Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/phpaisdk/ollama/llms.txt

Use this file to discover all available pages before exploring further.

Ollama embeddings use the native /api/embed endpoint rather than the OpenAI-compatible one. This means the request and response format follow Ollama’s own schema, and batching multiple inputs in a single call is supported natively — Ollama returns one embedding vector per input in the same response.
The embedding adapter calls the native /api/embed path, which is derived from nativeBaseUrl (not baseUrl). By default nativeBaseUrl is http://localhost:11434. See the configuration options page for how to change it.

​
Basic single embedding

use AiSdk\Generate;
use AiSdk\Ollama;

$result = Generate::embedding('Search query')
    ->model(Ollama::model('nomic-embed-text'))
    ->run();

$vector = $result->output->vector;

​
Batch embeddings

Pass an array of strings to embed multiple inputs in one request. The response contains one EmbeddingData entry for each input, in the same order: 
$result = Generate::embedding(['Search query', 'Document text'])
    ->model(Ollama::model('nomic-embed-text'))
    ->dimensions(256)
    ->providerOptions('ollama', [
        'truncate' => false,
        'keep_alive' => '10m',
    ])
    ->run();

$queryVector = $result->embeddings[0]->vector;
$documentVector = $result->embeddings[1]->vector;

​
Provider options

The following provider-specific options can be passed through providerOptions('ollama', [...]):
OptionTypeDescription
truncateboolPassed unchanged to Ollama. When false, inputs longer than the model context window are not silently truncated — Ollama returns an error instead.
keep_alivestringA duration string (e.g. '10m', '1h') controlling how long Ollama keeps the model loaded in memory after the request completes.
dimensionsintRequested output dimensions. You can also set this via the standard ->dimensions() fluent method. The installed model is the authority on which dimensions are actually supported.

​
Response fields

The returned EmbeddingResponse exposes the following fields:
  • $result->embeddings — An array of EmbeddingData objects, one per input. Each has a .vector property containing a float[] of embedding values.
  • $result->output->vector — Shorthand for $result->embeddings[0]->vector, the vector for the first (or only) input.
  • $result->usage->inputTokens — The number of prompt tokens evaluated, taken from Ollama’s prompt_eval_count field.
  • $result->providerMetadata['ollama'] — Provider-specific metadata returned by Ollama:
    • model — The model name as reported by Ollama, falling back to the requested model ID if the server omits it.
    • total_duration — Total request duration in nanoseconds.
    • load_duration — Time spent loading the model into memory, in nanoseconds.
Ollama must return exactly as many embedding vectors as there were input strings. If the count does not match, the adapter throws an InvalidResponseException. This can happen if Ollama silently drops inputs or returns a malformed response.

Build docs developers (and LLMs) love

Get started for freeTalk to us