AI Product Search in India

Shoppers increasingly expect to type a plain-language query — a cotton kurta under fifteen hundred rupees, size M — and get relevant results, not just keyword matches. Building that kind of search means combining unoblox's native embeddings and reranking endpoints with a chat model for query understanding, all billed in rupees through one endpoint.

The anatomy of AI-powered product search

Semantic product search generally has three parts: an embedding step that turns product titles and descriptions into vectors for similarity search, a reranking step that reorders the top candidates by relevance to the actual query, and, optionally, a chat model that parses a natural-language query into structured filters (price ceiling, size, category) before the search even runs. unoblox exposes all three as native endpoints.

Where each unoblox endpoint fits

• /v1/embeddings — embed your product catalog once, then embed each incoming query for similarity search against it
• /v1/rerank — reorder your top candidate results by relevance before showing them to the shopper
• /v1/chat/completions — parse a natural-language query into structured filters, or generate a short explanation of why a result matched

The embedding and reranking models in the catalog change as new ones are added — check /models for the current model and its live rupee rate rather than assuming a fixed price.

Chat models for query parsing

Model	₹ input / ₹ output (per 1M tokens)	Notes
deepseek-ai/deepseek-v4-flash	₹9.11 / ₹18.21	Lowest-priced paid option in this table — good for parsing high query volumes.
qwen/qwen3-235b-a22b-instruct-2507	₹9.11 / ₹55.67	Stronger structured-extraction quality at a similar input rate.
openai/gpt-5-mini	₹25.30 / ₹202.46	A balanced mid-tier option.
qwen/qwen3.8-27b	FREE (₹0)	Free for prototyping your query-parsing prompts.

Sample query-parsing call

curl https://api.unoblox.ai/v1/chat/completions \
  -H "Authorization: Bearer ub-gw-xxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Extract category, max price in rupees, and size from this search query as JSON: cotton kurta under fifteen hundred rupees size M"}]
  }'

Budgeting the pipeline

Because every step — embedding, reranking, and query parsing — is billed per million tokens in rupees with a GST tax invoice issued at each top-up, you can estimate the full pipeline's cost by multiplying expected query volume by each step's rate, without juggling separate vendor bills for each part.

Frequently asked questions

Do I need three separate vendors for embeddings, reranking, and chat? No — unoblox exposes all three as native endpoints behind one key and one rupee invoice.

Which embedding or reranking model does unoblox use? This changes as the catalog is updated — check /models for the current model and its live rate.

Can a chat model replace embeddings for search? Not efficiently at catalog scale — embeddings plus reranking are built for fast similarity search over large catalogs; use a chat model for the query-understanding step, not for scoring every product directly.

Is there a free way to prototype query parsing? Yes, qwen/qwen3.8-27b is free (₹0).

Is pricing in rupees across all three endpoints? Yes — embeddings, reranking, and chat completions are all priced per million tokens in rupees.

Do I need a new SDK for the chat portion? No — it's OpenAI-compatible, so existing chat-completions code works by pointing at https://api.unoblox.ai/v1.

Get started in rupees → https://unoblox.ai/sign-in