Examples¶

The public examples live under examples/. These files are meant to stay in version control:

examples/data/mock_data.csv
examples/prompts/
examples/templates/
examples/schemas/
examples/config/

Generated run outputs go under examples/runs/ and are ignored by Git.

Single example¶

examples/config/run_config.json demonstrates:

SingleTaskBase
SimpleTemplateBuilder
StructuredOutputParser
OpenAIBatchProvider
LocalArtifactStore

Each request covers one input row. The bundled schema returns a bare object, and StructuredOutputParser can infer the single row id from request context when it is omitted. The bundled CSV uses query_id as that row-id column.

In the resulting parsed/responses.jsonl, each single-item row contains query_id, request_id, fields, and metadata. It does not include group_id.

Grouped example¶

examples/config/run_config_2.json demonstrates:

GroupedTaskBase
fixed-size groups with group_size: 3
a schema that accepts up to 3 returned items

With the bundled 10-row CSV, the grouped example produces 4 requests.

In grouped runs, parsed/responses.jsonl keeps query_id at top level for each row and also includes group_id, so you can trace both the original row and the batch grouping.

Artifact layout¶

Each run uses the canonical layout below:

runs/{run_id}/
  config/run_config.json
  metadata/manifest.json
  metadata/summary.json
  tables/units.jsonl
  tables/groups.jsonl
  tables/requests.jsonl
  raw/raw_outputs.jsonl
  raw/raw_errors.jsonl
  parsed/parsed_requests.jsonl
  parsed/responses.jsonl
  parsed/failures.jsonl

What to inspect first¶

If a run does not behave as expected, check files in this order:

metadata/manifest.json
metadata/summary.json
raw/raw_outputs.jsonl
raw/raw_errors.jsonl
parsed/failures.jsonl
parsed/responses.jsonl