Changelog
What changed in each ParceLabs product release: new features, improvements, fixes and security updates.
v1.4.0 · PDF and photo catalogues, version-free paths only
Catalogue parsing now reads PDF files and photos, and the old /v1 paths announced in the previous release are gone.
Breaking changes
- The
/v1paths are removed and answer404: call/easydogana/classifyand every other route without a version in the path. The version is still the optionalParceLabs-Version: YYYY-MM-DDrequest header, unchanged.
New
- Catalogue parsing reads PDF files (up to 20 pages) and photos, page by page: the text of the page when it has one, the image otherwise. Each row also gets its English description and the language it was written in.
- A page too dense to read in one pass is refused with
422 PAGE_TOO_DENSEand its page number, so you can split the file, instead of a generic error.
Improved
- Catalogue parsing waits up to 240 seconds by default, enough for a dense page of about 200 rows.
- The weight unit printed in a column header or cell is used; kilograms are assumed only when no unit is printed anywhere.
- Freight, shipping, insurance, discount and total lines are left out of the parsed rows and counted as skipped.
- Landed-cost analysis recognises item value, declared value, customs value and goods value columns, and value columns with a currency code such as
value_eur, as the price.
Fixed
- For destinations outside the EU, the
dutyNotAdValoremwarning now follows the destination's own duty lines, so a row priced in full is no longer marked for review.
v1.3.0 · Versioning by header, catalogue parsing and multilingual TARIC search
The API version now travels in a request header instead of the path, the same way as every ParceLabs API. This release also adds catalogue parsing and a free-text TARIC search, and makes classification steadier.
Breaking changes
- Paths no longer carry a version:
/easydogana/v1/classifyis now/easydogana/classify, and the same applies to every route. The version is the optionalParceLabs-Version: YYYY-MM-DDrequest header; without it you get the oldest supported version,2026-10-23. The old/v1paths still answer for a short transition, marked withDeprecationandSunsetheaders and a link to the new path, and will be removed soon: move your calls now.
New
- Every response states the API version it was served with (
ParceLabs-Version). An unknown version is refused with400 UNSUPPORTED_VERSION, a retired one with410 VERSION_RETIRED. - Catalogue parsing: upload a spreadsheet or CSV and get its rows back as clean product data, with column recognition, origin country detection and optional description translation. Nothing is stored and nothing is classified.
- Free-text TARIC search without any AI step: fast lookups of tariff lines by words, in the language of the query.
Improved
- TARIC search understands each language better (stemming and stop words per language), with a fallback that still matches across languages.
- The 19 languages that the EU publishes only twice a year now follow the current tariff structure every month: new lines appear in English until their translation is published, retired lines disappear.
- Tariff lines carry their level and whether they are declarable, and legal references resolve on every line.
- Classification is steadier: repeated requests for the same product give the same answer far more often, and only attributes stated in the input are used.
- Catalogue analysis classifies on the product name together with its description and recognises weight columns with units.
Fixed
- A catalogue row the classifier could not handle is marked as failed instead of failing the whole request, and is told apart from a genuine no-match.
- Chat search tools read the right TARIC fields and duty lookups return the duty lines.
- EUR-Lex legal-base links follow the language of the tariff detail.
Security
- Size and concurrency limits on uploads, archives and the heaviest AI routes.
- Webhooks carry a timestamp-bound second signature, and the reference verifier requires it.
- Chat output is sanitised and untrusted text is kept apart from instructions.
- Product text and operator emails are kept out of the logs.
v1.2.0 · Corrected documentation, public OpenAPI and landed-cost analysis
The API documentation was audited against the code and corrected, and the public API description is now generated from the code itself.
New
- Public OpenAPI description of the classification API, with a stable operation id, request and response examples and the required scopes on every operation, ready for client generators.
- Suggestions returned by classify carry
feedback_confirmedandfeedback_match_typewhen an operator has confirmed the code for a similar product. - A declared HS code is never ignored silently: responses record what happened to it (
hint_outcome). - Catalogue analysis prices imports: duty, VAT and landed cost per row, per destination, with clear flags when a row could not be priced.
- Catalogues can now be uploaded as documents (PDFs and images); the engine reads them into rows.
Improved
- Documentation corrected to match actual behaviour: scopes, modules, MCP tools and the quick reference.
- Total landed cost covers every row that could be priced, and the number of unpriced rows is reported next to it.
- Destination country detection accepts free-text values and non-standard column headers.
- Dependencies upgraded across the stack.
Fixed
- The snapshot build sequence now stops at the first failed step, and a retry resumes from the step that failed.
- Location headers on snapshot operations include the service prefix.
Security
- The REST routes of the beta modules now require OAuth scopes and validate request bodies; they are no longer reachable without a token.
- Outgoing webhook signatures are computed over exactly the bytes that are sent.
v1.1 · Retrieval quality and calibrated confidence
Classification accuracy on our evaluation set improved sharply: the right code is now first in 58.1% of cases (was 29.9%) and within the first five in 65.0% (was 53.0%).
New
confidence_reasonon classify, refine and batch responses explains why a result was rated high, medium or low confidence.retrieval_modeanddegraded_stageson classification responses tell you when part of the pipeline was unavailable and the answer came from a reduced search.
Improved
- Confidence is now calibrated: higher confidence ratings are measurably more often right than lower ones.
- Better ranking of candidate codes: chapter hints from the product's material and form, and an AI re-ranking pass over the shortlist.
- An operator-confirmed code keeps its confidence instead of being diluted by the ranking pass.
v1.0 · TARIC classification engine
First production release: classify a product description into a TARIC / HS code, with the reasoning attached.
New
- Classification API that returns a ranked list of TARIC codes with evidence, the position of each code in the nomenclature hierarchy and the legal references behind it.
- Hybrid search over the TARIC nomenclature, combining keyword and semantic matching.
- Product normalization for free-text and structured product descriptions, with detection of descriptions too generic to classify.
- Multilingual nomenclature: when a language is missing from a new TARIC release the previous text is carried forward, and the response says so with the
language_carried_forwardandlanguage_unavailablewarnings.