milvus

Milvus Lite spawns and manages a native C++ milvus binary as a subprocess, eliminating the need for separate server infrastructure. The ServerManager component handles process lifecycle (startup, shutdown, cleanup), while the Python client communicates via gRPC to the MilvusServiceImpl endpoint. This single-process architecture uses SQLite for file-based persistence, enabling zero-configuration deployment in Jupyter notebooks, laptops, and edge devices without Docker or Kubernetes.

Milvus Lite provides a schema definition system that allows developers to declare collections with typed fields (vectors, scalars, text) before data insertion. The schema validation occurs at the MilvusProxy layer, enforcing field types, dimensions, and constraints. Collections are persisted in SQLite and indexed via the Index component, supporting multiple vector types (dense float32/float16, sparse vectors) and scalar fields (int, float, string, bool) with optional filtering capabilities.

Milvus Lite uses CMake-based conditional compilation to build optimized binaries for multiple platforms (Ubuntu x86_64/ARM64, macOS Intel/Apple Silicon), with platform-specific code paths and dependencies. The Python package build system (setup.py, pyproject.toml) downloads the appropriate precompiled binary (~50MB) during installation, eliminating the need for users to compile C++ code. The build system detects the target platform and architecture, selecting the correct binary variant automatically.

Milvus Lite executes vector similarity searches through the Query Processing layer, which accepts a query vector and returns ranked results based on configurable distance metrics (L2, IP, COSINE, HAMMING). The search operation supports optional scalar filtering via WHERE clauses, limit/offset pagination, and output field selection. The Index component maintains in-memory vector indexes (FLAT, IVF_FLAT, HNSW, etc.) that are queried during search, with results ranked by similarity score and optionally re-ranked by scalar fields.

Milvus Lite supports BM25 full-text search through sparse vector indexing, where text fields are tokenized and converted to sparse vector representations. The Index component creates sparse indexes that enable keyword-based retrieval with TF-IDF weighting. Sparse vectors can be searched independently or combined with dense vectors in hybrid search queries, with results ranked by BM25 relevance scores. This capability bridges traditional full-text search and modern vector search in a single system.

Milvus Lite enables hybrid search by combining results from multiple vector indexes (dense + sparse) or multiple dense indexes with different metrics, then re-ranking by weighted scores or scalar fields. The Query Processing layer executes parallel searches across indexes and merges results using configurable weighting strategies (e.g., 70% semantic relevance + 30% BM25 score). Re-ranking can apply scalar field sorting (e.g., recency, popularity) to refine final rankings without re-executing searches.

Milvus Lite's Index component creates and manages in-memory vector indexes (FLAT, IVF_FLAT, HNSW, etc.) that accelerate similarity search. Index creation is triggered explicitly via the create_index() API, specifying the index type, distance metric, and parameters (e.g., nlist for IVF, M/ef for HNSW). Indexes are built synchronously and stored in memory, with optional persistence to SQLite. The index selection strategy balances accuracy (FLAT is exact, HNSW is approximate) against query latency and memory consumption.

Milvus Lite provides CRUD (Create, Read, Update, Delete) operations through the Data Operations layer, supporting insert, upsert, delete, and query methods. Upsert combines insert and update semantics, replacing existing records by primary key or inserting new ones. Batch operations accept lists of records and process them efficiently through the gRPC service layer, with results returned as operation summaries (inserted count, deleted count, etc.). All operations are persisted to SQLite and reflected immediately in subsequent queries.

Milvus Lite implements identical Python API to Milvus Standalone and Distributed deployments, allowing the same code to run across all deployment types by changing only the connection URI. The MilvusClient class abstracts the connection details (local file path for Lite, HTTP endpoint for Standalone, gRPC for Distributed), while all collection, data, and search operations remain unchanged. This compatibility is achieved through a unified gRPC service layer that works identically whether the server is a subprocess or remote instance.

Milvus Lite supports scalar field filtering through WHERE clause expressions that are evaluated during search or query operations. The MilvusProxy layer parses filter expressions and applies them to scalar fields (int, float, string, bool) before or after vector search, depending on the query type. Filters support comparison operators (==, !=, <, >, <=, >=), logical operators (AND, OR, NOT), and range queries. Filtered results are returned with matching vectors and metadata, enabling precise data retrieval without separate post-processing.

Milvus Lite provides collection statistics and metadata through the MilvusClient API, exposing information such as row count, memory usage, index status, and field definitions. The ServerManager and MilvusLocal components track collection metadata in SQLite, while the gRPC service layer exposes this information through describe_collection() and get_collection_stats() methods. Statistics are updated synchronously after data operations, providing real-time visibility into collection state without separate monitoring systems.