Embeddings

embed turns one text into a vector (embedContent); embedMany turns many (batchEmbedContents, MAX_EMBED_BATCH texts per request), in order.

jennifer

import "@deckyard/gemini/" as gemini;
use io;
use linalg;
use os;

def c as gemini.Client init gemini.client(os.getEnv("GEMINI_API_KEY"));
def q as list of float init gemini.embed($c, "gemini-embedding-001", "How do cats purr?", "RETRIEVAL_QUERY", 768);
def docs as list of list of float init gemini.embedMany($c, "gemini-embedding-001",
    ["Purring comes from the larynx.", "Dogs bark."], "RETRIEVAL_DOCUMENT", 768);
for (def d in $docs) {
    io.printf("%f|prec=3\n", linalg.dot($q, linalg.normalize($d)));
}

taskType tunes the vector for its use: RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT, SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING, QUESTION_ANSWERING, FACT_VERIFICATION, CODE_RETRIEVAL_QUERY, or "". gemini-embedding-001 takes it; gemini-embedding-2 does not, and takes the task in the text instead (such as task: classification | query: ...).
dimensions shortens the vector (outputDimensionality, 128 to 3072, 768 or 1536 recommended); 0 keeps the model's own size (3072). A shortened vector is not normalised: normalise it before comparing by dot product.

embedMany returns every vector at once. A Jennifer list of float takes several hundred bytes per number, so 1000 texts at 3072 dimensions hold over a gigabyte: embed a large corpus in slices of a few hundred texts (or ask for fewer dimensions) and store each slice before the next.

gemini-embedding-2 also embeds images, audio, video and PDFs; send those parts with call (POST /models/gemini-embedding-2:embedContent).

#Embeddings

Embeddings