What the documents look like
Store cold-chain receiving for fresh products turns on four recurring documents: cold-chain temperature records, arrival inspection sheets, rejection notices, and supplier delivery notes. Compared with fruit purchasing, this work centers on transport temperature and arrival batch rather than origin price inquiry. The materials update per batch and move daily. Errors carry consequences such as spoiled ingredients, supplier disputes, and regulatory penalties. The comparison task requires checking these receiving materials line by line against contract terms. Citation tracing requires every comparison conclusion to point back to a specific source text and its version.
Retrieval output is not authorization. The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. That array must pass to the answering step, where each answer is checked against the cited original text. Source display is checked against traceable samples; a fluent answer or a visible thinking process does not replace source evidence. Document timeliness is tested with current and historical versions, and whether a citation is adopted still depends on review against the effective version. Knowledge bases built on different vector models cannot be associated together; the identifier is checked first and non-matching bases are excluded.
Constraints at the citation tracing step
A hypothetical drill sample is given here, clearly labelled, for checking purposes only. The source texts to be checked are: cold-chain temperature record LD-2025-0117-A, stating "arrival time 2025-01-17 06:20, carriage temperature 2.8 degrees Celsius, unloading temperature 3.1 degrees Celsius"; arrival inspection sheet YS-2025-0117-03, stating "inspection conclusion: pass, inspector: Zhao, date 2025-01-17"; and supplier delivery note SS-2025-0117-12, stating "item: chilled beef, quantity 200 kg, batch B20250116." The draft sentence to be judged is: "This batch of chilled beef passed inspection, the temperature record is complete, and storage is approved." The check compares whether the draft sentence agrees with the original text of the three materials and whether each citation traces back to the corresponding document number. No real customer, company, or identity number is used.
The search node outputs citations as an array, and a zero-length array still lets the downstream path run. An empty array is not an error termination and is not approval. Similarity scores and minimum relevance are filters on retrieval results; they do not isolate suppliers, entities, or permission scopes. A pending status is not an effective contract. Signing, reaching cooperation, and contract effectiveness are three distinct facts. Where an effective clause is not stated, the evidence gap stays open rather than being filled by assumption. Whether a citation is adopted depends on review against the effective version, not on the retrieval score alone.
The overall retrieval flow combines optional paths, and not every request exercises all of them. Image retrieval requires an image input and a model with the matching capability. Re-ranking requires a re-rank model and must be enabled by configuration. RRF is a fusion process for mixed results and is not a switch to be turned on. Thinking budget is a reasoning budget and is not the citation token limit.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the four bases for cold-chain temperature records, arrival inspection sheets, rejection notices, and supplier delivery notes | Check whether the four bases use the same vector model; exclude bases with a different model |
| Search mode | Hybrid retrieval | Compare with semantic retrieval and full-text retrieval, observing whether recalled content covers document numbers and temperature values |
| Citation limit | Set to 2000 tokens as the hypothetical starting point | Compare with 1000 and 3000, observing whether citations fully contain the three source texts |
| Minimum relevance | Set to 0.5 as the hypothetical starting point | Compare with 0.3 and 0.7, observing whether fragments unrelated to receiving are filtered out |
| Question optimization | Enabled | Compare with disabled, observing whether document numbers or batch information are completed in multi-turn follow-ups |
| Result re-ranking | Enabled | Compare with disabled, observing whether re-rank scores help order the receiving materials |
These starting values are drill settings, not documented defaults. The citation limit uses the documented unit of tokens. The minimum relevance value is a 0-1 filter threshold and applies only in semantic retrieval or when result re-ranking is used. Re-ranking produces a 0-1 score for relevance and can be used for filtering. The knowledge base search node can associate only knowledge bases with the same vector model; this requirement does not extend to business registration sites or other external systems.
Three common errors
- Treating retrieval output as authorized execution. Input: the citation array returned by the search node. Comparison: whether the cited content agrees with the contract terms. Observable result: whether a storage action is triggered before manual review is complete. When the citation array length is zero, the workflow still runs, and this must not be read as authorization.
- Treating minimum relevance as a permission or entity filter. Input: minimum relevance 0.5. Comparison: whether receiving materials from different suppliers are excluded by the similarity threshold. Observable result: whether a low-scoring record is wrongly treated as inaccessible. A relevance score does not perform permission isolation.
- Treating a pending-review status as an effective contract. Input: an arrival inspection sheet marked "pending review." Comparison: whether the contract terms are already effective. Observable result: whether pending review is misread as approved. Signing, reaching cooperation, and contract effectiveness are separate facts; where no effective clause is stated, the evidence gap remains.
How to verify the setup
- Input the three source texts from the hypothetical sample and the draft sentence. Compare the draft phrase "passed inspection, the temperature record is complete" against the temperature record values 2.8 and 3.1 degrees Celsius and the inspection sheet conclusion "pass." Observe whether the citation traces back to LD-2025-0117-A, YS-2025-0117-03, and SS-2025-0117-12. If the citation array is empty, check whether the downstream design handles an empty array.
- Input the supplier delivery note and a second batch inspection sheet from the hypothetical sample. Compare whether the same field, such as batch number, shows mutually exclusive values. Observe whether a case with no stated value or only related-stage evidence is marked as pending verification. A contradiction requires the same subject, the same period, and the same field with clearly exclusive values.
- Input the "pending review" status on the arrival inspection sheet from the hypothetical sample. Compare whether the contract terms are already effective. Observe whether pending review is misjudged as approved or rejected. Pending review does not demonstrate a lack of operating permission; record the evidence gap and keep the manual review step.
The materials needed for verification must sit within an accessible target retrieval scope or in a clearly defined manual input. Where they are missing, record the gap. Retrieval checks describe what is observed and how a check fails; they do not guarantee a first hit, full recall, or an empty result for an unrelated query. An empty-result path can be checked with a clearly stated paper input of an empty array against the downstream design, without claiming that a fictional rule always returns an empty array.