Gemini Embed

Gemini Embeddings API

Generate text embeddings using Google Gemini API via REST.

Prerequisites

• Environment variable GOOGLE_API_KEY must be set
• API endpoint: https://generativelanguage.googleapis.com/v1beta
• Model: gemini-embedding-001

Workflow

Phase 1: Determine Embedding Type

• Single Embedding: For one text input
• Batch Embedding: For multiple texts (more efficient)

Phase 2: Configure Task Type (Optional)

Choose based on use case:

• RETRIEVAL_QUERY: For search queries
• RETRIEVAL_DOCUMENT: For documents to be searched
• SEMANTIC_SIMILARITY: For comparing text similarity
• CLASSIFICATION: For text classification
• CLUSTERING: For grouping similar texts

Phase 3: Execute API Call

1. Single Text Embedding

Basic Embedding

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:embedContent?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "models/gemini-embedding-001",
      "content": {
        "parts": [{"text": "Hello world"}]
      }
    }'

With Task Type

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:embedContent?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "models/gemini-embedding-001",
      "content": {
        "parts": [{"text": "What is machine learning?"}]
      },
      "task_type": "RETRIEVAL_QUERY"
    }'

With Output Dimensionality Control

Truncate embeddings to a smaller size for efficiency:

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:embedContent?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "models/gemini-embedding-001",
      "content": {
        "parts": [{"text": "Hello world"}]
      },
      "output_dimensionality": 256
    }'

2. Batch Embedding

Process multiple texts in a single API call:

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:batchEmbedContents?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "requests": [
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "What is the meaning of life?"}]}
        },
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "How does the brain work?"}]}
        },
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "What is quantum computing?"}]}
        }
      ]
    }'

Batch with Task Type

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:batchEmbedContents?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "requests": [
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "Document about AI"}]},
          "task_type": "RETRIEVAL_DOCUMENT"
        },
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "Document about ML"}]},
          "task_type": "RETRIEVAL_DOCUMENT"
        }
      ]
    }'

Response Structure

Single Embedding Response

{
  "embedding": {
    "values": [
      -0.02342152,
      0.01676572,
      0.009261323,
      ...
    ]
  }
}

Batch Embedding Response

{
  "embeddings": [
    {
      "values": [-0.022374554, -0.004560777, ...]
    },
    {
      "values": [-0.007975887, -0.02141119, ...]
    },
    {
      "values": [-0.0047850125, 0.008764064, ...]
    }
  ]
}

Task Types Reference

Common Use Cases

RAG (Retrieval-Augmented Generation)

1. Embed documents with RETRIEVAL_DOCUMENT
2. Store embeddings in vector database
3. Embed user query with RETRIEVAL_QUERY
4. Find similar documents by vector similarity

Semantic Search

# Embed the query
curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:embedContent?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "models/gemini-embedding-001",
      "content": {"parts": [{"text": "How to train a neural network?"}]},
      "task_type": "RETRIEVAL_QUERY"
    }'

Text Similarity Comparison

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-embedding-001:batchEmbedContents?key=$GOOGLE_API_KEY" \
    -H 'Content-Type: application/json' \
    -d '{
      "requests": [
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "The cat sat on the mat"}]},
          "task_type": "SEMANTIC_SIMILARITY"
        },
        {
          "model": "models/gemini-embedding-001",
          "content": {"parts": [{"text": "A feline rested on the rug"}]},
          "task_type": "SEMANTIC_SIMILARITY"
        }
      ]
    }'

Then compute cosine similarity between the two embedding vectors.

Best Practices

1. Use batch embedding: More efficient for multiple texts
2. Set task_type: Improves embedding quality for specific use cases
3. Reduce dimensionality: Use output_dimensionality for smaller, faster embeddings when full precision isn't needed
4. Match task types: Use same task_type for queries and documents in RAG
5. Normalize vectors: Cosine similarity works best with normalized vectors