zvec

Usage Instructions

Before starting, understand the following:

1. Development Language: Python or Node.js?

   • Python: use pip install zvec
   • Node.js: use npm install @zvec/zvec

2. Use Cases:

   • RAG document retrieval system
   • Semantic search
   • Multimodal search (image + text)
   • Hybrid search (keywords + semantic)

3. Data Scale:

   • < 100k: use FLAT index (exact search)
   • 100k-10M: use HNSW index (recommended default)

   • 10M: use IVF index (memory optimized)

Decision Workflow

• User needs vector search functionality
  • Choose development language (Python/Node.js)
  • Determine use case
    • RAG system → use single-vector search + document chunk management
    • E-commerce search → use hybrid search (vector + filter)
    • Multimodal → use multi-vector search + weighted ranking
  • Design Schema (vector fields + scalar fields)
  • Select index type (HNSW/FLAT/IVF)
  • Implement data synchronization strategy

Default Recommendations

• Use create_and_open() / ZVecCreateAndOpen() to create Collection
• Use cosine similarity (COSINE) as default distance metric
• Use FP32 type for dense vectors
• Create InvertIndexParam index for filter fields

Validation Checklist

• Vector dimensions match Schema definition
• Scalar field types are correct
• Filter condition syntax is correct
• Call optimize() after large batch writes

Quick Start

Python:

import zvec

# Create Collection
schema = zvec.CollectionSchema(
    name="my_collection",
    fields=[
        zvec.FieldSchema(name="title", data_type=zvec.DataType.STRING),
    ],
    vectors=[
        zvec.VectorSchema(
            name="embedding",
            data_type=zvec.DataType.VECTOR_FP32,
            dimension=768,
            index_param=zvec.HnswIndexParam(
                metric_type=zvec.MetricType.COSINE
            ),
        ),
    ],
)

collection = zvec.create_and_open("./my_data", schema)

# Insert document
collection.upsert(zvec.Doc(
    id="doc_1",
    vectors={"embedding": [0.1] * 768},
    fields={"title": "Hello World"},
))

# Search
results = collection.query(
    vectors=zvec.VectorQuery(
        field_name="embedding",
        vector=[0.1] * 768,
    ),
    topk=10,
)