AgentSkillsCN

embeddings

向量嵌入配置与语义搜索

SKILL.md
--- frontmatter
name: embeddings
description: "Vector embeddings configuration and semantic search"
emoji: "🧬"

Embeddings - Complete API Reference

Configure embedding providers, manage vector storage, and perform semantic search.


Chat Commands

View Config

code
/embeddings                                 Show current settings
/embeddings status                          Provider status
/embeddings stats                           Cache statistics

Configure Provider

code
/embeddings provider openai                 Use OpenAI embeddings
/embeddings provider voyage                 Use Voyage AI
/embeddings provider local                  Use local model
/embeddings model text-embedding-3-small    Set model

Cache Management

code
/embeddings cache stats                     View cache stats
/embeddings cache clear                     Clear cache
/embeddings cache size                      Total cache size

Testing

code
/embeddings test "sample text"              Generate test embedding
/embeddings similarity "text1" "text2"      Compare similarity

TypeScript API Reference

Create Embeddings Service

typescript
import { createEmbeddingsService } from 'clodds/embeddings';

const embeddings = createEmbeddingsService({
  // Provider
  provider: 'openai',  // 'openai' | 'voyage' | 'local' | 'cohere'
  apiKey: process.env.OPENAI_API_KEY,

  // Model
  model: 'text-embedding-3-small',
  dimensions: 1536,

  // Caching
  cache: true,
  cacheBackend: 'sqlite',
  cachePath: './embeddings-cache.db',

  // Batching
  batchSize: 100,
  maxConcurrent: 5,
});

Generate Embeddings

typescript
// Single text
const embedding = await embeddings.embed('Hello world');
console.log(`Dimensions: ${embedding.length}`);

// Multiple texts (batched)
const vectors = await embeddings.embedBatch([
  'First document',
  'Second document',
  'Third document',
]);

Semantic Search

typescript
// Search against stored vectors
const results = await embeddings.search({
  query: 'trading strategies',
  collection: 'documents',
  limit: 10,
  threshold: 0.7,
});

for (const result of results) {
  console.log(`${result.text} (score: ${result.score})`);
}

Similarity

typescript
// Compare two texts
const score = await embeddings.similarity(
  'The cat sat on the mat',
  'A feline rested on the rug'
);

console.log(`Similarity: ${score}`);  // 0.0 - 1.0

Store Vectors

typescript
// Store embedding with metadata
await embeddings.store({
  collection: 'documents',
  id: 'doc-1',
  text: 'Original text',
  embedding: vector,
  metadata: {
    source: 'wiki',
    date: '2024-01-01',
  },
});

// Store batch
await embeddings.storeBatch({
  collection: 'documents',
  items: [
    { id: 'doc-1', text: 'First doc' },
    { id: 'doc-2', text: 'Second doc' },
  ],
});

Cache Management

typescript
// Get cache stats
const stats = await embeddings.getCacheStats();
console.log(`Cached: ${stats.count} embeddings`);
console.log(`Size: ${stats.sizeMB} MB`);
console.log(`Hit rate: ${stats.hitRate}%`);

// Clear cache
await embeddings.clearCache();

// Clear specific entries
await embeddings.clearCache({ olderThan: '7d' });

Provider Configuration

typescript
// Switch provider
embeddings.setProvider('voyage', {
  apiKey: process.env.VOYAGE_API_KEY,
  model: 'voyage-large-2',
});

// Use local model (Transformers.js)
// No API key required - runs locally via @xenova/transformers
embeddings.setProvider('local', {
  model: 'Xenova/all-MiniLM-L6-v2',  // 384 dimensions
});

Providers

ProviderModelsQualitySpeedCost
OpenAItext-embedding-3-small/largeExcellentFast$0.02/1M
Voyagevoyage-large-2ExcellentFast$0.02/1M
Cohereembed-english-v3GoodFast$0.10/1M
Local (Transformers.js)Xenova/all-MiniLM-L6-v2GoodMediumFree

Models

OpenAI

ModelDimensionsBest For
text-embedding-3-small1536General use
text-embedding-3-large3072High accuracy

Voyage

ModelDimensionsBest For
voyage-large-21024General use
voyage-code-21536Code search

Use Cases

Semantic Memory Search

typescript
// Store user memories
await embeddings.store({
  collection: 'memories',
  id: 'mem-1',
  text: 'User prefers conservative trading',
});

// Search memories
const relevant = await embeddings.search({
  query: 'what is user risk preference',
  collection: 'memories',
  limit: 5,
});

Document Similarity

typescript
// Find similar documents
const similar = await embeddings.findSimilar({
  text: 'How to trade options',
  collection: 'docs',
  limit: 5,
});

Best Practices

  1. Use caching — Avoid redundant API calls
  2. Batch requests — More efficient than single calls
  3. Choose dimensions wisely — Balance quality vs storage
  4. Monitor costs — Embeddings can add up
  5. Local for development — Use local model to save costs