Live demo

Paste a Spanish text, adjust the JSON schema and run one of the three pipelines submitted to GenSIE 2026. Fields the text does not answer should come back as null.

1

Choose an example or paste your own text

Text (Spanish)

2

Review the instruction and the schema Instruction

Edit JSON schema

3

Pick a pipeline and extract

Pipeline e226 — few-shot examples chosen by schema type, plus a per-field guide e231 — compact zero-shot prompt, tuned for F1 per token e232 — e226 hardened for unseen schemas: organizer examples, stricter type matching and a structural plus TF-IDF fallback

Result

About this demo

This is a JavaScript reimplementation of pipelines e226, e231 and e232 as submitted to GenSIE 2026 and described in the published paper. The prompts, the field guide, the text cleaning and the few-shot selection reproduce the submitted Python code and are checked against it by automated tests.
The post-processing is reduced: it keeps the null-alias and duplicate clean-up and the hallucination filter, and leaves out the normalisation layers in between. On the 24 test cases used to check it, the reduced version gave the same output as the full one in 21.
It runs on the open model gpt-oss-20b through Groq, not on the models evaluated in the paper and not on the fine-tuned models, so its results are not the published ones.
e226 takes its few-shot examples from the author's synthetic dataset, precomputed with the original selector; schemas that match a GenSIE development schema type use its generic examples instead. e232 takes them from the GenSIE starter and development sets by GIA-UH (MIT licence).
Your text and schema are sent to Groq to run the model and are not stored by this site.
The demo uses Groq's free tier: it can run about one e226 or e232 request per minute (e231 is much lighter) and a limited number per day. Other limits: texts up to 4,000 characters, schemas up to 30 fields and 4 levels of nesting, 10 requests per minute and 20 per day per connection.
The synthetic few-shot examples contain text from Spanish Wikipedia, CC BY-SA 4.0. The demo examples were written for this page; the news item and the clinical note are fictional.