Gemini embeddings

[ferdinandl007]

Description

Enable Gemini embedding models

How Has This Been Tested?

Manual test and unit test

Backporting (check the box to trigger backport action)

Note: You have to check that the action passes, otherwise resolve the conflicts manually and tag the patches.

• This PR should be backported (make sure to check that the backport attempt succeeds)
• [Optional] Override Linear Check

[vercel]

Someone is attempting to deploy a commit to the Danswer Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

PR Summary

This PR introduces Google's Gemini embedding models and updates the embedding infrastructure to use Google's Generative AI API instead of Vertex AI.

• Added gemini-embedding-001 model in /web/src/components/embedding/interfaces.tsx with 768 dimensions and $0.15/million pricing
• Replaced google-cloud-aiplatform with google-genai==1.15.0 in /backend/requirements/model_server.txt for Gemini support
• Modified /backend/model_server/encoders.py to use genai.Client with special batch size handling for Gemini models
• Added environment variables VERTEXAI_EMBEDDING_MODEL_LOCATION and VERTEXAI_EMBEDDING_MODEL_DIMENSION in /backend/shared_configs/configs.py
• Potential issue: Duplicate model entries in SupportedEmbeddingModel list need to be cleaned up

_{4 file(s) reviewed, 6 comment(s)
Edit PR Review Bot Settings | Greptile}

[greptile-apps]

style: Description makes unsubstantiated claim about being 'Worlds best performing embedding model'. Consider using more specific, verifiable performance claims.

Suggested change

	description: "Google's most recent text gemini embedding model. Worlds best performing embedding model",
	description: "Google's most recent text gemini embedding model. Offers strong performance across a wide range of tasks.",

[greptile-apps]

logic: Duplicate model entries for 'google/gemini-embedding-001' with different index names. This will cause confusion and potential issues. Keep only one entry with a consistent index name pattern.

Suggested change

	name="google/gemini-embedding-001",
	dim=768,
	index_name="danswer_chunk_google_gemini_embedding_001",
	),
	SupportedEmbeddingModel(
	name="google/gemini-embedding-001",
	dim=768,
	index_name="danswer_chunk_gemini_embedding_001",
	),
	name="google/gemini-embedding-001",
	dim=768,
	index_name="danswer_chunk_gemini_embedding_001",
	),

[greptile-apps]

logic: Duplicate model entries for 'google/text-multilingual-embedding-002' with different index names. Similar to the Gemini entries, this needs to be consolidated.

Suggested change

	name="google/text-multilingual-embedding-002",
	dim=768,
	index_name="danswer_chunk_google_multilingual_embedding_002",
	),
	SupportedEmbeddingModel(
	name="google/text-multilingual-embedding-002",
	dim=768,
	index_name="danswer_chunk_multilingual_embedding_002",
	),
	name="google/text-multilingual-embedding-002",
	dim=768,
	index_name="danswer_chunk_multilingual_embedding_002",
	),

[greptile-apps]

style: Gemini models require batch size of 1, which could severely impact performance. Consider adding a warning log when using Gemini models to inform users about potential performance implications.

[greptile-apps]

logic: embed_batch function could fail silently if embeddings.embeddings is None or malformed. Add explicit error checking.

Suggested change

	async def embed_batch(batch: list[str]) -> list[list[float]]:
	embeddings = await client.aio.models.embed_content(
	model=model,
	contents=batch,
	config=EmbedContentConfig(
	task_type=embedding_type,
	output_dimensionality=VERTEXAI_EMBEDDING_MODEL_DIMENSION
	)
	)
	return embeddings.embeddings
	async def embed_batch(batch: list[str]) -> list[list[float]]:
	embeddings = await client.aio.models.embed_content(
	model=model,
	contents=batch,
	config=EmbedContentConfig(
	task_type=embedding_type,
	output_dimensionality=VERTEXAI_EMBEDDING_MODEL_DIMENSION
	)
	)
	if not embeddings or not embeddings.embeddings:
	raise ValueError(f"Failed to get embeddings for batch of size {len(batch)}")
	return embeddings.embeddings

[greptile-apps]

style: Client initialization should be wrapped in try/except to handle invalid credentials or connection errors gracefully.

Suggested change

	client = genai.Client(
	vertexai=True,
	project=project_id,
	location=VERTEXAI_EMBEDDING_MODEL_LOCATION,
	credentials=credentials
	)
	try:
	client = genai.Client(
	vertexai=True,
	project=project_id,
	location=VERTEXAI_EMBEDDING_MODEL_LOCATION,
	credentials=credentials
	)
	except Exception as e:
	raise RuntimeError(f"Failed to initialize Vertex AI client: {e}") from e

[github-actions]

This PR is stale because it has been open 75 days with no activity. Remove stale label or comment or this will be closed in 15 days.

[github-actions]

This PR was closed because it has been stalled for 90 days with no activity.