Tuned catalog search as a service

You bring a catalog. We tune a domain-specific search model on it. You get a hosted /search endpoint. Integrate like Algolia — except your search actually understands Hindi, Hinglish, misspellings, and brand-as-generic queries.

Benchmark

Default SaaS = raw keyword search. Algolia, Typesense, Elastic, Meilisearch Cloud all behave similarly out-of-the-box on Indian catalogs. Numbers are Meilisearch-default as category proxy.

Who this is for

Mid-market e-commerce and marketplace teams in India running catalogs where:

• Customers type in Hindi, Hinglish, or Romanized Devanagari
• Customers misspell brand names routinely (break pad is the #1 misspelling)
• Customers search by symptoms (steering bhaari) not part names
• Customers use brand names generically (Amaron = any battery)

How it works

Stack: fine-tuned BGE-m3 multilingual embedder (open source, domain-adapted on HSN + ITI + NHTSA data) + Meilisearch BM25 with custom Indic tokenizer + 2,700-pair Hinglish bridge dictionary + class-routed cross-encoder reranker. Hybrid RRF fusion, tuned per query class.

All open-source base components. No proprietary embedding API dependency.

The bigger idea

The auto-parts endpoint is a reference implementation. The actual product I'm chasing: make catalog-search tuning agentic. You bring a catalog and 30 example queries; AI agents (Claude + Codex) walk the same 11 stages end-to-end; human in the loop only for taste decisions.

Playbook-as-code. I ran the auto-parts version manually so the playbook could exist. Read the 11-stage playbook.