Product & catalog search

The outcome

Search that understands intent, not just keywords. By combining semantic matching (meaning) with keyword matching (exact tokens) — hybrid search — and then reranking the top candidates, it handles synonyms (“sneakers” finds “running shoes”), typos, and natural-language queries (“warm jacket for hiking under £100”) over a product catalog or any large item set. It can go multimodal too: search by image to find visually similar products.

Why it matters: better relevance puts the right item at the top, which lifts conversion; understanding intent means far fewer dead-end “no results” pages; and exact-token matching still nails SKUs, model numbers, and brand names that pure semantic search would miss. One ranked list, the best of both signals.

What you build on Dodil

A vector store over your catalog that runs hybrid dense + BM25 search with reranking in a single engine, fed by embeddings from the same platform. Each piece maps to one product:

• Hybrid search + reranking — a vector store that does dense (semantic) and sparse BM25 (keyword) retrieval and fuses them, then reranks the top candidates for top-K precision. Use VBase  (managed Milvus, driven with the standard Milvus SDK) when you want to own the schema and index, or K3 Vector  when you want bucket-native search with one search mode that handles the rest.
• Embeddings + reranking models — Ignite Models : OpenAI/Cohere-compatible inference for embedding both your catalog items and incoming queries, and for the cross-encoder rerank step.
• Multimodal (optional) — embed product images alongside text so a query image returns visually similar items, reranked against a text description for “looks like X and matches Y.”

Why it’s faster and cheaper here

Without Dodil, intent-aware catalog search is several products glued together: a search vendor (or self-hosted Elasticsearch / OpenSearch) for keyword matching, plus a separate vector DB for semantic similarity, plus an embedding pipeline to turn items and queries into vectors, plus a reranker model to sharpen the top results — each a distinct account, integration, and bill, all wired together and operated by you.

On Dodil that collapses. One engine does dense + sparse (BM25) retrieval and fuses them, and reranks the top candidates — no separate keyword cluster, no separate vector DB, no fusion glue to maintain. The embeddings and the reranker come from the same platform, so there’s no third-party inference API to wire in. And it’s all reached with one IAM token — the same token is your Milvus token on VBase. The pieces are built to fit, so there’s nothing to integrate between them.

• Faster — one engine and one platform instead of four stitched-together services means search that understands intent ships in days, not weeks: less integration code, one auth, and reranking is a flag rather than a model to stand up and serve.
• Cheaper — fewer vendors and fewer bills, no fusion or embedding glue to own, and no idle search cluster or inference endpoint to keep warm and monitor.

These wins are structural — they come from removing vendors, glue, and idle capacity, not from a pricing trick.

Build it

Follow the step-by-step build in Semantic search, which stands up a vector store and runs a search both from the CLI and the Milvus SDK. To add the keyword and rerank tiers on top of that baseline:

• K3 Vector — the Hybrid + Rerank recipe  shows how one search mode runs dense + BM25 and an optional rerank flag, with guidance on when each tier helps.
• VBase — the Hybrid Search recipe  shows a Milvus schema with a dense field and an auto-populated BM25 sparse field, fused with an RRF or weighted reranker.

See also

• Use Cases — what teams build on Dodil, and why it ships faster and runs cheaper
• Semantic search — the step-by-step build behind this use case
• VBase  — managed Milvus when you want to own the schema and index
• K3 Vector  — bucket-native vector search with dense, BM25, and rerank
• Ignite Models  — OpenAI/Cohere-compatible embeddings and reranking
• Get an Access Token — the single IAM token used across all of it