tsevals

tsevals

TypeScript evals for LLM apps. Define evals with a typed API, run them through vitest, store every run in SQLite, browse them in a UI, diff them from the CLI.

What's included

• defineEval — typed API for datasets, tasks, named scorers
• A vitest reporter that captures per-row results to SQLite as evals run
• A React UI with score-over-time charts and per-row inspection
• A CLI: run, dev, ui, show, diff, list — all read-side commands emit JSON
• Versioning: tag a run with a note (--note "switched to haiku-4.5") and deltas are computed against the previous tagged version

Install

npm install --save-dev tsevals vitest

[!NOTE] Requires Node 22+ (uses the built-in node:sqlite module). vitest is a peer dependency.

Quick start

Create an eval file ending in .eval.ts and export default an eval definition:

// examples/sentiment.eval.ts
import { defineEval } from "tsevals";

export default defineEval<string, "positive" | "negative" | "neutral">({
  name: "sentiment",
  data: () => [
    { input: "I love this!",                 expected: "positive" },
    { input: "Worst purchase ever.",          expected: "negative" },
    { input: "It's fine, I guess.",           expected: "neutral"  },
  ],
  task: async (input) => {
    // your model / agent / pipeline
    return await classifySentiment(input);
  },
  scorers: {
    exactMatch: ({ output, expected }) => (output === expected ? 1 : 0),
    llmJudge: async ({ output, expected }) => ({
      score: await judge(output, expected),
      metadata: { rationale: "..." },
    }),
  },
});

Run them:

npx tsevals run

Open the UI:

npx tsevals dev   # watcher + UI on http://localhost:3939

API

defineEval(config)

defineEval<TInput, TOutput>({
  name: string,
  data: () => DataItem<TInput, TOutput>[] | Promise<...>,
  task: (input: TInput) => TOutput | Promise<TOutput>,
  scorers: Record<string, Scorer<TInput, TOutput>>,
})

• Named scorers: scorers is a record, so each scorer has a stable identity across runs (used for per-scorer deltas).
• Scorer return: number or { score: number, metadata?: unknown }. Metadata is stored per row and shown in the UI on click.
• Data is a function: lazy, async-capable.
• trialCount: optional integer. Re-runs the full task+scorers pipeline N times per row and averages the score. Use when the task or scorers carry sampling noise (LLM-as-judge, temperature > 0). Per-trial values are stored alongside the mean and surfaced in the UI.

Convention

• Files matching **/*.eval.{ts,tsx,mts,...} are picked up by tsevals run.
• Each file exports an eval as default.
• Eval files coexist with regular *.test.ts — vitest's normal test runner ignores .eval.ts files.

CLI

tsevals run    [pattern] [--watch] [--note "..."] [--json]
tsevals dev    [--port]
tsevals ui     [--port]
tsevals show   <id|latest|prev-version> [--full]
tsevals diff   <from> [to=latest]
tsevals list   [--limit N] [--versions]

Command	Description
run	Run all evals (or a name regex). Saves a row to SQLite.
run --watch	Vitest watch mode — re-runs on file change.
run --note "..."	Tag this run as a version with a description.
run --json	Emit a structured run summary to stdout (no TTY noise).
dev	UI server + file watcher + auto-rerun + live UI polling.
ui	UI server only (production / inspect-only).
show <ref>	Print a run as JSON. --full includes per-row data.
diff <from> [to]	Per-eval and per-scorer score deltas. Exits 1 on regression.
list	Recent runs as JSON. --versions for tagged-only.

Refs latest and prev-version work everywhere a runId is accepted.

Use with agents

Every read-style command emits JSON, exit codes are meaningful, and the loop is scriptable:

tsevals run --json | jq '.score'                        # post-change score
tsevals diff prev-version || revert_changes              # auto-revert on regression
tsevals show latest --full | jq '.evals[].results[]'     # inspect rows

A skill for AI coding agents ships at skills/tsevals/SKILL.md. Point your agent (Claude Code, Cursor, etc.) at it for the iteration workflow — when to tag versions, how to inspect regressions, useful jq snippets.

CLI exit codes:

• run — 0 if all rows passed, 1 if any failed
• diff — 0 if no scorer regressed (delta > -0.001), 1 otherwise
• show — 0 on success, 2 if the ref is not found

CI: gating on regression

Use diff against a named version to fail the build on regression:

# after a green run on main:
tsevals run --note "release-2.4"

# in PR CI:
tsevals diff release-2.4
# exit 0 = no scorer regressed
# exit 1 = at least one scorer dropped

prev-version works the same way against whatever the latest tagged run happens to be.

[!NOTE] LLM-based scorers carry sampling noise. Use trialCount on the eval definition (see API) to average across multiple trials before relying on a single delta.

Config

Optional. Drop a tsevals.config.{ts,mts,mjs,js,json} in your project root.

// tsevals.config.ts
import { defineConfig } from "tsevals";

export default defineConfig({
  dbPath: ".tsevals/runs.db",
});

Currently supported keys:

Key	Default	Notes
dbPath	.tsevals/runs.db	Where the SQLite history is stored. Relative paths resolve from the config file's directory.

.ts configs are loaded via jiti so you can use TypeScript syntax without a build step. .mjs / .js use native ESM import; .json is parsed directly.

How runs work

Each run produces a row in .tsevals/runs.db (SQLite, schema-migrated automatically):

• A runs row: id, started/finished timestamps, duration, optional note
• A eval_results row per (data row × eval), with input/output/expected/scores/duration

A run with a non-empty note is a version. The UI's score chart and diff prev-version use versions as the comparison baseline.

Tag at runtime with --note "...", or after the fact via the inline note editor on each run in the UI.

Storage

• Database: .tsevals/runs.db in the working directory (gitignored by default)
• Backed by node:sqlite (Node 22+ built-in, zero native deps)
• Schema migrations are versioned and applied on first connection per process; re-running them is a no-op
• Inspect directly: sqlite3 .tsevals/runs.db

Thanks

The defineEval shape and the vitest-reporter approach are heavily inspired by evalite. Go check it out.