원클릭으로 Manus에서 모든 스킬 실행

vector-search

스타11

포크1

업데이트2026년 7월 24일 10:01

VelociDB vector search implementation - F32_BLOB(n)/VECTOR(n) columns, vector32 literals, distance functions, and exact parallel KNN. Use when modifying src/vector.rs, adding distance metrics, extending vector SQL syntax, or debugging dimension-mismatch, KNN-ordering, or vector-parsing issues.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

niklabh

niklabh/velocidb

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Vector Search

Turso/libSQL-compatible exact (brute-force) vector search. There is no ANN index; every query scans candidate rows, in parallel above 1024 rows.

Data flow

Stage	Where	Detail
Column type	`DataType::Vector(u32)` in `src/types.rs`	parsed from `F32_BLOB(n)` / `VECTOR(n)` in `DataType::from_str`
Value	`Value::Vector(Vec<f32>)`	row-format tag 5 in `src/btree.rs`
Literal	`parse_vector_constructor` in `src/vector.rs`	accepts `vector32('[...]')`, `vector('[...]')`, `vector64('[...]')`, quoted `'[...]'`, bare `[...]`
Insert validation	`execute_insert` in `src/executor.rs`	dimension must equal the declared `n`; NULL allowed
Schema persistence	`save_schema`/`load_schema` in `src/storage.rs`	type byte 5 followed by `dim: u32`

SQL surface

CREATE TABLE docs (id INTEGER PRIMARY KEY, embedding F32_BLOB(3));
INSERT INTO docs VALUES (1, vector32('[1.0, 0.0, 0.0]'));

SELECT id, vector_distance_cos(embedding, vector32('[1,0,0]'))
FROM docs
ORDER BY vector_distance_cos(embedding, vector32('[1,0,0]'))
LIMIT 5;

Distance functions: vector_distance_cos (1 - cosine similarity), vector_distance_l2 (euclidean), vector_distance_dot (negated dot product so lower = more similar for ALL metrics; keep that convention).
Distance expressions are parsed by parse_distance_expr in src/vector.rs — the parser (src/parser.rs) deliberately keeps them as opaque strings in OrderBy.column and the SELECT column list; the executor interprets them. If you add a function name, register it in DistanceMetric::from_function_name.
ORDER BY ascending + LIMIT k triggers the top-k KNN path (vector::knn, select_nth_unstable_by); descending does a full sort.
Rows with NULL or wrong-dimension vectors get distance f64::INFINITY (sort last) rather than erroring mid-scan.

Programmatic APIs

Sync: Database::vector_search(table, column, &[f32], k, DistanceMetric)
Async: AsyncConnection::vector_search(...)
Both return Vec<(f64 distance, Row)> ascending.

Parser pitfalls (learned the hard way)

vector32('[1, 2]') contains commas and parens. Value splitting (parse_values) and SELECT-column splitting (split_top_level_commas) are quote/paren/bracket aware — never replace them with a naive split(',').
The INSERT regex captures VALUES greedily to the final ); a lazy [^)]+ would cut nested calls short.
ORDER BY is located with a quote-aware scan (find_order_by), not a regex, because the expression may contain anything.

Tests

Unit tests live in src/vector.rs; end-to-end coverage in tests/advanced_features_tests.rs (roundtrip, dimension enforcement, KNN ordering, projection distances, reopen persistence, 1500-row parallel KNN).

cargo test --lib vector && cargo test --test advanced_features_tests

이 저장소의 다른 Skills

같은 저장소

async-io-model

niklabh/velocidb

How VelociDB's async API and parallel query execution work - the tokio spawn_blocking model in src/async_api.rs and the rayon thresholds in the executor. Use when modifying async_api.rs, adding async methods, tuning parallelism, or debugging blocked runtimes, stalled futures, or async test hangs.

2026-07-2411

cdc

niklabh/velocidb

VelociDB Change Data Capture - the in-memory change log in src/cdc.rs and its executor hooks. Use when modifying cdc.rs, adding CDC hooks to new write paths, changing ChangeEvent fields, or debugging missing/duplicate/out-of-order change events.

2026-07-2411

testing

niklabh/velocidb

How to run and write VelociDB tests - test suite layout, commands, and patterns for unit, integration, crash-recovery, property-based, and async tests. Use when adding features, fixing bugs, writing new tests, or deciding which suites must pass before a change is done.

2026-07-2411

code-quality

niklabh/velocidb

VelociDB code conventions - error handling, module boundaries, the experimental-modules policy, locking style, and documentation expectations. Use when writing or reviewing any VelociDB code, deciding where new code lives, or cleaning up warnings.

2026-07-2411

debugging

niklabh/velocidb

Debugging VelociDB - REPL-based reproduction, tracing, inspecting db/WAL files, and a map of common failure modes to their causes. Use when investigating bugs, corruption errors, deadlocks/hangs, wrong query results, or data that disappears after reopen.

2026-07-2411

sql-parser

niklabh/velocidb

Extending VelociDB's regex-based SQL parser in src/parser.rs - statement dispatch, quote/paren-aware splitting helpers, value parsing, and known fragility. Use when adding SQL syntax, new statements or operators, changing WHERE/ORDER BY/VALUES parsing, or debugging "Invalid ... syntax" / mis-split errors.

2026-07-2411

name	vector-search
description	VelociDB vector search implementation - F32_BLOB(n)/VECTOR(n) columns, vector32 literals, distance functions, and exact parallel KNN. Use when modifying src/vector.rs, adding distance metrics, extending vector SQL syntax, or debugging dimension-mismatch, KNN-ordering, or vector-parsing issues.

Vector Search

Turso/libSQL-compatible exact (brute-force) vector search. There is no ANN index; every query scans candidate rows, in parallel above 1024 rows.

Data flow

Stage	Where	Detail
Column type	`DataType::Vector(u32)` in `src/types.rs`	parsed from `F32_BLOB(n)` / `VECTOR(n)` in `DataType::from_str`
Value	`Value::Vector(Vec<f32>)`	row-format tag 5 in `src/btree.rs`
Literal	`parse_vector_constructor` in `src/vector.rs`	accepts `vector32('[...]')`, `vector('[...]')`, `vector64('[...]')`, quoted `'[...]'`, bare `[...]`
Insert validation	`execute_insert` in `src/executor.rs`	dimension must equal the declared `n`; NULL allowed
Schema persistence	`save_schema`/`load_schema` in `src/storage.rs`	type byte 5 followed by `dim: u32`

SQL surface

CREATE TABLE docs (id INTEGER PRIMARY KEY, embedding F32_BLOB(3));
INSERT INTO docs VALUES (1, vector32('[1.0, 0.0, 0.0]'));

SELECT id, vector_distance_cos(embedding, vector32('[1,0,0]'))
FROM docs
ORDER BY vector_distance_cos(embedding, vector32('[1,0,0]'))
LIMIT 5;

Distance functions: vector_distance_cos (1 - cosine similarity), vector_distance_l2 (euclidean), vector_distance_dot (negated dot product so lower = more similar for ALL metrics; keep that convention).
Distance expressions are parsed by parse_distance_expr in src/vector.rs — the parser (src/parser.rs) deliberately keeps them as opaque strings in OrderBy.column and the SELECT column list; the executor interprets them. If you add a function name, register it in DistanceMetric::from_function_name.
ORDER BY ascending + LIMIT k triggers the top-k KNN path (vector::knn, select_nth_unstable_by); descending does a full sort.
Rows with NULL or wrong-dimension vectors get distance f64::INFINITY (sort last) rather than erroring mid-scan.

Programmatic APIs

Sync: Database::vector_search(table, column, &[f32], k, DistanceMetric)
Async: AsyncConnection::vector_search(...)
Both return Vec<(f64 distance, Row)> ascending.

Parser pitfalls (learned the hard way)

vector32('[1, 2]') contains commas and parens. Value splitting (parse_values) and SELECT-column splitting (split_top_level_commas) are quote/paren/bracket aware — never replace them with a naive split(',').
The INSERT regex captures VALUES greedily to the final ); a lazy [^)]+ would cut nested calls short.
ORDER BY is located with a quote-aware scan (find_order_by), not a regex, because the expression may contain anything.

Tests

cargo test --lib vector && cargo test --test advanced_features_tests