File formats

The JSONL shapes Kno reads: cases, pools, and mined transcripts, one object per line, stable ids, and what each field means.

All eval and pool files are JSONL: one JSON object per line.

Cases (evals)

{
  "id": "refund-01",
  "input": "How do I get a refund?",
  "expected": "Refunds are processed within 5 business days."
}

Field	Meaning
`id`	Stable identifier, the holdout is keyed on it; never reuse an id for a different case
`input`	What the agent receives
`expected`	What the goal scores against

Eval sources beyond bare JSONL: langsmith:<name>, langfuse:<name>, braintrust:<name>, and hf:<org>/<name>/<config>/<split> (Hugging Face datasets-server).

Pools (candidate assets)

{
  "id": "refund-policy-v3",
  "content": "Refunds are processed within 5 business days.",
  "kind": "knowledge"
}

Field	Meaning
`id`	Stable asset identifier
`content`	The asset body, a document, example, policy, instruction
`kind`	Asset kind (e.g. `knowledge`)

Pools also load from CSV (csv:pool.csv), Markdown (md:dir/ or md:file.md), and Hugging Face (hf:<org>/<name>/<config>/<split>:<kind>, the pool kind is part of the address). A bare path always means JSONL.

State

Runs, per-case scores and traces, assets, portfolios, and gaps live in kno.db, a SQLite file in the working directory (--db to override). There is no ~/.kno, no hidden global state.