What the documents look like
Producer export tax refunds combine the offset of domestic sales tax with the refund of export tax in one calculation. That structure requires the documents to align on the same filing period and the same filing entity. In store service and internal training, the materials most often retrieved are the producer export tax refund declaration detail sheet, export contracts, the sales detail ledger, and export customs declarations. These materials update monthly. A retrieval result that returns only one of them can supply a fragment with no corresponding counterpart.
For knowledge base retrieval and recall, each document carries three qualifiers: period, entity, and field. Between one export customs declaration and the sales detail ledger for the same period, the free-on-board price difference requires manual reconciliation. A retrieval result supplies candidate text and does not replace that reconciliation step. The monthly update frequency of the materials does not by itself justify a batch synchronization or transaction-triggered mechanism in the retrieval configuration.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. The handling step is to check the vector model identifier first, exclude the knowledge bases that do not match, and record the association result actually observed. The retrieval step does not perform permission decisions. A similarity score and the minimum relevance setting act on content relevance only; they cannot be used to isolate unrelated business lines or unauthorized records.
Full-text retrieval and semantic retrieval apply under different conditions. Queries for numbers, models, and proper nouns suit full-text retrieval or hybrid retrieval. Natural-language questions suit semantic retrieval or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a merge process, not a standalone switch. Result reranking requires a rerank model and must be enabled under its setting. Image retrieval requires image input and the corresponding model capability. The optional paths shown in the documented flow diagram do not mean every request executes all of them.
The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call.
Configuration and implementation checks
The values below are starting points for this hypothetical exercise. They are used to compare against neighboring values and are not official defaults.
Hypothetical exercise sample: a producer export enterprise in a store service and internal training setting. The source material text reads: "In this enterprise's July 2024 producer export tax refund declaration detail sheet, the export sales amount field is filled as 1.2 million yuan; the export contract for the same month stipulates a free-on-board price of 1.18 million yuan; the sales detail ledger for the same month records export sales of 1.2 million yuan." The draft sentence to be assessed is: "The enterprise's July 2024 export sales amount has been confirmed at the contract free-on-board price, and no further check against the customs declaration is needed." That draft sentence lacks customs declaration evidence and remains unverified.
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one producer export tax refund material base using the same vector model | Check the vector model identifier first, compare against non-matching bases, and exclude them |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone and full-text retrieval alone for fragment completeness |
| Citation limit | 2000 tokens | Compare against 1000 and 4000 and observe whether detail sheet fields are truncated |
| Minimum relevance | 0.4 | Takes effect only under semantic retrieval or result reranking; compare against 0.3 and 0.5 |
| Question optimization | Enabled | Compare whether the entity is altered across multi-turn follow-up questions |
| Result reranking | Enabled under its setting | Compare candidate text ordering against the disabled state |
Three common errors
- Treating a retrieval score as a permission filter. The input is the same query; the comparison is between the similarity score and record ownership; the observable result is that a high or low score does not decide whether a record belongs to the current entity. The correct step is to check the access scope separately.
- Forcing knowledge bases with different vector models into one association. The input is two knowledge bases with different vector model identifiers; the comparison is the model identifier; the observable result is that a mismatch requires exclusion, with the actual association observation recorded and no inference about missed recall or redundant results.
- Treating the draft sentence under assessment as a verified conclusion. The input is the draft sentence above; the comparison is between the draft sentence and the source material text; the observable result is that when the source text does not state customs declaration information, the sentence is unsupported and cannot be written as a denial or a conflict.
How to verify the setup
- Verification one: Input the hypothetical sample source material text and the draft sentence. Compare the phrase "confirmed at the contract free-on-board price" in the draft sentence against the three source materials. The observable result is that the source text does not provide a customs declaration field, so the sentence is judged unverified and the retrieval output should preserve that gap.
- Verification two: Input "July 2024 producer export tax refund declaration detail sheet export sales amount." Compare the fragments returned by hybrid retrieval against those returned by semantic retrieval alone. The observable result is that hybrid retrieval can hit both the detail sheet field and the contract free-on-board price wording, while semantic retrieval alone may return only semantically similar fragments.
- Verification three: Input a clearly defined empty-result paper input of an empty array. Compare how the downstream design handles a citation content length of zero. The observable result is that the output chain still runs through, without claiming that a fictional rule retrieval necessarily returns an empty array.
This page addresses only producer export tax refunds, store service and internal training, and knowledge base retrieval and recall. When operations staff hand over duties, the authorization state of materials, actions, and requests must be checked separately; retrieval and forms are not authorization execution.