Qdrant Vector Database Skill

Quick Start Workflow

When working with Qdrant:

1. Check if Qdrant is configured

- Look for QDRANTURL and QDRANTAPI_KEY in .env - For local: http://localhost:6333 - For cloud: https://xxx.qdrant.io

1. For collection creation

- Define vector size (1536 for OpenAI ada-002) - Choose distance metric (Cosine for semantic similarity) - Set up HNSW parameters for performance

1. For content ingestion

- Chunk text into 800-character segments with 200-char overlap - Generate embeddings with OpenAI text-embedding-ada-002 - Upsert vectors with metadata (chapter, section, file path)

1. For semantic search

- Convert user query to embedding - Search with score threshold (>= 0.7 for relevance) - Return top 5 results with metadata

Standard Patterns

Client Setup

import { QdrantClient } from '@qdrant/js-client';

export const qdrant = new QdrantClient({
  url: process.env.QDRANT_URL,
  apiKey: process.env.QDRANT_API_KEY,
});

Collection Configuration

await qdrant.createCollection('textbook_chunks', {
  vectors: {
    size: 1536, // OpenAI ada-002
    distance: 'Cosine',
  },
  hnsw_config: {
    m: 16,
    ef_construct: 100,
  },
});

Best Practices

For Physical AI textbook RAG:

• Collection name: textbook_chunks
• Vector size: 1536 (OpenAI ada-002 embeddings)
• Chunk size: 800 characters with 200 overlap
• Score threshold: 0.7 minimum for relevance
• Batch size: 100 vectors per upsert operation
• Metadata: Always include chapter, section, file path

Knowledge Base

For detailed information, see:

• Docker Setup → references/docker-setup.md
• Collection Management → references/collections.md
• Embedding Generation → references/embeddings.md
• Search Patterns → references/search-patterns.md
• Performance Tuning → references/performance.md