熟练
- 作者仓库星标 275
- 作者仓库 skilld
- 领域
- 写作
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @skilld-dev · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- macOS · Linux · Windows
- 运行环境要求
- Python
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
# asg017/sqlite-vec `sqlite-vec`
**Version:** 0.1.7
**Tags:** latest: 0.1.7, alpha: 0.1.7-alpha.13
**References:** [package.json](./.skilld/pkg/package.json) — exports, entry points • [README](./.skilld/pkg/README.md) — setup, basic usage • [Docs](./.skilld/docs/_INDEX.md) — API reference, guides • [GitHub Issues](./.skilld/issues/_INDEX.md) — bugs, workarounds, edge cases • [Releases](./.skilld/releases/_INDEX.md) — changelog, breaking changes, new APIs
## Search
Use `skilld search` instead of grepping `.skilld/` directories — hybrid semantic + keyword search across all indexed docs, issues, and releases. If `skilld` is unavailable, use `npx -y skilld search`.
```bash
skilld search "query" -p sqlite-vec
skilld search "issues:error handling" -p sqlite-vec
skilld search "releases:deprecated" -p sqlite-vec
```
Filters: `docs:`, `issues:`, `releases:` prefix narrows by source type.
<!-- skilld:api-changes -->
## API Changes
This section documents version-specific API changes — prioritize recent major/minor releases.
- BREAKING: DELETE operations now properly clear vector data and free space — v0.1.7 changed behavior from only setting validity bits. Code using DELETE statements may see different storage behavior [source](./.skilld/releases/v0.1.7.md:L16)
- NEW: Distance column constraints in KNN queries — v0.1.7 adds support for `>`, `>=`, `<`, `<=` constraints on the distance column, enabling pagination-like patterns without requiring large k values [source](./.skilld/releases/v0.1.7.md:L17)
- NEW: Metadata columns in vec0 virtual tables — v0.1.6 added ability to declare metadata columns that can be filtered in WHERE clauses of KNN queries alongside vector matching [source](./.skilld/releases/v0.1.6.md:L13-27)
- NEW: Partition keys for internal index sharding — v0.1.6 added `partition key` syntax to internally shard vector indexes by column values [source](./.skilld/releases/v0.1.6.md:L23-24)
- NEW: Auxiliary columns with `+` prefix — v0.1.6 added support for auxiliary columns (prefix with `+`) that are unindexed but available for fast lookups in KNN query results [source](./.skilld/releases/v0.1.6.md:L31-33)
- BREAKING: `vec_npy_each` table function removed from default entrypoint — v0.1.3 moved this experimental function out due to CVE-2024-46488 security mitigation; affected code using untrusted SQL or the rare `vec_npy_each` function [source](./.skilld/releases/v0.1.3.md:L9)
**Also changed:** Static linking support for SQLite 3.31.1+ · `serialize_float32()` / `serialize_int8()` Python functions added
<!-- /skilld:api-changes -->
<!-- skilld:best-practices -->
## Best Practices
- **Use two-column re-scoring pattern for binary quantization** — store both quantized and full-precision vectors; query coarse index with quantized vectors, then re-score top candidates with full precision to recover quality lost from extreme dimensionality reduction [source](./.skilld/docs/binary-quant.md#re-scoring)
- **Combine `vec_slice()` with `vec_normalize()` for Matryoshka embeddings** — truncating dimensions requires subsequent normalization to maintain embedding quality and semantic meaning [source](./.skilld/docs/matryoshka.md#matryoshka-embeddings-with-sqlite-vec)
- **Prefer scalar quantization over binary quantization for moderate storage savings** — trade off storage efficiency against quality loss; `vec_quantize_float16` (2 bytes per value) and `vec_quantize_int8` (1 byte per value) offer better quality retention than binary quantization for many use cases [source](./.skilld/docs/scalar-quant.md#L1:26)
- **Use partition keys to shard large vector datasets** — declare a `partition key` column in `CREATE VIRTUAL TABLE` to internally shard the vector index on that column, improving query performance by reducing search scope [source](./.skilld/releases/v0.1.6.md#L23:24)
- **Combine metadata columns (indexed) with auxiliary columns (unindexed) for efficient filtering** — use regular metadata columns for dimensions you filter on in KNN WHERE clauses; prefix columns with `+` to store related data without indexing overhead [source](./.skilld/releases/v0.1.6.md#L26:33)
- **Use distance constraints instead of oversampling for pagination** — as of v0.1.7, apply `distance > threshold` or `distance < threshold` constraints in WHERE clauses to paginate through KNN results without fetching excess candidates [source](./.skilld/releases/v0.1.7.md#L17)
- **Monitor the k value limit when performing large KNN queries** — the default maximum k is 4096 (configurable) to prevent memory exhaustion; be aware that kNN results are materialized in memory and internally use O(n²) complexity on k [source](./.skilld/issues/issue-157.md#L22:33)
- **Rely on v0.1.7+ for automatic DELETE cleanup** — vector space is now reclaimed when enough vectors are deleted to clear a chunk (~1024 vectors); previous versions only marked entries as deleted without freeing space [source](./.skilld/releases/v0.1.7.md#L16)
- **Select embedding models with quantization support for better results** — models like `nomic-embed-text-v1.5`, `mxbai-embed-large-v1`, and OpenAI's `text-embedding-3` are specifically trained to maintain quality after quantization and Matryoshka truncation [source](./.skilld/docs/binary-quant.md#L114:125)
<!-- /skilld:best-practices -->
章节与文件索引
按说明文件中的顺序排列。
搜索
API变更
最佳实践
文件/命令 -> sqlite-vec · skilld search · .skilld/ · skilld · npx -y skilld search · docs: · issues: · releases:
内容校验值 -> 8b59958da96a
当前原文的自动摘译,尚未人工审校。
搜索
使用
skilld search而不是 grep.skilld/目录——跨所有索引文档、问题和版本的混合语义 + 关键字搜索。如果skilld不可用,请使用npx -y skilld search。过滤器:docs:、issues:、releases:前缀按源类型缩小。API变更
本节记录了特定于版本的 API 更改 - 优先考虑最近的主要/次要版本。重大突破:删除操作现在可以正确清除矢量数据和可用空间 - v0.1.7 改变了仅设置有效位的行为。使用 DELETE 语句的代码可能会看到不同的存储行为源新:KNN 查询中的距离列约束 — v0.1.7 添加了对距离列上的
>、>=、<、<=约束的支持,无需大 k 值即可启用类似分页的模式源最佳实践
使用两列重新评分模式进行二进制量化——同时存储量化向量和全精度向量;使用量化向量查询粗索引,然后以全精度重新评分顶级候选,以恢复极端降维源造成的质量损失将
vec_slice()与vec_normalize()组合用于俄罗斯套娃嵌入 - 截断维度需要后续标准化以维持嵌入质量和语义源