Vector Storage & RAG

YH-UI provides a complete vector storage abstraction layer and production-ready RAG support.

createInMemoryVectorStore - In-Memory Vector Storage

Suitable for development and small-scale scenarios, uses cosine similarity for retrieval.

import { createInMemoryVectorStore } from '@yh-ui/ai-sdk'

const vectorStore = createInMemoryVectorStore()

// Add document vectors
await vectorStore.add([
  {
    id: 'doc-1',
    content: 'Vue 3 is a progressive JavaScript framework',
    embedding: [0.1, 0.2, 0.3 /* ... */],
    metadata: { source: 'vue-docs' }
  }
])

// Similarity search
const results = await vectorStore.search([0.1, 0.2, 0.3], {
  topK: 5,
  threshold: 0.7
})

console.log(results[0].content)

API

Method	Description
add(documents, namespace?)	Add document vectors
search(queryEmbedding, options)	Similarity search
delete(ids, namespace?)	Delete by IDs
clear(namespace?)	Clear namespace

External Vector Storage Interfaces

YH-UI provides configuration types for Pinecone, Weaviate, and Qdrant. You can implement your own adapters.

import type {
  VectorStoreConfig,
  PineconeVectorStoreConfig,
  WeaviateVectorStoreConfig,
  QdrantVectorStoreConfig
} from '@yh-ui/ai-sdk'

// Pinecone configuration example
const pineconeConfig: PineconeVectorStoreConfig = {
  apiKey: 'your-api-key',
  indexName: 'your-index',
  environment: 'us-east1-aws'
}

createProductionRAG - Production-Ready RAG

Build complete RAG pipeline: document loading, vectorization, storage, retrieval, and generation.

import { createProductionRAG, createInMemoryVectorStore } from '@yh-ui/ai-sdk'

// Create RAG system
const rag = createProductionRAG({
  namespace: 'my-knowledge-base',
  topK: 5,
  similarityThreshold: 0.5,
  includeSources: true,
  vectorStore: createInMemoryVectorStore(),
  embed: async (text) => {
    // Call embedding API to get vectors
    const response = await fetch('/api/embed', {
      method: 'POST',
      body: JSON.stringify({ text })
    })
    return response.json()
  }
})

// Add documents (auto-vectorization)
await rag.addDocuments([
  { content: 'Vue 3 core features include reactive system', metadata: { source: 'vue' } },
  { content: 'Composition API is a new feature of Vue 3', metadata: { source: 'vue' } }
])

// Query
const result = await rag.query('What are the new features of Vue 3?', async (prompt) => {
  const response = await fetch('/api/chat', {
    method: 'POST',
    body: JSON.stringify({ prompt })
  })
  return response.json()
})

console.log(result.answer) // LLM generated answer
console.log(result.sources) // Referenced sources

API

Method	Description
addDocuments(documents)	Add documents, auto-vectorize
retrieve(query, k?)	Retrieve documents only
query(question, llm)	Retrieve + generate answer
clear()	Clear knowledge base

Return Values

interface RAGResult {
  answer: string // LLM generated answer
  sources: Array<{
    // Referenced sources
    content: string
    metadata: Record<string, unknown>
    score: number
  }>
  contextUsed: string // Context used (for debugging)
}

TIP

Use with @yh-ui/ai-sdk's Document Loaders for better results.

Important Notes

1. The embed function needs to be implemented by yourself or call embedding APIs (e.g., OpenAI Embedding, Azure OpenAI Embedding)
2. For production, it is recommended to use cloud vector stores like Pinecone, Weaviate
3. Set similarityThreshold appropriately; too high may result in no retrieval results