Plugin → web app

Migration Checklist

Every capability from the ChatGPT plugin export, mapped to its status here.

Implemented

done
JSON import & input detection
Distinguishes RAG array, wrapped records and raw crawl envelope.
done
Exact eleven-key schema validation
Keys, types, snake_case IDs, lang, 3–5 variants, 4-field metadata, Malaysia-time crawled_at_my_tz, 30–70-word search_text, source links.
done
Audit ledger
Every input entry is approved / hold / excluded with reasons; merge mappings retained.
done
06 270 phone exclusion
Same regex as the original crawler; +606 229 2182 and emails preserved.
done
Website crawler (port of crawl_website.py)
Depth 4 / 300 attempts / 5 redirects / 0.5 s / 20 s / 8 MiB, robots, www scope, EN+MS, document-route filter, priority, checkpoints.
done
Deduplication
Duplicate IDs, identical answers, near-duplicate questions with reviewer-controlled merge.
done
RAG JSON export
Approved-only top-level array, web_<entity>_rag_<DDMMYY>.json, KL date.
done
Q&A workbook export
Real XLSX: one Q&A sheet, 4 text columns, frozen header, filter, reopened and reconciled 1:1.
done
Word audit summary
Genuine .docx, ≤200 words, built from audit evidence only.
done
Bilingual RAG generation (Stage 2)
Lovable AI converts each crawled page under the original SKILL.md, DBP and Fact & Figure rules. Metadata/source_url/lang come from the fetch; every record must quote verbatim evidence or it is held; 06 270 numbers scrubbed.

partial
Vector readiness
Deterministic checks + local BM25 retrieval with must_match. No embeddings or deployed vector-store test.
partial
Knowledge Manager persistence
Stored in this browser (IndexedDB). Shared, multi-user storage needs Lovable Cloud.
partial
Crawler runtime
Runs from the open browser tab; closing it stops the crawl (checkpoint kept). No JavaScript page rendering.
partial
Workbook citation rule
Original instructions conflict; resolved by an explicit user toggle in Export Center.
partial
Audit Word review
Content verified; rendered visual layout review is not automated.

partial
Semantic evidence review
Verbatim-evidence check is automatic; contradiction/currentness review across pages still needs a human reviewer.
missing
DBP Malay sign-off
DBP rules are applied during generation, but authoritative DBP terminology lookup has no service adapter; human sign-off required.
missing
Generation from uploaded legacy RAG/XLSX
Re-converting existing records (e.g. legacy FAIX entries lacking provenance) is not supported; they stay held.
missing
XLSX input import
Uploading XLSX as source input is not yet supported (JSON only).
missing
Staff directory / glossary snapshots
Not in the export; cannot be recreated.
missing
Fact & Figure label verification by rendering
Needs a headless browser/screenshot service.

done
Lovable AI
Enabled; used for bilingual generation.
missing
Lovable Cloud database
Cloud is enabled, but datasets still live in the browser; shared storage and sign-in are not built yet.
missing
Vector store + embeddings
Your Development Vector Store (provider and credentials not specified).
missing
Rendering service (optional)
For JavaScript pages and statistic cards.

done
Automated tests (14 passing)
Phone rule, document routes, URL normalisation, cfemail, robots, parser contacts, crawl engine with redirects/robots/scope, schema rules, FAIX sample (53 entries held), dedup, naming, XLSX round-trip.
done
Live AI generation
One real generation on FAIX page text: 3 records, 0 schema issues, 06 270 fax removed, +606 229 2182 kept.
partial
Live crawl
Engine tested with simulated site responses only; a live crawl of a UTeM site has not been run in this session.
missing
Retrieval quality
No deployed vector-store or Playground test.