What the documents look like
Trade returns and claims work handles reverse logistics and liability determination after export. Materials arrive from customer email, quality inspection reports, customs systems and insurer claim systems. Each case forms its own group. A return agreement, a claim letter, a quality report and a return customs declaration commonly come from different issuers in different languages, and the same consignment can carry different reference numbers across those files.
Two retrieval behaviours follow from this. Multilingual retrieval depends on the vector model mapping meaning across languages, so a Chinese question can match an English passage. Reference numbers and declaration numbers depend more on exact keyword hits. These two query types pull chunking in different directions. When a liability condition is separated from the consignment it governs, a retrieved fragment cannot be traced back to its source text.
Constraints at the document chunking step
One set of source data can map to multiple indexes. Adding indexes does not change the source content; it increases the number of ways that content can be recalled. A paragraph should keep its structure and meaning intact, and a condition, its explanation and the consignment information it constrains belong in the same data block.
Whether an image can be retrieved depends on whether an index was created at ingestion time, and on whether a vision model or an image-capable vector model is configured. Hybrid retrieval merges multiple recall paths through RRF as part of the retrieval process. The minimum relevance value takes effect only under semantic retrieval or result re-ranking; it filters low-scoring results and does not perform permission or entity isolation. Search covers the whole library, while a collection only classifies data. Upload success and completed training or indexing are different states.
Configuration and implementation checks
Hypothetical drill sample: a trade after-sales staff member receives a scanned return agreement. The source text reads: "Party A agrees to accept the returned goods of consignment A-227, and the return reason is subject to Party B's March 2024 quality inspection report; this agreement takes effect on the date both parties sign." The same case file contains a draft claim letter with one sentence written as: "Both parties have reached a compensation amount for consignment A-227 and payment has been completed." The question is whether that draft sentence is supported by the available materials.
| Configuration item | Value or check action in this sample | Basis and comparison method |
|---|---|---|
| Search mode | Use full-text retrieval for reference A-227, semantic retrieval for liability wording | Full-text retrieval suits numbers and proper nouns; semantic retrieval suits natural language |
| Minimum relevance | Hypothetical trial starting value 0.5, compared against 0.3 and 0.7 | This value takes effect only under semantic retrieval or re-ranking; the starting value is hypothetical |
| Citation limit | Set by tokens, compared against counting chunks | Documents control the citation limit by tokens; no fixed number is given |
| Chunk check | Inspect whether the conditional clause and A-227 sit in one block | Separating a condition from what it constrains weakens traceability |
| Index status | Check the source data against the indexes field and whether training completed | Upload success does not equal an updated index |
| Image index | Confirm whether ingestion created an image description or image vector for the scan | Depends on model capability and whether an index was built at ingestion |
Three common errors
- Treating RRF in hybrid retrieval as a standalone switch that can be turned on. RRF is the merge step that combines multiple recall paths; configuration describes the search mode and re-ranking settings, and no such switch is documented.
- Using the minimum relevance value to exclude unrelated business or unauthorized records. That value is a similarity filter threshold, not a permission or entity filter, and retrieval itself does not complete authorization.
- Reading upload success as retrieval already updated. The source data, the indexes field and the actual training state need separate checking, and they can disagree.
How to verify the setup
- Enter the agreement source text and reference A-227 from the hypothetical sample, running full-text retrieval and semantic retrieval separately. Compare whether the returned content points back to the same consignment and the same agreement. If the number query does not hit that consignment, record it as unsupported rather than inferring a cause.
- Enter the draft sentence "Both parties have reached a compensation amount for consignment A-227 and payment has been completed" and compare it item by item against the agreement source text on party, object and state. The agreement states only acceptance of returned goods and an effective condition; it states no compensation amount and no completed payment. The sentence is judged unsupported.
- Enter the paper input empty result [] together with a pending handover request, and compare how the downstream design handles an empty citation. If the returned content is empty, record it as no cited content rather than asserting that a rule necessarily returns an empty array. For libraries across different vector models, check the identifier first and exclude non-matching ones; do not run cross-model retrieval to compare missed recalls.