Product Search That Understands Thai
Why Thai breaks ordinary store search, and how hybrid keyword and semantic search finds the right products from slang, mixed-language and descriptive queries.
Thai product search AI is search that understands how Thai shoppers actually type: no spaces between words, shortened slang, English brand names written in Thai letters, and descriptions instead of product names. Standard keyword search often returns nothing for these queries. Search built on language models and Thai-aware text processing can match meaning, so the shopper finds the item without guessing your catalogue wording.
This article is for e-commerce and merchandising teams in Thailand whose search box returns "no results" more often than it should. It explains why Thai breaks ordinary search, how modern approaches fix it, what to test, and what can go wrong.
Why Thai breaks ordinary search
Most search engines were designed for languages that put spaces between words. Thai does not. A phrase like "รองเท้าผ้าใบสีขาว" (white canvas shoes) has no spaces, so the engine must decide where words begin and end. If it splits wrongly, it matches the wrong products or none.
Other problems add up:
- Word segmentation. Different tools split the same text differently, so the catalogue and the query may be cut in different places.
- Spelling variation. Shoppers drop tone marks, repeat letters for emphasis or misspell borrowed words.
- Transliteration. A brand may be typed in English, in Thai letters, or both. "ไอโฟน" and "iPhone" mean the same item.
- Mixed language. Queries such as "เสื้อ oversize สีดำ" combine Thai and English in one phrase.
- Descriptive queries. A shopper writes "กระเป๋าใส่โน้ตบุ๊กกันน้ำ" rather than the product title you used.
None of this is exotic. It is how people type on phones every day.
Three approaches, and when each fits
| Approach | How it works | Strength | Weakness |
|---|---|---|---|
| Keyword search with Thai tokenization | Splits text into words and matches terms | Predictable, fast, easy to tune | Misses synonyms and descriptive queries |
| Semantic (vector) search | Converts text to numbers that capture meaning, then finds the nearest products | Handles synonyms, typos, descriptions | Can return loosely related items; harder to explain |
| Hybrid search | Combines keyword and semantic results, then re-ranks | Best coverage for most catalogues | More parts to set up and monitor |
For most Thai online stores, hybrid search is the practical choice. Keywords keep exact matches such as model numbers and sizes reliable. Semantic matching rescues the queries that keywords miss.
What a good Thai product search AI pipeline does
Think of search as a short chain of steps. Each can be tested separately.
- Normalize the query. Clean spacing, unify spelling variants and map common transliterations to the same form.
- Segment Thai text. Use a tokenizer built for Thai, and use the same one for your catalogue and for queries.
- Understand intent. Detect color, size, brand, category and attributes such as material or use. A query for black oversize shirts should filter or boost on those attributes.
- Retrieve candidates. Run keyword and semantic retrieval together.
- Re-rank. Order results using relevance, availability and business rules, such as hiding out-of-stock items.
- Learn from behavior. Use clicks and add-to-cart signals, carefully, to improve ordering over time.
Your catalogue is the foundation. Titles stuffed with keywords and missing attributes hurt every step. Clean, structured data helps more than any model.
Synonyms, slang and brand names
A synonym list is still valuable, even with semantic search. It gives you direct control for cases that matter commercially. Build it from your own data:
- Review zero-result queries and the top searches each week.
- Group variants: Thai and English names, nicknames, common misspellings.
- Add category words shoppers use that differ from your internal terms.
- Keep a record of why each rule exists, so someone can remove it later.
Language models can draft candidate synonyms and attribute tags at catalogue scale, in Thai and English. Have a merchandiser review them before they go live, since a wrong synonym can surface wrong products across thousands of queries.
Beyond text: images and conversation
Some shoppers search with a photo or describe an item in a chat. The same meaning-based approach extends to these cases. Image search can match a photo to similar products, and a LINE sales agent can use product search behind the scenes to answer "do you have something like this?" Treat these as later steps after text search is stable. Our AI for retail and e-commerce page shows how language agents, recommendations and demand forecasting fit together, and our post on AI customer service on LINE covers the chat side.
What search data tells your merchandising team
Search is also a listening tool. The queries shoppers type show what they want in their own words, including products you do not stock. Review them every week with the people who buy and describe your range.
- Zero-result queries point to missing products, missing synonyms or catalogue text that does not match how people talk.
- Repeated reformulations, where a shopper tries three phrasings in a row, show queries that fail quietly.
- High-exit queries show where results appeared but nothing was clicked, which often means the ranking is off.
Feed what you learn back into titles, attributes and synonym rules. This weekly loop is simple, and it keeps the system improving after launch rather than drifting.
How to test it
Do not judge search by trying a few queries yourself. Build a small test set.
- Collect a few hundred real queries from logs, including common, rare and zero-result ones.
- For each, have a merchandiser mark which products should appear in the top results.
- Run your current search and the new approach on the same set.
- Compare how often a correct product appears near the top, and review the misses by hand.
- Repeat after every change, so you notice regressions.
Add edge cases deliberately: no-space phrases, mixed Thai-English, misspellings, slang and very short queries. Type them the way shoppers do on a phone, not as tidy text.
Privacy and data handling
Search logs can include personal information, especially if shoppers type names or phone numbers into the box. Under Thailand's PDPA, limit what you store, set retention periods, and avoid sending raw logs to outside services without checking how they are processed. If you use personalization based on browsing history, make sure consent is in place. Check current guidance from Thailand's PDPC and confirm with your own counsel. This is general information, not legal advice.
What can go wrong
Plausible but wrong results. Semantic search may return items that sound related but are not what the shopper wants, such as accessories instead of the product. Keep strong keyword and attribute filters in the mix.
Ignoring stock. Showing sold-out items at the top frustrates shoppers. Include availability in ranking.
Tokenizer mismatch. If query and catalogue are segmented differently, matches fail silently. Use one tokenizer and test it.
Over-personalization. Ranking that changes too much between users makes results hard to trust and hard to debug.
No monitoring. Catalogues and trends change. Review zero-result queries and click patterns every week.
When not to rebuild search
If your catalogue is small and your products have simple names, a well-tuned keyword search with a good synonym list may be enough. If your product data is incomplete, fix that before adding models. Better data usually improves results more than a new engine.
How to start
- Pull the numbers you already have. Zero-result queries, top searches and exits after search.
- Build the test set described above.
- Fix the catalogue basics. Titles, attributes, variants and images.
- Pilot hybrid search on one category and compare against your current results.
- Review weekly, adjust synonyms, and expand to more categories.
You can also see how we approach the sector on our retail and e-commerce page and across our solutions.
Next step
If shoppers are leaving after a failed search, our consultants can review a sample of your queries and catalogue and suggest where to start. Contact us.
Frequently asked questions
Why does normal search struggle with Thai?
Thai has no spaces between words, so the engine must guess where words end. Spelling variation, transliterated brands and mixed Thai-English queries make matching harder.
Is semantic search better than keyword search?
Each has strengths. Keywords handle exact terms like sizes and model numbers; semantic search handles synonyms and descriptions. Most stores do best with a hybrid of both.
How do we test search quality?
Collect a few hundred real queries, have merchandisers mark the products that should appear, and compare engines on the same set. Repeat after every change.
Do we need clean product data first?
Yes. Clear titles, attributes and variants improve results more than a new engine. Fix catalogue basics before adding models.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
