Product

Say it however you'd
actually say it.

Not a translation tool. Sequena figures out what someone means — in Telugu, half-English, a typo, a voice note, a sentence with no product name in it at all — and hands your application a clean, structured answer it already knows how to use.

What gets extracted

One sentence carries a lot more than a product name.

A shopper saying "5 kilo kandi pappu kavali" isn't just naming toor dal — they're also telling you the quantity, the unit, and the intent. Sequena pulls out all of it, not just the noun.

Language & entity

language dialect / regional variation transliteration product / entity category brand model

Attributes & constraints

quantity unit size color variant material specifications price / budget

Preference & evidence

preferences reviews ratings ranking constraints availability

Intent & context

negation context previous-order reference multi-product intent comparison intent discovery / ordering intent conversational context
Multimodal by design

Type it, say it, or show it a photo. Same result.

This was never about multimodal AI for its own sake — it's that people don't only type. They talk, and sometimes they'd rather just point a camera at something.

🎙Voice
↓
ASRSpeech-to-text
⌨Text
↓
goes straight in
🖼Image
↓
👁Vision processing
↓   ↓   ↓
Sequena understanding
↓
✓Structured intent
Not a black box

Every match comes with its receipt.

Sequena doesn't just hand back a ranked list and ask you to trust it. Every result points to the exact sentence, spec line, or review that justified it — so "why did it pick this?" always has a real answer.

She typed: "laptop bag with separate shoe compartment"
Sequena foundTravelPro Laptop Backpack
Why this one?
✓descriptionThe listing's description says "a dedicated shoe compartment to keep footwear separate from clothes" — word for word what she asked for.
✓titleThe title just says "Travel Laptop Backpack" — a keyword search would never have found this.
How ranking works

Addition, not a mystery.

A result ranks higher because it checked more of the boxes below — not because a model quietly decided so. That's what makes the evidence above possible to show at all.

final_score = semantic_match        // title, description, specs, reviews
            + category_match
            + evidence_target_match  // she specifically asked about reviews/specs
            + attribute_match        // brand, waterproof, 5G, etc.
            + constraint_score       // price / rating / pack size satisfied
            - exclusion_penalty      // she said "not X" and X showed up anyway
            + quality_signal         // rating × review volume
In practice

Two real queries, heard correctly.

He said:

"5 kilo kandi pappu kavali"

Sequena heard: 5kg of toor dal.

{
  "language": "te-en",
  "intent": "product_search",
  "entity": "Toor Dal",
  "quantity": { "value": 5, "unit": "kg" }
}

She said:

"good cotton saree for my mother, under 1500"

Sequena heard: a gift, not just a product.

{
  "intent": "product_search",
  "product": "saree",
  "audience": "mother",
  "material": "cotton",
  "budget": { "max": 1500, "currency": "INR" }
}

Sequena doesn't own the catalog either of these results comes from — it hands this structured answer to whatever search or ranking system the application already runs.

See it live

Try it yourself — it's not a mockup.

Type a real regional-language query into the actual engine, running live in your browser, and watch it work it out in real time.

Try a real query →