How does syntaqlite compare?

There are many SQL tools out there. This page tests them head-to-head on SQLite-specific SQL, the kind of syntax that trips up generic SQL parsers, and shows raw, reproducible results.

Generated on arm64-darwin with syntaqlite 0.1.0 on 2026-03-14. To reproduce: tools/run-comparison --setup && tools/run-comparison --all. Per-statement breakdowns and reproduction details are in the detailed results.

Parsing

What we test: The corpus contains 40 statements covering advanced SQLite syntax, including UPSERT, RETURNING, STRICT tables, window frames with EXCLUDE, numeric underscores, IS NOT DISTINCT FROM, recursive CTEs, and more. Each statement is first validated against sqlite3 itself (the ground truth), then run through every parser. A tool scores "correct" only if it agrees with sqlite3.

Accuracy

Speed

We benchmark two file sizes: a 40-statement file (startup-dominated) and that file repeated 30× (throughput-dominated).

bench.sql (1×)

bench_30x.sql (30×)

Formatting

What we test: Round-trip semantic preservation. Each of the same 40 statements is formatted, then we run EXPLAIN on both the original and formatted SQL and compare the bytecode sqlite3 produces. Identical bytecode means sqlite3 will execute the exact same operations, so the formatter preserved semantics, not just validity. Tools that crash or refuse to format score "refused". Tools whose output produces different bytecode score "corrupt".

Accuracy

Speed

bench.sql (1×)

bench_30x.sql (30×)

Analysis

What we test: We check whether each tool can identify unknown tables, bad column references, incorrect function arity, and CTE column mismatches without running the query. The corpus contains 24 cases against a known schema, with 15 intentional errors and 9 valid queries. Each tool's verdict is checked against sqlite3.

Accuracy

Schema: users, orders, products, order_items. Ground truth: sqlite3.

Diagnostic quality

The test query contains two errors: its CTE declares three columns while its SELECT produces two, and ROUND is misspelled as ROUDN:

syntaqlite finds both errors in one pass, with source locations and a did-you-mean suggestion:

error: table 'monthly_stats' has 2 values for 3 columns
  --> /var/folders/rx/t6_rqmqx0f15l7kgp7yjhcbc0000gn/T/tmpico33e5u.sql:30:3
   |
30 |   monthly_stats(month, revenue, order_count) AS (
   |   ^~~~~~~~~~~~~
warning: unknown function 'ROUDN'
  --> /var/folders/rx/t6_rqmqx0f15l7kgp7yjhcbc0000gn/T/tmpico33e5u.sql:42:3
   |
42 |   ROUDN(ms.revenue / ms.order_count, 2) AS avg_order
   |   ^~~~~
   = help: did you mean 'round'?

sqlite3 stops at the first error (runtime execution):

Error: in prepare, table monthly_stats has 2 values for 3 columns

Speed

bench.sql (1×)

bench_30x.sql (30×)

LSP (Language Server)

What we test: We start each language server, open a test document, and probe for completions, hover, diagnostics, and formatting. Results are from actual LSP protocol responses, not self-reported feature lists.

Key difference: sqls requires a live database connection for its features. syntaqlite and sql-language-server work offline.

Features

¹ sqls requires a live database connection. Completion and hover results come from the connected database schema, not static analysis. Without a database, these features return no results.

Completion depth

Startup + response speed

Time from server start → document open → diagnostics received → exit: