One Lambda fans a query into a prefix match and a 512-dimension cosine search

The lexical query can leave for DynamoDB as soon as the request arrives. The semantic one waits for a vector. Amazon Titan Text Embeddings V2 has to turn the query text into 512 numbers first, and only then can SearchVectors compare it against what is stored ^[9]. The two branches do overlap in time, so the clock is set by the slower one, and the semantic branch is two network calls deep before any comparison happens ^[21].

The projection is the part worth copying. The vector index carries the product attributes the search response needs, so a SearchVectors result answers the request without a further read ^[15]. Take the projection away and the flow goes back to the familiar one, where you collect identifiers from the vector result and then read the items back by key ^[26]. You declare that saving in the index definition ^[15].

You cannot change the embedding settings casually. The project pins amazon.titan-embed-text-v2:0, 512 dimensions, and normalization enabled ^[13]. Products and queries must use the same model, the same number of dimensions, and the same normalization ^[12]. Titan V2 also offers 256 and 1,024 dimensions, according to the Bedrock documentation cited in the tutorial ^[14], so 512 is a decision about the whole corpus: move to another size and every stored product has to be embedded again before a single query gets better ^[22].

Ordering comes from the distance function configured on the index, and DynamoDB does the nearest-neighbour step with an approximate index ^[10]. The tutorial uses cosine distance, where 0 means identical vectors and lower values mean closer ^[11]. One query run through a comparator with the sign backwards looks exactly like a broken index. The walkthrough does not publish latency or recall numbers, so treat it as a shape to copy; the author's production case is help article titles, and the tutorial indexes products ^[17].

Adoption cost is version floors and IAM. Node.js 22 or later, npm 11 or later, permissions for DynamoDB, Lambda, API Gateway, CloudFormation and IAM, plus bedrock:InvokeModel, plus an account already bootstrapped for CDK ^[16]. Bedrock model access is usually enabled by default in new accounts, though availability still varies by Region ^[20]. The table runs on on-demand capacity ^[8]. Indexing and querying are separate flows, and nothing is searchable until its embedding has been generated and stored ^[18].

The walkthrough describes the project and the AWS services as of September 2026, about a month after general availability ^[19][23]. The author's advice is to review the official AWS documentation before using the design in production ^[19].