gemini-embeddings（Gemini embeddings 搜索）

Gemini Embeddings

Generate high-quality text embeddings for semantic search, similarity analysis, clustering, and RAG (Retrieval Augmented Generation) applications through executable scripts.

When to Use This Skill

Use this skill when you need to:

• Find semantically similar documents or texts
• Build semantic search engines
• Implement RAG (Retrieval Augmented Generation)
• Cluster or group similar documents
• Calculate text similarity scores
• Power recommendation systems
• Enable semantic document retrieval
• Create vector databases for AI applications

Available Scripts

scripts/embed.js

Purpose: Generate embeddings and calculate similarity

When to use:

• Creating vector representations of text
• Comparing text similarity
• Building semantic search systems
• Implementing RAG pipelines
• Clustering documents

Key parameters:

Output: Embedding vectors or similarity scores

Workflows

Workflow 1: Single Text Embedding

node scripts/embed.js "What is the meaning of life?"

• Best for: Basic embedding generation
• Output: Vector with 3072 dimensions (default)
• Use when: Storing single document vectors

Workflow 2: Semantic Search

# 1. Generate embedding for query
node scripts/embed.js "best practices for coding" --task RETRIEVAL_QUERY > query.json

# 2. Generate embeddings for documents (batch)
node scripts/embed.js "Coding best practices include version control" "Clean code is essential" --task RETRIEVAL_DOCUMENT > docs.json

# 3. Compare and find most similar (calculate similarity separately)

• Best for: Building search functionality
• Task types: RETRIEVAL_QUERY, RETRIEVAL_DOCUMENT
• Combines with: Similarity calculation for ranking

Workflow 3: Text Similarity Comparison

node scripts/embed.js "What is the meaning of life?" "What is the purpose of existence?" "How do I bake a cake?" --similarity

• Best for: Comparing multiple texts, finding duplicates
• Output: Pairwise similarity scores (0-1)
• Use when: Need to rank text similarity

Workflow 4: Dimensionality Reduction for Efficiency

node scripts/embed.js "Text to embed" --dim 768

• Best for: Faster storage and comparison
• Options: 768, 1536, or 3072 (default)
• Trade-off: Lower dimensions = less accuracy but faster

Workflow 5: Document Clustering

# 1. Generate embeddings for multiple documents
node scripts/embed.js "Machine learning is AI" "Deep learning is a subset" "Neural networks power AI" --json > embeddings.jsonl

# 2. Process embeddings with clustering algorithm (your code)
# Use scikit-learn, KMeans, etc.

• Best for: Grouping similar documents, topic discovery
• Task type: CLUSTERING
• Combines with: Clustering libraries (scikit-learn)

Workflow 6: RAG Implementation

# 1. Create document embeddings (one-time setup)
node scripts/embed.js "Document 1 content" "Document 2 content" --task RETRIEVAL_DOCUMENT --dim 1536

# 2. For each query, find similar documents
node scripts/embed.js "User query here" --task RETRIEVAL_QUERY

# 3. Use retrieved documents in prompt to LLM (gemini-text)
node skills/gemini-text/scripts/generate.js "Context: [retrieved docs]. Answer: [user query]"

• Best for: Building knowledge-based AI systems
• Combines with: gemini-text for generation with context

Workflow 7: JSON Output for API Integration

node scripts/embed.js "Text to process" --json

• Best for: API responses, database storage
• Output: JSON array of embedding vectors
• Use when: Programmatic processing required

Workflow 8: Batch Document Processing

# 1. Create JSONL with documents
echo '{"text": "Document 1"}' > docs.jsonl
echo '{"text": "Document 2"}' >> docs.jsonl

# 2. Process with script or custom code
python3 << 'EOF'
import json
from google import genai

client = genai.Client()

texts = []
with open("docs.jsonl") as f:
    for line in f:
        texts.append(json.loads(line)["text"])

response = client.models.embed_content(
    model="gemini-embedding-001",
    contents=texts,
    task_type="RETRIEVAL_DOCUMENT"
)

embeddings = [e.values for e in response.embeddings]
print(f"Generated {len(embeddings)} embeddings")
EOF

• Best for: Large document collections
• Combines with: Vector databases (Pinecone, Weaviate)

Parameters Reference

Task Types

Dimensionality Options

Similarity Scores

Output Interpretation

Embedding Vector

• Format: List of float values (768, 1536, or 3072)
• Range: Typically -1.0 to 1.0
• Normalized for cosine similarity
• Can be stored in vector databases

Similarity Output

Pairwise Similarity:
  'What is the meaning of life?...' <-> 'What is the purpose of existence?...': 0.8742
  'What is the meaning of life?...' <-> 'How do I bake a cake?...': 0.1234

• Higher scores = more similar
• Use threshold (e.g., 0.7) for matching

JSON Output

[[0.123, -0.456, 0.789, ...], [0.234, -0.567, 0.890, ...]]

• Array of embedding vectors
• One per input text
• Ready for database storage

Common Issues

"google-genai not installed"

npm install @google/genai@latest dotenv@latest

"numpy not installed" (for similarity)

pip install numpy

"Invalid task type"

• Use available tasks: SEMANTIC_SIMILARITY, RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, CLASSIFICATION, CLUSTERING
• Check spelling (case-sensitive)
• Use correct task for your use case

"Invalid dimension"

• Options: 768, 1536, or 3072 only
• Check model supports requested dimension
• Default to 3072 if unsure

"No similarity calculated"

• Need multiple texts for similarity comparison
• Use --similarity flag
• Check that at least 2 texts provided

"Embedding size mismatch"

• All embeddings must have same dimensionality
• Use consistent --dim parameter
• Recompute if dimensions differ

Best Practices

Task Selection

• SEMANTIC_SIMILARITY: General text comparison
• RETRIEVAL_DOCUMENT: Storing documents for search
• RETRIEVAL_QUERY: Querying for similar documents
• CLASSIFICATION: Categorization tasks
• CLUSTERING: Grouping similar content

Dimensionality Choice

• 768: Real-time applications, high volume
• 1536: Balanced choice for most use cases
• 3072: Maximum accuracy, offline processing

Performance Optimization

• Use lower dimensions for speed
• Batch multiple texts in one request
• Cache embeddings for repeated queries
• Precompute document embeddings for search

Storage Tips

• Use vector databases (Pinecone, Weaviate, Chroma)
• Normalize vectors for consistent comparison
• Store metadata with embeddings
• Index for fast retrieval

RAG Implementation

• Precompute document embeddings
• Use RETRIEVAL_DOCUMENT for docs
• Use RETRIEVAL_QUERY for user questions
• Combine top results with gemini-text

Similarity Thresholds

• 0.9+: Exact duplicates or near-duplicates
• 0.7-0.9: Same topic/subject
• 0.5-0.7: Related concepts
• <0.5: Different topics

Related Skills

• gemini-text: Generate text with retrieved context (RAG)
• gemini-batch: Process embeddings in bulk
• gemini-files: Upload documents for embedding
• gemini-search: Implement semantic search (if available)

Quick Reference

# Basic embedding
node scripts/embed.js "Your text here"

# Semantic search
node scripts/embed.js "Query" --task RETRIEVAL_QUERY

# Document embedding
node scripts/embed.js "Document text" --task RETRIEVAL_DOCUMENT

# Similarity comparison
node scripts/embed.js "Text 1" "Text 2" "Text 3" --similarity

# Dimensionality reduction
node scripts/embed.js "Text" --dim 768

# JSON output
node scripts/embed.js "Text" --json

Reference

• Get API key: https://aistudio.google.com/apikey
• Documentation: https://ai.google.dev/gemini-api/docs/embeddings
• Vector databases: Pinecone, Weaviate, Chroma, Qdrant
• Cosine similarity: Standard for embedding comparison