve top-k closest documents

Use the Reranker API to improve the top results

A reranker is best for:

• taking a shortlist of candidate documents
• reordering them by relevance to the query
• improving precision for the final results

Typical flow:

1. Use embeddings/vector search to get, say, top 20–100 candidates
2. Send those candidates plus the query to the reranker
3. Return the reranked top few results

Practical recommendation

• If you need scalable semantic search: start with embeddings
• If you want higher-quality final ranking: add a reranker on top
• If you must choose only one: choose embeddings for the core search system

Rule of thumb

• Embeddings = recall
• Reranker = precision

If you want, I can also suggest a good architecture for semantic search depending on your dataset size and latency requirements.