zvec（zvec 搜索）

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

1. Use Cases:

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

1. 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,
)

Node.js:

import { ZVecCreateAndOpen, ZVecCollectionSchema, ZVecFieldSchema, ZVecVectorSchema, ZVecDataType, ZVecHnswIndexParams, ZVecMetricType } from "@zvec/zvec";

const schema = new ZVecCollectionSchema({
  name: "my_collection",
  fields: [new ZVecFieldSchema({ name: "title", dataType: ZVecDataType.STRING })],
  vectors: [new ZVecVectorSchema({
    name: "embedding",
    dataType: ZVecDataType.VECTOR_FP32,
    dimension: 768,
    indexParams: new ZVecHnswIndexParams({ metricType: ZVecMetricType.COSINE }),
  })],
});

const collection = ZVecCreateAndOpen("./my_data", schema);

Core Concepts

Data Model

Collection

• Similar to a table in relational databases, a container for storing, organizing, and querying data
• Each Collection has a Schema defining its structure
• Each Collection is independently persisted in a dedicated directory on disk

Document

• Basic unit of data storage, similar to a row in a relational table
• Contains three core components:

- id: unique string identifier - vectors: named vector collection (supports dense and sparse vectors) - fields: named scalar field collection

Schema

• Dynamic Schema: scalar fields and vectors can be added or removed at any time
• Strong type system: each field must declare a DataType

Vector Types

Dense Vector

• Fixed-length real-valued embeddings
• Types: VECTOR_FP16, VECTOR_FP32, VECTOR_INT8
• Suitable for: semantic understanding, context capture

Sparse Vector

• High-dimensional representation with only a few non-zero dimensions
• Types: SPARSE_VECTOR_FP32, SPARSE_VECTOR_FP16
• Suitable for: keyword matching, BM25 scoring

Index Types

Available Topics

Python

• Quick Start - Quick start with Zvec Python API
• Collection Management - Create, open, and manage Collections
• Data Operations - Insert, update, and delete documents
• Vector Search - Single-vector, multi-vector, and hybrid search
• RAG System - Build document retrieval system
• Hybrid Search - Vector similarity + scalar filtering
• Multimodal Search - Image + text joint search

Node.js

• Quick Start - Quick start with Zvec Node.js API
• Collection Management - Create, open, and manage Collections
• Data Operations - Insert, update, and delete documents
• Vector Search - Single-vector, multi-vector, and hybrid search
• RAG System - Build document retrieval system
• Hybrid Search - Vector similarity + scalar filtering
• Multimodal Search - Image + text joint search

General

• Configuration - Global configuration and initialization
• Data Model - Zvec data model overview
• Embedding - Text embedding functions (Python only)
• Reranker - Result reranking functions (Python only)
• API Cheatsheet - Python & Node.js API quick reference
• Troubleshooting - Common issues and solutions

Available Topics

Python

• Collection Management
• Data Operations
• Hybrid Search
• Multimodal Search
• Quick Start
• Rag System
• Vector Search

Node.js

• Collection Management
• Data Operations
• Hybrid Search
• Multimodal Search
• Quick Start
• Rag System
• Vector Search

General

• Configuration
• Data Model
• Embedding
• Reranker
• Api Cheatsheet
• Troubleshooting