Photo to Listing: Barcode Decoding and LLM Re-rank
Explore the architecture of intent-longbox v0.2.1, a pipeline turning comic book photos into Shopify drafts using barcodes and LLM reranking.
Stock photo for illustration only, not from the actual event
- intent-longbox v0.2.1 concluded on 2026-09-01 after eleven commits
- Uses 12-digit UPC-A plus 5-digit supplement for precise identification
- LLM acts as a ranker, never as the source of truth
- Shopify products are always published as drafts for review
A comic shop employee photographing a back issue faces a tough identification problem, as a first-print variant and a common reprint share the exact same cover picture with different economics. This challenge led to the development of intent-longbox, a photo-to-listing pipeline that wrapped up version v0.2.1 on September 1, 2026, following eleven commits.
Research prior to building the architecture established a clear rule: LLM vision alone cannot handle issue-exact and variant-exact identification. Every working incumbent runs image-similarity retrieval against a reference corpus, ensuring the model functions strictly as a ranker rather than the source of truth.
Stock photo for illustration only, not from the actual event
The pipeline prioritizes deterministic components first through the following structured flow:
- Barcode decoding
- Candidate retrieval
- LLM reranking
- Human confirmation
- Condition and price assessment
- Shopify draft creation
Nothing is published without human oversight, as Shopify product entries land as drafts for owner review. The pilot store utilizing this setup is Gotham City Limit, running free of charge.
The deterministic front-end consists of a barcode parser completely free of machine learning models. Post-1990 comics carry a 12-digit UPC-A code identifying the series, accompanied by a 5-digit supplement encoding the issue number, cover variant, and printing batch.
"A 9.4 is a claim a phone photo does not entitle anyone to make."
intent-longbox architecture
The database architecture enforces an append-only audit trail where nearly all tables are immutable, preventing accidental data corruption from late-night hotfixes or migration scripts. Furthermore, numeric grading has been entirely removed from the schema and UI in favor of letter grade ranges, because a smartphone photo does not provide sufficient detail to justify a precise numerical grade.
For quality control, the VisionProvider interface mandates structured evidence fields where the model must report readable cover details such as the issue number, price box text, and publisher logo era. These details are cross-validated against top candidate metadata, and any contradiction causes the model to forfeit its fast path.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment