What the documents look like
Cross-border product compliance and labelling work centres on product safety and label evidence. The materials handled here include product test reports, English-language labels, MSDS files, and platform prohibited-item lists. They are used before listing and before import. Errors at this stage can lead to platform delisting, customs seizure, or consumer claims. The materials are updated irregularly; each new product launch can bring a new version.
For contract and transaction material comparison in this category, test reports, label text, MSDS files, and prohibited-item lists must each be parsed into retrievable data chunks before they can be compared against corresponding fields in contracts or transaction materials. The quality of document parsing and chunking determines whether a comparison can locate the correct passage. A warning sentence separated from the model number it qualifies, for example, produces a retrieved fragment that cannot support any field-level comparison.
Constraints at the document chunking step
In FastGPT, a knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections exist for classification management and do not affect search results. One group of data can map to multiple vectors, and the indexes field records an array of vector IDs. During retrieval, vectors are recalled first, then the original data is located in the MongoDB dataset.datas table by vector ID. Records corresponding to the same original data are merged, and the vector score takes the highest value. If chunking splits a condition, an explanation, or a qualifier away from the content it constrains, the comparison will retrieve an incomplete fragment.
Full-text retrieval on vector stores such as PostgresSQL uses MongoDB $text with jieba tokenisation. Hybrid retrieval combines semantic and full-text results through RRF, then applies ReRank for secondary ordering. The citation limit is controlled by tokens, not by top k. The minimum relevance value applies only under semantic retrieval or when result reranking is used; it is a filter threshold and not a final ordering rule. Retrieval is not authorisation: a similarity setting does not exclude unrelated suppliers or unauthorised records, and any entity or permission restriction requires a separate downstream implementation.
Hypothetical drill sample: a supplier provides an English label whose source text reads "Product: Children's Toy Car. Model: TC-2024. Warning: Not suitable for children under 3 years. Batch: B2401." A draft sentence awaiting determination reads "This product label permits use by children under 3 years." The comparison object is the label warning field, the period is batch B2401, and the subject is this supplier's model. The source warning and the draft sentence take mutually exclusive values on the same field, which constitutes a contradiction.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Model and batch are exact identifiers requiring full-text recall; label descriptions require semantic recall |
| Citation limit | To be set against actual chunk length | Compare with the token count of the sample passage so the complete warning sentence is covered |
| Minimum relevance | To be set against observed scores | Compare with the retrieval score of the sample to filter low-relevance fragments |
| Question optimisation | Enabled for short questions | The sample draft sentence is short; complete the subject before retrieval |
| Result reranking | Enabled | When candidates are numerous, filter by the 0-1 score |
| Chunking method | Segmented by label field | Warning, model, and batch each form an independent data chunk |
These values are hypothetical starting points for this drill and are not official default values.
Three common errors
- Splitting the warning sentence and the model number into different chunks, so retrieval returns only the model chunk and the comparison cannot reach the warning field. Check: input the sample model number and warning keyword, compare whether the returned chunk contains the complete warning sentence.
- Treating minimum relevance as a permission or entity filter, assuming a higher value excludes other suppliers' records. Check: input the sample supplier identifier, compare whether returned records still contain other subjects, and confirm that filtering requires a separate implementation.
- Treating a successful upload as an updated index and comparing directly after a material change. Check: input the changed batch number, compare whether the returned content is the old version, and confirm that actual training and index status must be verified.
How to verify the setup
- Verification one: input the sample label source text and the draft sentence, compare the values on the warning field. The condition is satisfied if retrieval returns a chunk containing "Not suitable for children under 3 years" and the draft sentence is marked as contradictory. This is a paper expectation and has not been tested.
- Verification two: input the sample model number "TC-2024", compare whether full-text retrieval hits the chunk containing that identifier. The condition is satisfied if the returned chunk contains both the model and the batch. This is a paper expectation and has not been tested.
- Verification three: input the empty result
[], compare whether the downstream design handles it as no match. The condition is satisfied if the flow records the gap without generating a comparison conclusion. This is a paper expectation and has not been tested.