AI Search understands the intent behind a query, not just its tokens. A query like users created recently returns records from PostgreSQL, MySQL, MongoDB, and Redis simultaneously, even though none of those records contain the words "created", "recently", or "users" as a search term. The engine understands what those words mean and finds the semantically closest data.

Like regular Search, AI Search is zero-config, always tenant-isolated, and queries all databases at once. The difference is the matching engine underneath.

# Keyword search - only finds exact word matches
$ tdb search --tenant demo --query "users created recently"
0 results in 48ms

# AI Search - understands meaning, returns semantically relevant records
$ tdb ai-search --tenant demo --query "users created recently"

DATABASE    COLLECTION  ID   SCORE  CREATED       CONTENT
PostgreSQL  users       52   0.379  Mar 05 10:04  email: [email protected] | name: CrossDB PG
MySQL       users       -    0.374  -             email: [email protected] | name: MySQL User
PostgreSQL  users       -    0.371  -             email: [email protected] | name: UniqueID PG
Redis       users       -    0.371  -             email: [email protected] | name: HashUser
20 results in 326ms

When a query arrives, it is first converted into a 768-dimension vector - a numerical representation of its meaning. That vector is used to find semantically similar documents in the index. Simultaneously, a standard keyword phase runs BM25 term matching on the same index. Both phases produce independent ranked lists. The final results are a weighted blend of both scores, controlled by the weights parameter.

AI Search operates on semantic vectors, not on surface text. Because the embedding model maps meaning across languages into the same vector space, a query written in one language will find semantically relevant records written in another. The same tenant data is queryable regardless of what language the query arrives in.

# Query in a different language - returns English-language records
$ tdb ai-search --tenant demo \
    --query "query written in any language" \
    --weights keyword=0.0,semantic=1.0

DATABASE    COLLECTION  ID   SCORE  CONTENT
PostgreSQL  users       54   0.375  email: [email protected] | name: Pedro Garcia
PostgreSQL  users       52   0.363  email: [email protected] | name: CrossDB PG
Redis       users       -    0.361  email: [email protected] | name: CrossDB Redis
PostgreSQL  users       53   0.358  email: [email protected] | name: Final4 PG
20 results in 287ms

The weights field takes two values that must sum to 1.0. The first is the keyword weight, the second is the semantic weight. The final score for each result is computed as:

final_score = (bm25_score * keyword_weight) + (vector_score * semantic_weight)

A higher keyword weight makes exact word presence matter more. A higher semantic weight makes conceptual closeness matter more. The default 0.5/0.5 is a balanced starting point.

AI Search requests are rate limited by plan. If you exceed the limit you receive a rate_limited error. Regular keyword search is not subject to the same limits.

{
  "success": false,
  "code": "rate_limited",
  "error": "AI Search rate limit exceeded. Retry after 60 seconds."
}