[Platform] Add support for Eden AI gateway

[welcoMattic]

Q	A
Bug fix?	no
New feature?	yes
Docs?	yes
Issues	-
License	MIT

Eden AI is an AI gateway exposing hundreds of models from many
providers behind a single API, using provider/model identifiers. This adds a
symfony/ai-eden-ai-platform bridge covering both halves of its v3 API.

OpenAI-compatible endpoints

Chat completions (/v3/chat/completions, including streaming and tool calling) and
embeddings (/v3/embeddings) reuse the Generic bridge:

use Symfony\AI\Platform\Bridge\EdenAi\Factory;
use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;

$platform = Factory::createPlatform($apiKey);

$messages = new MessageBag(Message::ofUser('What is the Symfony framework?'));
echo $platform->invoke('openai/gpt-4o-mini', $messages)->asText();

$vectors = $platform->invoke('openai/text-embedding-3-small', 'Some text')->asVectors();

Expert models

The /v3/universal-ai endpoints expose non-LLM features, addressed as
feature/subfeature/provider[/model]: OCR, document parsing (invoices, resumes, identity
documents), text-to-speech, speech-to-text, image analysis (object detection, explicit
content) and image generation.

use Symfony\AI\Platform\Bridge\EdenAi\Ocr\Result\OcrResult;
use Symfony\AI\Platform\Message\Content\Audio;
use Symfony\AI\Platform\Message\Content\ImageUrl;

// OCR: raw text and bounding boxes
$result = $platform->invoke('ocr/ocr/google', new ImageUrl('https://example.com/scan.jpg'), [
    'language' => 'en',
]);

$ocr = $result->asObject();
\assert($ocr instanceof OcrResult);
echo $ocr->getText();

// Document parsing: structured data out of an invoice
$parsing = $platform->invoke('ocr/financial_parser/affinda', new DocumentUrl('https://example.com/invoice.pdf'), [
    'language' => 'en',
])->asObject();

// Text-to-speech
$platform->invoke('audio/tts/amazon/neural', 'Welcome to Symfony AI!')->asFile('welcome.mp3');

// Speech-to-text: the async job is polled transparently, and a local file is
// uploaded through /v3/upload first
echo $platform->invoke('audio/speech_to_text_async/deepgram', Audio::fromFile('./audio.mp3'))->asText();

// Image generation
$platform->invoke('image/generation/stabilityai', 'A red apple on a white table')->asFile('apple.png');

Expert models accept their input as a direct file URL, as a file ID, or as Audio,
Document and Image content, which is transparently uploaded through /v3/upload
beforehand. Speech-to-text runs on /v3/universal-ai/async, whose job is polled until it
reaches a terminal state. Options that are not root-level request fields (fallbacks,
provider_params, show_original_response, webhook_receiver,
user_webhook_parameters) are forwarded inside the input object, so per-feature
parameters like language, document_type or voice are passed as invocation options.

Notes

The catalog ships a curated subset of the available models; any other one can be
registered through the $additionalModels constructor argument. Capabilities are derived
from the metadata /v3/models exposes per entry: supports_response_schema maps to
OUTPUT_STRUCTURED, supports_function_calling to TOOL_CALLING, and the
input_modalities to the matching INPUT_* cases.

Gateways like Eden AI answer authentication failures with {"detail": "..."} rather than
an OpenAI-style error payload. The Generic completions converter already threw an
AuthenticationException on those, it just could not find the reason in that shape and
fell back to its generic Authentication failed. message. It now reads both shapes, so
the gateway's own wording reaches the exception.

Endpoints, payload keys, the upload contract and every output shape were checked against
the live API, and the thirteen examples under examples/edenai/ were all run against it
and recorded as replay cassettes under examples/tests/fixtures/edenai/, so every converter
is covered offline by ExamplesReplayTest.
One discrepancy is worth flagging for future readers: the output_schema advertised by
GET /v3/info for image/generation describes a single item without the enclosing
items list the endpoint really returns, which the converter documents.

Bridge documentation is in docs/components/platform/edenai.rst, and 115 tests cover the model
catalog, the contract normalizers, both model clients and every result converter.

cc @Guikingone

[wachterjohannes]

Structurally this holds up: scaffolding, deptrac, splitsh and the bundle wiring all match the sibling bridges, and the two things that looked like omissions (no base_url in the bundle config, two catalogs) are what the siblings do. Details inline.

• Record and replay is the direction we want to take bridges, so that a converter is pinned against the provider's real shapes rather than hand-written fixtures. This bridge is the strongest case for it so far, eight converters at once, and you already ran every example against the live API, so examples/runner --record edenai would come almost for free. ExamplesReplayTest iterates over the cassettes, so without one the bridge stays silently uncovered.

On scope I am deliberately not deciding: a gateway with eight expert-feature families and its own result types is a maintenance call, so I would like @chr-hertel to weigh in before this moves.

Verified locally, all green.

[chr-hertel]

@welcoMattic i rebased this and adopted the async job capabilities from #2408 - somehow still not all example run on my side?

can you please give it a final test before we bring it in?

[welcoMattic]

@chr-hertel I've fixed the branch and re-run examples scripts against the real Eden AI API.
All good on my side.

[chr-hertel]

Thank you @welcoMattic.