What the documents look like
Fresh-food supplier claims pursue quality responsibility for goods at arrival. They do not handle ordinary in-store shrinkage, which belongs to a different process such as near-expiry markdown handling. The working file for one claim commonly includes a quality exception note, photo evidence, a deduction notice and a claim reconciliation statement. These materials come from quality inspection systems, WMS receiving records, supplier contracts and finance reconciliation.
In cross-border work, materials for the same arrival batch may be written in different languages. A claim basis has to trace back to a specific batch, a specific receiving record and a specific deduction line. Updates occur per batch and per month. Citation tracing in this setting means locating the original entry inside multilingual materials and stating the status that entry actually carries. A retrieval hit shows that content was found; it does not show that a supplier agreed, that a deduction was executed or that a reconciliation closed.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. When the array is empty the output path still runs. That observation means this retrieval returned no citation; it does not mean the workflow terminated with an error.
Semantic retrieval supports cross-language understanding, so a Chinese question can match an English knowledge point. The retrieval result itself does not constitute authorization, approval or business confirmation. The minimum relevance value is a filter threshold and takes effect only in semantic retrieval or when result reranking is used. It does not separate records by supplier, by permission or by entity. Hybrid retrieval merges multiple result paths through RRF. The citation limit controls the maximum amount of content cited per search, measured in tokens. Image retrieval requires an image input and the corresponding model capability. Reranking requires a rerank model and must be enabled in settings. The optional paths listed in the documentation do not mean every request executes all of them.
Hypothetical drill sample: the original quality inspection record for one fresh-food arrival reads, "12 March 2026, batch A-0312, accepted quantity 200 cases, sampling found 12 cases of wilted leafy greens, judged a quality exception, quality exception note QC-0312 issued." The corresponding deduction notice reads, "Under quality exception note QC-0312, deduct RMB 2,400 from batch A-0312; notice status pending supplier confirmation." A draft sentence now awaits review: "The supplier has confirmed the RMB 2,400 deduction for batch A-0312, and the claim reconciliation is complete." To judge this draft, compare the subject, batch, amount and status fields. The original text shows the deduction notice still pending supplier confirmation, while the draft claims supplier confirmation and a completed reconciliation. That assertion is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Select multilingual claim libraries sharing one vector model | Check whether the selected knowledge base identifiers and vector model identifiers agree, and exclude those that do not |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone for hits on identifiers and document numbers |
| Citation limit | Set to 3,000 tokens as the hypothetical starting point for this trial | Compare against the actual length of the sample materials and observe whether citations are truncated |
| Minimum relevance | Set to 0.3 as the hypothetical starting point for this trial | Compare against returned relevance scores and observe whether low-scoring entries are filtered |
| Result reranking | Enable reranking and check rerank scores | Compare against ordering without reranking and observe changes in candidate order |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare against retrieving with the current question alone and observe whether references are completed |
These values are hypothetical starting points for this trial, used to compare against sample material length and actual scores. They are not official defaults.
Three common errors
- Treating a pending status as confirmed. In the sample, the deduction notice status is pending supplier confirmation, while the draft sentence states the supplier has confirmed. The input is the deduction notice text, the comparison field is status, and the observable result is that the status values do not agree.
- Treating a retrieval hit as authorized execution. Retrieval returns a citation array, and that array represents content found; it does not represent an approved claim or an executed deduction. The input is the retrieval output, the comparison object is the authorization status field, and the observable result is that the authorization field is missing or still pending review.
- Linking knowledge bases that use different vector models. The input is the vector model identifier of each knowledge base, the comparison checks whether those identifiers agree, and the observable result is that bases with differing identifiers should not be linked together; record only the identifier difference.
How to verify the setup
- Input the sample quality inspection record text and the draft sentence; compare the subject, batch, amount and status fields; observe whether the draft sentence values agree with the original. If the status claimed in the draft does not match the original, judge it unsupported.
- Input an explicit paper-based empty result array [] and compare whether the downstream node still runs; observe whether the downstream design handles an empty citation as intended. This check does not claim that a fictitious rule retrieval necessarily returns an empty array.
- Input the sample deduction notice text; compare the cited content returned by retrieval against the original wording; observe whether citations correspond line by line to original entries. If cited content differs from the original or an original entry is missing, record the citation source as pending verification.