Vector search for real-world intent
Search by meaning,
not just keywords.
Search that understands Thai, and understands what people are asking. Send records to intent-search over HTTP and search them without running an index; every hit says whether it matched the words, the meaning, or both.
Minimal setup
Send records as they are, up to 5,000 per call. The schema is inferred.
Built for developers
HTTP and JSON under
/v1. A storefront query takes 9.8 ms p95 at 1M records, on one machine.Thai-first search
Thai is cut into words with a dictionary and a language model, not at spaces.
Ergonomic Mesh Chair Proscore 0.962
High-breathability lumbar sync mechanism, posture correction
matched: meaningErgonomicsActive Noise-Cancelling Headphonesscore 0.918
40dB hybrid ANC, multi-point focus soundscapes
matched: meaningAudio & FocusMotorized Dual-Motor Standing Deskscore 0.884
Anti-collision presets, smooth sit-to-stand transitions
matched: meaningWorkspaceMechanical Pomodoro Focus Timerscore 0.851
Visual time-blocking dial with ambient warm light cues
matched: meaningProductivity
How it works
From sign-up to results ranked by words and meaning in four steps.
Register and copy your keys
Your first dataset is created at sign-up with two keys: a service key for your server and a client key for the browser. Revoking a key replaces it.
Send records as they are
Up to 5,000 records per call, in the shape your database already holds them. The schema is inferred, and
POST /v1/syncpublishes the batch to the index.POST /v1/documentsSearch with what people type
Pass the raw text from the search box. Thai is cut into words before matching, and full records come back whole, including fields that were not searched.
POST /v1/searchTurn on meaning
Enable vectors and intent-search embeds every record with bge-m3. Search with
"mode": "hybrid"to merge the keyword and vector rankings. You never build a vector yourself.PUT /v1/vectors
Frequently asked questions
What is intent-search?
intent-search is a search service for developer teams. Send records over HTTP and search them without running an index: Thai is cut into words with a dictionary and a language model, keyword (BM25) and vector rankings merge in one query, and every hit says whether it matched words, meaning or both.
How does intent-search split Thai text into words?
With a dictionary and a unigram language model, not at spaces. It takes the shortest path through a lattice of character clusters, with costs from the Thai National Corpus, and scores a word F1 of 0.8224 on VISTEC-2021 and 0.7717 on BEST-2010. It also normalizes tone-mark order, repeated characters and ๆ.
Can it find results that share no words with the query?
Yes, once vectors are on. intent-search embeds every record and every hybrid query with bge-m3, so อยากดูเด็กลง ("I want to look younger") finds เซรั่มลดริ้วรอย, an anti-wrinkle serum, with no word in common. Hybrid search merges the keyword and vector rankings with reciprocal rank fusion, and each hit is labelled words, meaning or both.
How fast is a search?
A storefront query takes 9.8 ms at p95 over 1M records and 35.7 ms over 10M, where 13 of 20 query shapes finish under 50 ms. Vector search takes 0.31 ms; embedding the query with bge-m3 adds 18 to 30 ms. nsbench measured these numbers on one aarch64 macOS machine.
Does it correct typos?
Partly. Near spellings match only through an explicit fuzzy clause, which is off unless you add it; over 10M records a fuzzy query measured 218 ms. When nothing matches, intent-search fixes a query typed on the wrong keyboard layout, Thai for English or English for Thai.
Is there an SDK?
No. intent-search is plain HTTP and JSON under /v1, so any language with an HTTP client can call it, and the quick start gives a TypeScript fetch helper to copy. The query DSL is shaped after Elasticsearch's, so existing query bodies work. An AI agent can integrate it from your service key and the docs' /llms.txt.
Does each dataset get its own index?
Yes. Every dataset has its own physical index. Each dataset starts with two keys: a service key for your server and a client key for the browser, and revoking a key replaces it. Synonyms are set per dataset and apply at query time, with no reindex.
What does the Free tier include?
One dataset of up to 10,000 records, 10,000 searches and 1,000 vector searches a month, type-ahead suggestions, 20 requests per second and 30 days of usage history, with no overage. Starter, Growth and Scale raise the limits, up to 10M records per dataset and 1,000 requests per second. Prices are not set yet.