What the documents look like
Fresh-food supplier claims pursue quality responsibility for goods on arrival. They do not address ordinary in-store shrinkage, which sits outside this claim process. The working documents are quality exception forms, photographic evidence, deduction notices and claim reconciliation statements. These materials come from the quality inspection system, WMS receiving records, supplier contracts and financial reconciliation. Updates arrive per batch and monthly, so copy drafting must check whether a cited material belongs to the current batch or to a historical version.
This page covers one task: drafting product and marketing copy for fresh-food supplier claims, with citation tracing. The drafting step takes a proposed sentence and asks which material text supports it. The category difference matters here. A claim about arrival quality turns on recorded inspection values for a specific batch, while shrinkage handling turns on store-level loss figures. Copy that describes claim outcomes must therefore trace back to batch-level inspection evidence rather than to general quality language.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model, and it outputs an array of cited content whose length may be zero. Even when nothing is retrieved, the output path still runs. The retrieval output must be handed to the answering step, where each answer is checked line by line against the cited source text. The minimum relevance value is a filter threshold that takes effect only under semantic retrieval or result re-ranking. It does not perform permission or entity isolation, and retrieval itself does not constitute authorization. The same-vector-model requirement applies to the linked knowledge bases and does not extend to business registration sites or other external systems.
Assumed paper exercise sample: Material text one reads "On 12 March 2024, batch A-grade spinach arrived and passed receiving inspection; the quality inspection system records a leaf wilting ratio of 18%, exceeding the 10% ceiling agreed in the contract." Material text two reads "Photographic evidence for this batch is numbered P-0312-07, captured on 12 March 2024 at 09:40, showing the wilting state of the leaves." The draft sentence to be judged reads "This batch of spinach has been inspected and fully meets quality standards, so it can be purchased with confidence." The draft sentence and material text one take mutually exclusive values on the wilting ratio field, so the sentence is unsupported. Material text two records only a wilting state and does not state that the batch fully meets standards, so it is also unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the quality inspection record base and the claim document base | Both bases must share the same vector model; verify the identifier first, then link |
| Search mode | Hybrid retrieval | Quality exception forms carry batch numbers and proper nouns, so semantic and full-text recall are both needed |
| Minimum relevance | 0.5 | Assumed starting point; compare against 0.3 and 0.7 to see whether cited content covers the material text |
| Citation ceiling | 2000 tokens | Assumed starting point; compare 1000 and 4000 to see whether a key field is truncated |
| Result re-ranking | Enabled | When candidates are numerous, re-rank text and use the re-rank score for filtering |
| Question optimization | Disabled | The draft sentence has a clear subject; enabling it adds a model call and latency |
Three common errors
- Treating the draft sentence "fully meets quality standards" and material text one "wilting ratio 18%" as two descriptions of the same field. They take mutually exclusive values on the wilting ratio field, which makes them contradictory, and the sentence must be marked unsupported.
- Treating an empty retrieval output array as a workflow error. The cited content length may be zero, and the empty array must be handled by the downstream step. An empty array does not indicate retrieval failure or a business stop.
- Treating minimum relevance as a permission or entity filter. The value is only a filter threshold. It does not exclude unrelated businesses or unauthorized records, and retrieval does not constitute authorization.
How to verify the setup
- Input material text one and the draft sentence, then compare the value in the wilting ratio field. If the draft sentence claims "fully meets" while the source records 18% against a 10% ceiling, the result is a contradiction and the sentence must be returned for revision.
- Input material text two and the draft sentence, then compare whether the photographic evidence number and capture time support the sentence. If the source shows only a wilting state and does not state "fully meets," the result is unsupported.
- Use a paper input of an empty array to check the downstream design. Observe whether the output path runs and whether the downstream step handles empty citations. Do not claim that a fabricated rule will return an empty array.
Across these checks, the same subject, the same period and the same field must be compared. A contradiction requires mutually exclusive values in one field. A source that does not state a value, or that shows only a related stage, is unsupported or pending verification. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a needed record is missing from the accessible retrieval scope or from a clearly identified manual input, record the gap and leave it open.