@ainative/skill-zerodb-workflows

Official AINative Studio skill for ZeroDB vector database workflows, semantic search patterns, RLHF feedback collection, and AI memory management.

Installation

npm install @ainative/skill-zerodb-workflows

Or install directly in your AINative Studio skills directory:

cd ~/.ainative/skills
git clone https://github.com/AINative-Studio/ainative-skills

What This Skill Covers

This skill provides comprehensive patterns and best practices for:

• Vector Database Operations: Storing, searching, and managing embeddings at scale
• Semantic Search: Hybrid search combining vector similarity and metadata filters
• Memory Management: Context window optimization for AI agents
• RLHF Workflows: Collecting and analyzing human feedback for model improvement
• Performance Optimization: Caching, batching, and query optimization strategies

Quick Start

1. Basic Vector Storage and Search

import { ZeroDBClient } from '@zerodb/client';
import { getEmbedding } from './embeddings';

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

// Store a document with semantic embedding
await client.vector.upsert({
  id: 'doc_1',
  embedding: await getEmbedding('How to implement OAuth authentication'),
  metadata: {
    title: 'OAuth Guide',
    category: 'security',
    timestamp: Date.now()
  }
});

// Search for similar documents
const results = await client.vector.search({
  embedding: await getEmbedding('authentication best practices'),
  topK: 5,
  filters: { category: 'security' }
});

2. Conversation Memory Management

import { ConversationMemory } from '@ainative/skill-zerodb-workflows/memory';

const memory = new ConversationMemory(process.env.ZERODB_API_KEY!);

// Store conversation turn
await memory.storeTurn('session_123', 'user_456', {
  role: 'user',
  content: 'How do I optimize database queries?',
  timestamp: Date.now()
});

// Retrieve recent context
const context = await memory.getRecentContext('session_123', 10);

3. RLHF Feedback Collection

import { FeedbackCollector } from '@ainative/skill-zerodb-workflows/rlhf';

const collector = new FeedbackCollector(process.env.ZERODB_API_KEY!);

// Submit user rating
await collector.submitFeedback({
  prompt_id: 'prompt_123',
  response_id: 'resp_456',
  user_id: 'user_789',
  rating: 4,
  feedback_type: 'quality',
  timestamp: Date.now()
});

// Analyze feedback trends
const stats = await collector.getFeedbackStats('resp_456');
console.log(`Avg rating: ${stats.avg_rating}`);

Skill Structure

zerodb-workflows/
├── SKILL.md                           # Main skill file with quick reference
├── references/
│   ├── api-endpoints.md              # Complete ZeroDB API documentation
│   ├── vector-search.md              # Advanced search patterns
│   ├── memory-management.md          # Context optimization strategies
│   └── rlhf-workflows.md             # Feedback collection patterns
├── package.json                       # NPM package configuration
└── README.md                          # This file

Reference Documentation

API Endpoints (references/api-endpoints.md)

• Authentication and client setup
• Vector operations (upsert, search, delete)
• Metadata filtering syntax
• Table operations for structured data
• File storage and retrieval
• Error handling and retry patterns

Vector Search (references/vector-search.md)

• Semantic search fundamentals
• Hybrid search (vector + metadata)
• Multi-table search and result merging
• Search result reranking strategies
• Performance optimization (caching, batching)
• Quality monitoring and analytics

Memory Management (references/memory-management.md)

• Short-term conversation context
• Long-term knowledge retention
• Context window optimization
• Token-aware context building
• Memory pruning strategies
• Multi-session management

RLHF Workflows (references/rlhf-workflows.md)

• Simple rating systems
• Comparative feedback (pairwise comparisons)
• Multi-dimensional feedback analysis
• Implicit feedback (behavioral signals)
• Training dataset construction
• Quality control best practices

Use Cases

1. Building a RAG System

// Store your knowledge base
for (const doc of documents) {
  await client.vector.upsert({
    id: doc.id,
    embedding: await getEmbedding(doc.content),
    metadata: { title: doc.title, category: doc.category }
  });
}

// Retrieve relevant context for user query
const context = await client.vector.search({
  embedding: await getEmbedding(userQuery),
  topK: 5
});

2. AI Agent Memory

// Store agent observations
await memory.storeTurn(sessionId, userId, {
  role: 'assistant',
  content: 'I noticed you prefer TypeScript for backend work',
  timestamp: Date.now()
});

// Retrieve relevant memories for next interaction
const relevantMemories = await memory.getRelevantContext(
  'What language should I use?',
  userId,
  { topK: 3 }
);

3. Model Improvement Pipeline

// Collect feedback
await collector.submitFeedback({...});

// Analyze patterns
const problems = await analyzer.identifyProblemPatterns();

// Build training dataset
const dataset = await builder.buildDataset({
  minRating: 4.0,
  limit: 5000
});

await builder.exportToJSONL(dataset, './training.jsonl');

Best Practices

Performance

• Use batch operations for inserting multiple vectors
• Implement embedding caching to reduce API calls
• Set appropriate topK values (5-20 typical)
• Monitor search latency and optimize queries

Data Quality

• Always include rich metadata for hybrid search
• Store conversation turns immediately (don't batch)
• Implement memory pruning for old/irrelevant data
• Use minimum similarity scores to filter poor matches

Security

• Never store API keys in metadata
• Implement proper authentication and authorization
• Anonymize sensitive user data
• Follow data retention policies

Monitoring

• Track search quality metrics (avg score, result count)
• Monitor engagement signals (copy rate, regeneration rate)
• Analyze feedback trends over time
• A/B test changes to measure impact

Requirements

• Node.js >= 18.0.0
• ZeroDB account and API key
• Embedding model (OpenAI, Anthropic, or local)

Environment Setup

export ZERODB_API_KEY="your_api_key_here"
export EMBEDDING_MODEL="text-embedding-3-small" # or your preferred model

TypeScript Support

This skill includes full TypeScript type definitions for all patterns and examples.

import type {
  VectorUpsertRequest,
  VectorSearchRequest,
  FeedbackData,
  ConversationTurn
} from '@ainative/skill-zerodb-workflows';

Contributing

Found a bug or have a pattern to share? Open an issue or PR at: https://github.com/AINative-Studio/ainative-skills

License

Apache-2.0

Support

• Documentation: https://docs.zerodb.ai
• AINative Studio: https://ainative.studio
• Discord: https://discord.gg/ainative

Related Skills

• @ainative/skill-api-design - RESTful API patterns
• @ainative/skill-typescript-backend - Backend architecture
• @ainative/skill-testing-patterns - Testing strategies

Made with ❤️ by AINative Studio