feat: add Amazon Nova 2 multimodal embeddings support

[gabrielkoo]

Summary

• Adds amazon.nova-2-multimodal-embeddings-v1:0 to SUPPORTED_BEDROCK_EMBEDDING_MODELS
• Implements NovaEmbeddingsModel using the taskType/singleEmbeddingParams request format from the Nova 2 embeddings docs
• Supports batch inputs (one API call per text), custom dimensions (256/512/1024/2048/3072, default 3072), and float/base64 encoding formats

Test plan

Start the gateway:

cd src && uvicorn api.app:app --port 8000

Then run the following (requires pip install openai):

import os
from openai import OpenAI

BASE_URL = os.environ.get("BASE_URL", "http://localhost:8000/api/v1")
MODEL = "amazon.nova-2-multimodal-embeddings-v1:0"
client = OpenAI(base_url=BASE_URL, api_key=os.environ.get("AWS_BEDROCK_BEARER_TOKEN", "dummy"))

# Single text
resp = client.embeddings.create(model=MODEL, input="Hello, world!")
assert len(resp.data[0].embedding) == 3072
print(f"[PASS] single text -> dim={len(resp.data[0].embedding)}")

# Batch
resp = client.embeddings.create(model=MODEL, input=["semantic search", "vector database", "RAG pipeline"])
assert len(resp.data) == 3
print(f"[PASS] batch (3 texts) -> dim={len(resp.data[0].embedding)}")

# Custom dimensions
resp = client.embeddings.create(model=MODEL, input="test", dimensions=256)
assert len(resp.data[0].embedding) == 256
print(f"[PASS] custom dimensions -> dim={len(resp.data[0].embedding)}")

# base64 encoding
resp = client.embeddings.create(model=MODEL, input="base64 test", encoding_format="base64")
assert isinstance(resp.data[0].embedding, str)
print("[PASS] base64 encoding")

# Model listed
assert MODEL in [m.id for m in client.models.list().data]
print("[PASS] model listed in /models")

[zxkane]

Thanks for adding Nova 2 multimodal embeddings support! The overall structure follows the existing patterns well. A few issues to address before merge:

Critical

1. VALID_DIMENSIONS values are incorrect per AWS documentation

According to the Nova 2 embeddings docs and the embeddings schema reference, the allowed embeddingDimension values are 256, 384, 1024, 3072.

The current code {256, 512, 1024, 2048, 3072} includes two invalid values (512, 2048) and omits one valid value (384).

# Fix:
VALID_DIMENSIONS = {256, 384, 1024, 3072}

2. str(item) fallback silently embeds garbage

else:
    texts.append(str(item))

If an unexpected type reaches this branch, it silently converts to its string representation (e.g., "{'key': 'value'}") and returns HTTP 200 with meaningless embeddings. This is worse than an error — users won't know their results are wrong. Should raise an error instead:

else:
    raise HTTPException(
        status_code=400,
        detail=f"Unsupported input item type: {type(item).__name__}. Expected str, int, or list of ints.",
    )

Important

3. Stale comment in schema.py:185

The dimensions field comment says # not used. but this PR makes it actively functional for Nova. Please update it, e.g.:

dimensions: int | None = None  # Used by Nova embeddings; ignored by other models.

4. getattr is unnecessary for a Pydantic model field

dimensions = getattr(embeddings_request, "dimensions", None)

EmbeddingsRequest defines dimensions: int | None = None as a Pydantic field. Direct access works and is consistent with how all other attributes are accessed in this codebase:

dimensions = embeddings_request.dimensions

5. Batch failure loses context

Since Nova processes each text as a separate API call, a failure mid-batch gives no indication of which item failed. Consider using enumerate and including the index in error context:

for idx, text in enumerate(texts):
    # ... on error, include idx in the error detail

Suggestions (optional)

• Move dimension validation before the loop (it's constant across texts but validated per-text currently)
• Add a comment explaining why embeddingPurpose is hardcoded to "GENERIC_INDEX" (Nova supports 9 different purposes)
• Consider using isinstance(item, list) instead of isinstance(item, Iterable) for more precise type matching

[zxkane]

Thanks for the contribution @gabrielkoo! 🎉 Great work adding Nova 2 multimodal embeddings support. The implementation is clean, well-documented, and consistent with the existing codebase. Merging now!

[pgodzin]

@zxkane @gabrielkoo note that this still only supports text embeddings, despite being a multimodal embedding model that supports image and video input.

see #159