SymphonyQG

Category: SDKs & Libraries
Brand: gouyt13
Source: https://github.com/gouyt13/SymphonyQG

SymphonyQG is a research codebase and ANN indexing method that integrates vector quantization with graph-based indexing to build efficient approximate nearest neighbor (ANN) indexes for high-dimensional vector search. It is aimed at vector databases and similarity search workloads where combining compact quantized codes with navigable graphs improves recall–latency trade-offs and reduces memory usage.

---

Features

Core Capabilities

• Approximate Nearest Neighbor (ANN) search for high-dimensional vectors.
• Hybrid quantization + graph approach ("quantized graph") to balance recall, latency, and memory footprint.
• Configurable similarity metric (e.g., choose distance metric when creating the index).
• Bounded graph degree via degree_bound parameter to control graph connectivity and index size.

Library Structure

• data/
  • Contains datasets and pre-built indices used for experiments and reproduction.
• symqglib/
  • index/
    • fastscan/: helper functions for FastScan-style operations.
    • qg/: implementation of the quantized graph index.
  • third/: third-party dependencies needed by the library.
  • utils/: common utility functions.
• python/
  • Python bindings exposing the core C++ ANN index to Python.
• reproduce/
  • Scripts and code to reproduce experimental results, including configuration and dataset handling.
• test/
  • Tests for validating core functionality (implied by directory name).

Python API (Recommended Interface)

Index Construction

symphonyqg.Index(index_type, metric, num_elements, dimension, degree_bound=32)

• Parameters
  • index_type: type of ANN index to build (e.g., specific quantization/graph variant; exact values in repo docs/code).
  • metric: distance/similarity metric (e.g., L2, inner product; details in repo).
  • num_elements: number of vectors the index will hold.
  • dimension: dimensionality of each vector.
  • degree_bound (optional, default = 32): upper bound on graph node degree to control index sparsity and memory.

Methods

• build_index(data, EF, num_iter=3, num_threads=ALL_THREADS)

  • Builds the ANN index from training data.
  • data: input dataset, shape (num_elements, dimension), dtype=float32.
  • EF: construction/search breadth parameter (controls exploration during graph-based search/build).
  • num_iter: number of construction iterations (default 3).
  • num_threads: number of threads to use (default ALL_THREADS).

• save(filename)

  • Saves the built index to disk.

• load(filename)

  • Loads a previously saved index from disk.

• set_ef(EF)

  • Adjusts the runtime search parameter EF to trade off latency vs. recall without rebuilding the index.

• search(query, k)

  • Performs ANN search.
  • query: query vector(s), shape (dimension,) or (1, dimension), dtype=float32.
  • k: number of nearest neighbors to retrieve.

Examples & Reproducibility

• Python examples: basic usage via Python bindings (creation, build, search, save/load) indicated in README.md.
• Real-world datasets: example configurations and scripts under ./reproduce.
• Datasets description: additional details in ./data/README.md.
• Reproduction guide: steps for experiments in ./reproduce/README.md.

C++ Usage

• C++ examples are provided (mentioned under “C++ examples”) for users who want to integrate SymphonyQG directly at the C++ level instead of Python.

License

• A LICENSE file is present in the repository; consult it directly for the exact open-source license terms.

---

Pricing

SymphonyQG is an open-source research codebase hosted on GitHub. No pricing or paid plans are specified in the available content.