Thanks for the pointers in your earlier reply! I’ve tried to implement a pages + blocks workflow using a toy example. I also used ChatGPT to help me draft the recipe.
Data (dataset/catalog.jsonl):
{"catalog_id":"CAT001","curator":"ops-team","catalog_items":[{"seq":1,"sku":"A-100","title":"Acme SuperWidget 3000","original_text":"super widget v3k","suggested_text":"Acme SuperWidget 3000"},{"seq":2,"sku":"B-200","title":"BoltMaster Kit (42pc)","original_text":"bolts kit 42 pieces","suggested_text":"BoltMaster Kit, 42-Piece"}]}
Recipe (recipe.py):
import prodigy
from prodigy.components.loaders import JSONL
from prodigy.util import set_hashes
from typing import Iterable, Dict, List
def _html_catalog_table(items: List[Dict], highlight_seq: str = None) -> str:
"""Context table with optional highlight for the current row."""
rows = []
for it in items:
seq = it.get("seq", "")
sku = it.get("sku", "")
title = it.get("title", "")
orig = it.get("original_text", "")
sug = it.get("suggested_text", "")
highlight = (str(seq) == str(highlight_seq)) if highlight_seq is not None else False
row_style = "background-color:#fff7cc;" if highlight else "" # pale yellow
rows.append(
f"<tr style='{row_style}'>"
f"<td style='padding:6px 8px; border-bottom:1px solid #eee'>{seq}</td>"
f"<td style='padding:6px 8px; border-bottom:1px solid #eee'>{sku}</td>"
f"<td style='padding:6px 8px; border-bottom:1px solid #eee'>{title}</td>"
f"<td style='padding:6px 8px; border-bottom:1px solid #eee'>{orig}</td>"
f"<td style='padding:6px 8px; border-bottom:1px solid #eee'>{sug}</td>"
f"</tr>"
)
header = (
"<table style='width:100%; border-collapse:collapse; font-size:14px'>"
"<thead>"
"<tr style='text-align:left'>"
"<th style='padding:6px 8px; border-bottom:2px solid #ccc; width:60px'>Seq</th>"
"<th style='padding:6px 8px; border-bottom:2px solid #ccc; width:120px'>SKU</th>"
"<th style='padding:6px 8px; border-bottom:2px solid #ccc'>Title</th>"
"<th style='padding:6px 8px; border-bottom:2px solid #ccc'>Original</th>"
"<th style='padding:6px 8px; border-bottom:2px solid #ccc'>Suggested</th>"
"</tr>"
"</thead><tbody>"
)
return header + "".join(rows) + "</tbody></table>"
def _iter_tasks(source: str) -> Iterable[Dict]:
"""
One parent task per catalog (pages). Each page = one item:
- Show full table as context with current row highlighted
- Editable text field with the current suggestion pre-filled
"""
for cat in JSONL(source):
items = cat.get("catalog_items") or []
if not isinstance(items, list) or not items:
continue
catalog_id = cat.get("catalog_id", "")
curator = cat.get("curator", "")
pages: List[Dict] = []
for it in items:
seq = it.get("seq", "")
sku = it.get("sku", "")
title = it.get("title", "")
orig = it.get("original_text", "") or ""
sug = it.get("suggested_text", "") or ""
context_table_html = _html_catalog_table(items, highlight_seq=seq)
page_html = (
"<div style='text-align:left; width:100%'>"
f"<div style='margin:0 0 8px 0; color:#666'>"
f"<strong>Catalog:</strong> {catalog_id}"
f" | "
f"<strong>Curator:</strong> {curator}"
f"</div>"
f"{context_table_html}"
"<hr style='margin:12px 0; border:none; border-top:1px solid #ddd'/>"
f"<div style='padding:6px 0; color:#666'><strong>Item Number:</strong> {seq}</div>"
f"<div style='padding:6px 0'><strong>SKU:</strong> {sku}</div>"
f"<div style='padding:6px 0'><strong>Title:</strong> {title}</div>"
f"<div style='padding:6px 0'><strong>Original Text:</strong> {orig}</div>"
"</div>"
)
page = {
"view_id": "blocks",
"config": {
"blocks": [
{"view_id": "html", "html": page_html},
{
"view_id": "text_input",
"field_id": "edited_text",
"field_label": "Normalized Text",
"placeholder": "Enter the normalized/catalog-ready text…",
"rows": 2
},
]
},
"edited_text": sug, # prefill
"text": sug or f"{catalog_id} item {seq}",# ensure hashable
# store meta-like data under a non-special key to avoid footer chips
"task_info": {
"catalog_id": catalog_id,
"seq": seq,
"sku": sku,
"title": title,
"original_text": orig,
"suggested_text": sug,
},
}
page = set_hashes(page, input_keys=("text",))
pages.append(page)
if not pages:
continue
parent = {
"view_id": "pages",
"pages": pages,
"text": f"Catalog {catalog_id}",
"task_info": {
"catalog_id": catalog_id,
"curator": curator,
"num_items": len(pages),
},
}
parent = set_hashes(parent, input_keys=("text",))
yield parent
@prodigy.recipe(
"catalog.normalize_review",
dataset=("Dataset to save annotations", "positional", None, str),
source=("Path to JSONL catalog file", "positional", None, str),
)
def catalog_normalize_review(dataset: str, source: str):
"""
Multi-annotator, no-overlap review/edit workflow:
- One catalog per task (pages), one page per item.
- Each catalog routed to a single session (no overlap).
- Record session user as `reviewer`.
"""
stream = _iter_tasks(source)
def before_db(examples):
for eg in examples:
session = (
eg.get("session_id")
or eg.get("_session_id")
or eg.get("session")
or ""
)
for p in eg.get("pages") or []:
orig = (p.get("task_info", {}) or {}).get("suggested_text", "")
edited_val = p.get("edited_text", "")
final = edited_val if isinstance(edited_val, str) and edited_val.strip() != "" else orig
p["final_text"] = final
p["edited"] = (final != orig)
p["reviewer"] = session
return examples
return {
"dataset": dataset,
"stream": stream,
"view_id": "pages",
"before_db": before_db,
"config": {
"lang": "en",
"auto_count_stream": True,
"buttons": ["accept", "reject", "ignore"],
# route whole catalogs to a single annotator
"feed_overlap": False,
"exclude_by": "input",
# hide meta chips just in case
"show_meta": False,
# (optional) comment out the keymap section; left out to keep MRE minimal
# If needed, demonstrate focus bug persists even without keybindings.
},
}
I’m getting the desired UI (screenshot below). The Normalized Text field is meant for the reviewer to edit whenever they feel the suggested description could be improved:
Problem: In the text_input block, the user loses focus after one keystroke. The annotator must click back into the field to continue typing.
Questions:
- Is there a workaround to keep focus so annotators can type normally?
- Also, could you please let me know if I’m doing something in a way that’s not the Prodigy way, and if there are improvements you’d recommend for this workflow design?
Thanks again,
Bilal
