What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against short shelf life, and it is not handled as a joint-operation revenue share. The working documents for this category are fresh-food order forms, damage and disposal forms, near-expiry handling tables, and temperature records. Updates follow a daily and shift rhythm, and the underlying sources include POS sales, the headquarters ordering system, store inventory, and cold-chain temperature records.
For copy drafting, this means the item, the status, and the date range cited in a proposed sentence must come from the same batch of documents. A record from a previous shift does not stand in for the current order form. This page covers knowledge-base retrieval and recall inside product and marketing copy drafting. It does not extend into identity certification or report analysis.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector model identifiers cannot be associated together. The correct action is to check the vector model identifier, exclude the ones that do not match, and record what was actually observed. Running a cross-model retrieval to compare missed recall is not part of this step.
Semantic retrieval, full-text retrieval, and hybrid retrieval serve different situations. Exact queries such as item numbers and product names can use full-text or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion process, not a switch that can be turned on. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold. It does not perform permission or entity isolation. Referenced content is output as an array, and its length can be 0, which means the downstream chain still runs even when nothing is found. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled by configuration. The optional paths listed in the overall flow do not mean every request executes all of them.
Hypothetical paper exercise sample. Source text one: “Fresh-food order form, 10 March 2025, Store A, item: tuna rice ball, order quantity 120 units, arrival date 11 March 2025, shelf life to 13 March 2025.” Source text two: “Near-expiry handling table, 11 March 2025, Store A, tuna rice ball, status: normal sale, handling deadline 12 March 2025.” Proposed copy sentence to be judged: “This store’s tuna rice balls are made fresh on the day, while stocks last.” The comparison is between “made fresh on the day” in the proposed sentence and the arrival date and shelf life recorded in the sources. The sources do not state a production time, so the wording is not supported. A separate handover request reads, “Please confirm whether Store A’s tuna rice balls have been approved for extended sale.” The sources do not record an extension approval, so this remains pending verification.
Configuration and implementation checks
The table below concerns this sample and lists only settings already described in the official material, or implementation checks. The numeric values are starting points for this hypothetical exercise, and each states what it is compared against.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Two bases, fresh-food order forms and near-expiry handling tables, with the same vector model identifier | Association requires matching identifiers; compare against bases with a different identifier and exclude them |
| Search mode | Hybrid retrieval | Item names need exact hits while status descriptions need semantic matching; compare against semantic retrieval alone |
| Reference limit | Hypothetical starting point of 800 tokens | Compare against the combined length of the order form and near-expiry table text, and observe whether references are truncated |
| Minimum relevance | Hypothetical starting point of 0.35 | Label relevant and irrelevant samples first, then compare whether the near-expiry status sentence survives filtering |
| Question optimization | Enabled | Compare whether the subject is altered across multi-turn follow-ups; a single clear question can be compared with it disabled |
| Result reranking | Enabled by configuration | When candidates are numerous, compare the rerank score against the original order and confirm the model is available |
Three common errors
- Treating minimum relevance as a permission or entity filter. The input is a retrieval question for Store A; the comparison is whether returned content includes records from other stores; the observable point is that a relevance value does not distinguish subject ownership, and out-of-scope content must be blocked by a downstream implementation.
- Forcing knowledge bases with different vector models into one association. The input is two bases with different vector model identifiers; the comparison is whether the node permits associating them together; the observable point is that a mismatched identifier should be excluded, without inferring missed recall or redundancy as an operating result.
- Treating a pending review or a quotation validity period as an approval or a supply term. The input is a handover request reading “extension of sale pending confirmation”; the comparison is whether the sources contain an approval field; the observable point is that an unstated field is recorded as pending verification, and no authorization clause is supplied.
How to verify the setup
- For the proposed sentence “This store’s tuna rice balls are made fresh on the day, while stocks last,” input source text one and source text two, compare “made fresh on the day” against the arrival date and shelf life fields, and observe that the sources state no production time, so the wording is judged unsupported.
- For the handover request “Please confirm whether Store A’s tuna rice balls have been approved for extended sale,” input the near-expiry handling table text, compare whether an extension approval field exists, and observe that it is not recorded, so the item is logged as pending verification with the evidence gap preserved.
- For empty-result handling, input an explicit paper empty reference array [], compare whether the downstream chain continues to execute, and observe that the chain still runs when the reference length is 0, without claiming that a fictional rule retrieval necessarily returns an empty array.