Skip to content
CAI
Software that uses CAICheck a score

RyanCodrai/turbovec

74.3

Strong · 14 November 2026

40k

lines of production code

Rust

with Python

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Turbovec is a high-performance vector search library that provides a Rust core with Python bindings for storing, indexing, and querying embeddings. It supports incremental, crash-safe persistence with format conversion and offers configurable similarity metrics like cosine and dot product. The system is designed for integration with major LLM frameworks such as LangChain and LlamaIndex, featuring robust validation, cross-platform performance optimization, and comprehensive durability testing.

Features

Initial Python bindings for TurboQuantIndex with hardened argument validation

This change introduces the Python binding layer for the TurboQuantIndex vector store, exposing the Rust core to Python via PyO3 and NumPy. The bindings implement strict input validation, replacing opaque panics with descriptive Python exceptions (ValueError/TypeError) for argument mismatches, and introduce a hardened mask-reading mechanism that bypasses Python method lookups to prevent potential filter bypasses from ndarray subclasses. It also includes parallel buffer copying utilities for efficient data ingestion.

turbovec-python/src · high confidence

Interruptible long operations and per-store similarity modes

Long search and add operations are now interruptible via Python-side batch chunking, allowing Ctrl-C to be serviced between slices rather than waiting for the entire batch to complete. Additionally, stores now support configurable similarity modes (cosine or dot\_product), allowing users to choose between normalized cosine similarity (default) and raw inner-product scoring, with the mode persisted in the side-car file.

turbovec-python/python · high confidence

New benchmark suites for insertion and persistence latency

Added benchmark scripts in benchmarks/suite to measure online-update insertion latency (single-vector and 100-vector batches) and persistence performance (save, load, and round-trip) for TurboQuantIndex. These suites compare TurboQuant against FAISS PQ on x86 and ARM architectures, using both multi-threaded and single-threaded execution, across 1536 and 3072 dimensions with 2-bit and 4-bit quantization.

benchmarks/suite · high confidence

New benchmarking infrastructure for performance validation and visualization

This change introduces a new suite of tools to automate benchmark execution, data handling, and result visualization. The \whm.py\ script provides a weighted harmonic mean scoring harness to evaluate performance across six operations (search, insert, delete, save, load, load\_search) on both ARM and x86 architectures, including single-threaded variants, ensuring that improvements in a target operation do not regress others. The \download\_data.py\ utility manages the acquisition of benchmark datasets (GloVe and OpenAI embeddings) with atomic file writes to prevent corruption. Finally, \create\_diagrams.py\ generates SVG charts in the project's visual style, displaying speed comparisons between TurboQuant and FAISS, as well as recall metrics, to provide clear performance insights for users.

benchmarks · high confidence

New diagnostic and benchmarking examples for index conversion and kernel performance

The examples directory now includes several new tools to aid in development and performance tuning. The \convert\ example allows users to convert index files between format versions v5, v6, and v7, or to report the version of an existing index. New micro-benchmarks (\insert\_bench\, \load\_bench\_v7\) provide isolated metrics for insertion/removal latency and file load times. Additionally, architecture-specific probes (\kernel\_roofline\, \kernel\_roofline\_neon\, \probe\_2bit\_sdot\, \probe\_2bit\_vnni\, \probe\_2bit\_lutstream\, \stream\_bw\) and cross-arch parity tests (\kernel\_xtest\, \encode\_hash\) are added to measure CPU instruction throughput, memory bandwidth, and ensure deterministic encoding consistency across x86\_64 and aarch64 platforms.

turbovec/examples · high confidence

Structured benchmarking framework for performance optimization

The \benchmarks/hillclimb\ directory now provides a formalized framework for tracking and executing performance optimizations. It introduces structured goal definitions (\GOAL\_2bit.md\, \GOAL\_mutate.md\, \GOAL\_persist.md\) that specify measurement rigs, correctness gates, and win conditions for specific optimization targets like 2-bit search, live-index mutation, and whole-file persistence. These goals are supported by detailed results logs (\LOG.md\, \LOG\_2bit.md\, \LOG\_mutate.md\, \LOG\_persist.md\, \LOG\_search.md\, \LOG\_sync.md\) that document the hypothesis-driven process, including specific performance wins, refutations, and infrastructure notes for the benchmarking environment.

benchmarks/hillclimb · high confidence

turbovec 1.0.0: stable v7 format, sync() incremental saves, and format conversion

The first stable release (1.0.0) establishes v7 as the only on-disk format for both the Rust crate and Python package, refusing to load older v5/v6 files to ensure format stability. It introduces sync() for incremental, crash-safe saves that write only changed rows, and provides a new convert utility to migrate existing v5 or v6 index files forward to v7 or back. This release also includes several critical bug fixes, such as resolving aarch64 search performance regressions, preventing unbounded memory allocation during load, and fixing codes buffer doubling on small adds.

(repo-wide) · high confidence

Behavioural changes

Rust core crate with v7 index format and legacy conversion

The library has been restructured into a pure Rust core crate (\turbovec/src\) with Python bindings, introducing the v7 index format as the native read/write standard. This format uses an incremental, headerized slot-array container with atomic save protocols and CRC-32C integrity checks. To support existing data, a new \convert\ module allows seamless migration between v5, v6, and v7 index files in any direction without re-encoding. The core also introduces stable external IDs via \IdMapIndex\, robust input validation that returns descriptive errors instead of panicking, and cross-platform reproducible quantization codebooks.

turbovec/src · high confidence

Support direct cargo builds on macOS

Users can now run \cargo build\ directly on macOS without encountering symbol resolution failures. A new \build.rs\ script was added to automatically emit the correct linker arguments (specifically \-undefined dynamic\_lookup\) required for Python extension modules, ensuring that symbols from the Python interpreter resolve correctly at load time.

turbovec-python · high confidence

Test coverage

Added adversarial durability and performance regression tests; Added downstream smoke test to verify public API stability; Expanded test coverage for Python integration stores and core bindings.

Dependencies

Initial release of turbovec 1.0.0 with Rust-Python workspace structure

This change introduces the initial 1.0.0 release of the turbovec vector search library, establishing a Rust workspace that includes the core \turbovec\ crate and its \turbovec-python\ bindings. The update pins specific dependency versions to ensure deterministic encoding and index stability: \rand\_chacha\ is fixed at 0.3.1 to prevent silent index corruption from stream changes, and \statrs\ is pinned at 0.17.1 to maintain consistent calibration anchors. The Python bindings require Rust 1.89 due to stabilized AVX-512 intrinsics and declare optional integrations for LangChain, LlamaIndex, Haystack, and Agno.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 75 → 74 (-0.9)
  • Rubric changed (rubric-2026.10.5 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 86 → 81 (-4.5)
  • Architecture 96 → 96 (+0.0)
  • Maturity 67 → 68 (+0.9)
  • Readiness 77 → 77 (+0.0)
  • Security 83 → 76 (-6.5)
  • Performance 100 → 100 (+0.0)

Resolved (11)

  • Duplicated block (11 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (11 lines × 2) (turbovec/src/id_map.rs)
  • Duplicated block (11–12 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (12 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (13 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (13–18 lines × 4) (turbovec/src/search.rs)
  • Duplicated block (20–25 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (5 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (6 lines × 2) (turbovec/src/search.rs)
  • Hotspot: turbovec-python/src/lib.rs (turbovec-python/src/lib.rs)
  • Hotspot: turbovec/src/error.rs (turbovec/src/error.rs)

New (45)

  • Duplicated block (11 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (11–12 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (11–15 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (12 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (13–18 lines × 4) (turbovec/src/search.rs)
  • Duplicated block (19–21 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (20–25 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (3–5 lines × 4) (turbovec/src/lib.rs)
  • Duplicated block (6 lines × 2) (turbovec/src/search.rs)
  • Duplicated block (6 lines × 3) (turbovec/src/search.rs)
  • Duplicated block (7 lines × 2) (turbovec-python/src/lib.rs)
  • Duplicated block (7 lines × 2) (turbovec/src/search.rs)
  • Edited copies of one member (3 members, 130+ corresponding lines) (turbovec/src/rotation.rs)
  • Edited copies of one member (3 members, 17+ corresponding lines) (turbovec/src/search.rs)
  • Edited copies of one member (4 members, 12+ corresponding lines) (turbovec/src/pack.rs)
  • Edited copies of one member (4 members, 3+ corresponding lines) (turbovec-python/python/turbovec/agno.py)
  • Edited copy of a member (11 corresponding lines) (turbovec/src/io_v7.rs)
  • Edited copy of a member (11 corresponding lines) (turbovec/src/io_v7.rs)
  • Edited copy of a member (14 corresponding lines) (turbovec-python/src/lib.rs)
  • Edited copy of a member (15 corresponding lines) (turbovec-python/python/turbovec/agno.py)
  • …and 25 more

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

RyanCodrai/turbovec was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 14 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit c3b043b992290ff223b903046af7f6683768241c — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.65+preprod-e1a3b69fc8ee.