zerodb-workflows

ZeroDB矢量数据库最佳实践、语义搜索模式、RLHF工作流程和内存管理。适用于使用ZeroDB API、矢量搜索或AI内存系统时使用。

📖 中文导读:ZeroDB矢量数据库最佳实践、语义搜索模式、RLHF工作流程和内存管理。适用于使用ZeroDB API、矢量搜索或AI内存系统时使用。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: zerodb-workflows description: ZeroDB vector database best practices, semantic search patterns, RLHF workflows, and memory management. Use when working with ZeroDB APIs, vector search, or AI memory systems. version: 1.0.0 author: AINative Studio license: Apache-2.0 keywords: - zerodb - vector-database - semantic-search - rlhf - memory-management category: database

ZeroDB Workflows & Best Practices

This skill provides patterns and best practices for working with ZeroDB, AINative's vector database system for AI memory, semantic search, and RLHF workflows.

When to Use This Skill

  • Creating ZeroDB projects or tables
  • Implementing vector search functionality
  • Managing AI agent memory and conversation context
  • Collecting RLHF feedback data for model improvement
  • Optimizing semantic similarity queries
  • Debugging vector search results and relevance
  • Building RAG (Retrieval Augmented Generation) systems
  • Storing and retrieving embeddings at scale

Core Concepts

Vector Storage

ZeroDB stores high-dimensional embeddings (384, 768, 1024, 1536 dimensions) for semantic search and similarity matching. Each vector includes:

  • Embedding: Dense vector representation of text/data
  • Metadata: Arbitrary JSON data for filtering and context
  • ID: Unique identifier for retrieval and updates

Memory Management

Efficient context window management for AI agents using vector similarity to retrieve relevant conversation history. Key patterns:

  • Store conversation turns as vectors with metadata (timestamp, user_id, session_id)
  • Search by semantic similarity to find relevant context
  • Prune old/irrelevant memories to maintain context quality
  • Use hybrid search (vector + metadata filters) for precise retrieval

RLHF Workflows

Collect human feedback on AI responses for model improvement and fine-tuning:

  • Store prompt-response pairs with feedback ratings
  • Track improvement metrics over time
  • Identify failure patterns for targeted training
  • Build datasets for reinforcement learning

Quick Start Examples

1. Vector Upsert with Metadata

import { ZeroDBClient } from '@zerodb/client';

const client = new ZeroDBClient({ apiKey: process.env.ZERODB_API_KEY });

// Store conversation memory
await client.vector.upsert({
  id: 'msg_12345',
  embedding: await getEmbedding('User asked about authentication'),
  metadata: {
    type: 'conversation',
    user_id: 'user_123',
    session_id: 'session_abc',
    timestamp: Date.now(),
    content: 'User asked about authentication',
    role: 'user'
  }
});

2. Semantic Search with Filters

// Find relevant conversation history
const results = await client.vector.search({
  embedding: await getEmbedding('How do I implement OAuth?'),
  topK: 5,
  filters: {
    user_id: 'user_123',
    type: 'conversation',
    timestamp: { $gt: Date.now() - 86400000 } // Last 24 hours
  }
});

// Build context for AI prompt
const context = results.map(r => r.metadata.content).join('\n');

3. RLHF Feedback Collection

// Store AI response with feedback tracking
await client.rlhf.feedback({
  prompt_id: 'prompt_123',
  response_id: 'resp_456',
  rating: 4, // 1-5 scale
  feedback_type: 'quality',
  metadata: {
    model: 'claude-3-sonnet',
    latency_ms: 1250,
    prompt_tokens: 1024,
    completion_tokens: 512,
    user_comment: 'Good response but could be more concise'
  }
});

Architecture Patterns

Memory-First Design

Always consider:

  1. What information needs to be retrieved later?
  2. How will you search for it (semantic, metadata, hybrid)?
  3. What metadata is needed for filtering?
  4. How long should memories persist?

Search Quality

Optimize for relevance:

  • Use meaningful embeddings (not just keywords)
  • Include rich metadata for hybrid search
  • Experiment with topK values (5-20 typical)
  • Monitor search latency and quality metrics

Scalability

Plan for growth:

  • Batch operations when inserting multiple vectors
  • Use pagination for large result sets
  • Implement caching for frequently accessed data
  • Monitor vector count and storage usage

Common Pitfalls

❌ Storing vectors without metadata - Makes filtering impossible ✅ Store rich metadata for every vector

❌ Using too few search results (topK=1) - Misses relevant context ✅ Use topK=5-10 and rerank if needed

❌ Ignoring embedding dimensions - Different models need different dimensions ✅ Match embedding model output to ZeroDB dimension config

❌ Not handling search errors - Network/API failures happen ✅ Implement retry logic and fallbacks

Reference Files

See the references/ directory for detailed patterns:

  • api-endpoints.md - Complete ZeroDB API reference with examples
  • vector-search.md - Advanced search query patterns and optimization
  • memory-management.md - Context window optimization strategies
  • rlhf-workflows.md - Feedback collection and analysis patterns

Best Practices Checklist

  • All vectors include meaningful metadata
  • Search queries use appropriate topK values
  • Error handling implemented for API calls
  • Batch operations used for multiple inserts
  • Memory pruning strategy defined
  • Search quality metrics monitored
  • RLHF feedback includes model/prompt metadata
  • Embedding dimensions match model output
  • @ainative/skill-api-design - RESTful API patterns
  • @ainative/skill-typescript-backend - TypeScript service architecture
  • @ainative/skill-testing-patterns - Testing database integrations

Support