What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against short shelf life, using POS sales, the headquarters ordering system, store inventory and cold-chain temperature records to plan next-day replenishment. This category does not follow the concession-counter model used by large-format fresh departments. The documents handled here are fresh-food order forms, damage and loss reports, near-expiry handling sheets, and temperature records. Updates arrive daily and by shift.
For contract and transaction material comparison, the parties to check are the store and the supplier. The objects are the ordered quantity, received quantity, damage quantity and temperature value for the same batch, the same date range and the same field. Retrieval only brings the relevant materials back. Whether a material has been approved, signed or has taken effect is a separate fact. A retrieval result and a filled-in form do not together constitute authorization.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. During implementation, first check the vector model identifier, exclude the ones that do not match, and record what was actually observed. Do not run a cross-model retrieval to compare missed recalls.
Semantic retrieval, full-text retrieval and hybrid retrieval are compared separately against the test question. No single mode is assumed to suit all fresh-food materials. Hybrid retrieval merges semantic and full-text results through RRF. RRF is a result-fusion process, not a switch that can be turned on. Reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the overall flow diagram do not mean every request executes all of them. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold. It does not perform permission or entity isolation, and it does not exclude unrelated businesses or unauthorized records. The reference cap is measured in tokens; test with materials of different lengths to see whether content is truncated. Question optimization performs coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject has been altered.
Hypothetical paper exercise sample: Material One is a fresh-food order form for a store dated 10 March 2025, stating store number A-07, supplier S-12, 120 units of rice balls ordered, 118 units received. Material Two is the temperature record for the same day, stating a refrigerated cabinet temperature of 4 degrees Celsius, with the recorder noting "reviewed." The draft sentence to be judged is: "Store A-07 ordered 120 units from S-12 on 10 March 2025, and all were received." The pending request is: "Please confirm the damage quantity and near-expiry handling status of this batch of fresh food."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the two knowledge bases for fresh-food order forms and temperature records | Check that their vector model identifiers match before associating; exclude those that do not match |
| Search mode | Start with hybrid retrieval | Compare recalled content for the same test question against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Hypothetical starting point of 2000 tokens | Compare order forms and temperature records of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, compare results before and after filtering, then adjust the threshold |
| Question optimization | Enabled for this exercise | Compare the original question with the optimized question, and check whether the store number, date and supplier were altered |
| Result reranking | Enabled for this exercise | Compare candidate order before and after reranking, and check the relationship between the rerank score and question completeness |
The values above are hypothetical starting points for this exercise, not official defaults.
Three common errors
- Treating a retrieval hit as an authorization conclusion. Retrieval returning the order form and the temperature record only means the materials were recalled. Material Two records only the temperature and a review note; it does not state the damage quantity or the near-expiry handling status. The draft sentence's "all were received" conflicts with Material One's received quantity of 118 units. The damage and near-expiry status remain unsupported.
- Associating knowledge bases built on different vector models. Check the vector model identifier first, exclude those that do not match, and record the observation. Do not infer the runtime result of associating them together.
- Treating minimum relevance as a permission or entity filter. This value filters only by relevance. It does not exclude unrelated businesses or unauthorized records, and retrieval itself does not constitute authorization.
How to verify the setup
- Input the original text of Material One and the draft sentence. Compare the ordered quantity of 120 units with the received quantity of 118 units. Observe whether the system identifies that "all were received" is unsupported. If retrieval does not recall Material One, record it as unsupported rather than as a negation.
- Input the original text of Material Two and the pending request. Compare the temperature record field with the damage and near-expiry fields. Observe whether the damage quantity and near-expiry handling status remain unstated. If retrieval returns an empty array, check the downstream design against the empty reference; do not claim that a rule-based retrieval necessarily returns empty.
- Input the identifiers of two knowledge bases built on different vector models. Compare whether the identifiers match. Observe whether the non-matching one is excluded and recorded. Do not run a cross-model retrieval to compare missed recalls.