What the documents look like
Channel diversion reviews address regional and price order, and they do not handle rebate amounts. The materials in this category are diversion complaint forms, anti-counterfeit code records, penalty notices and inspection photographs. These documents commonly carry goods-source codes, shop names, batch identifiers and inspection times. A single goods-source code can appear across several materials, so retrieval has to gather the fragments that concern the same object before any comparison is meaningful. Anti-counterfeit code records and dealer shipment records supply the movement side, while complaint forms and inspection photographs supply the observation side.
Inventory procurement and supplier materials belong to the supporting-evidence group within the same review. They sit in different collections from complaint forms and penalty notices. In FastGPT the smallest search unit is the library, and collections serve only to classify data and do not affect search results. Retrieval therefore operates across the whole library rather than across one collection. The update rhythm of these materials is weekly or irregular, and this page does not depend on any batch synchronization mechanism.
Constraints at the retrieval step
When several materials concern the same goods-source code or the same shop, semantic retrieval can return fragments that are close in meaning while referring to different objects. The vector retriever uses the PG Vector plugin of PostgresSQL with an HNSW index, and MongoDB holds the original data. One group of data can map to several vectors through the indexes field, and when several vectors resolve to the same original data the highest score is kept. Hybrid retrieval combines semantic and full-text recall through RRF, and reranking produces a score between 0 and 1. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used. Retrieval returns reference content as an array whose length can be zero, so the downstream link still runs when nothing is found.
Retrieval returns reference content and does not perform authorization checks. A hit on a penalty notice fragment does not mean the operator may view it or carry out a penalty, and the signing and delivery of a penalty notice remain downstream business actions.
Hypothetical paper exercise sample. The source materials read: "Complaint form A-07 records that goods-source code 8812-04 was sold below the suggested price on 12 March at a shop in East China; the dealer shipment record shows that this code was dispatched to the South China region; the inspection photograph is annotated as taken on 20 March at an East China storefront." The draft sentence to be assessed is: "The goods corresponding to goods-source code 8812-04 were dispatched by the South China dealer and sold at an East China storefront." The retrieval input is the goods-source code together with the shop description. The comparison is between the flow-direction field in the shipment record and the flow direction stated in the draft sentence. If the record does not state that this code entered East China, the draft sentence is not supported and the gap remains an evidence gap.
Configuration and implementation checks
| Configuration item | Setting relevant to this page | How to check |
|---|---|---|
| Associated knowledge libraries | Select libraries containing diversion complaint forms, anti-counterfeit code records, penalty notices and inspection photographs | Check the vector model identifier of each library and exclude those whose models differ |
| Search mode | Compare semantic, full-text and hybrid retrieval against the sample question | Test full-text retrieval with an exact code such as the goods-source code |
| Reference limit | Measured in tokens; the hypothetical starting value for this exercise is 2000 | Compare complaint fragments of different lengths to see whether content stays complete |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant fragments first, then adjust the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a model and runs when enabled in the settings | Check whether the rerank score is used to filter candidates |
The reference limit is set to 2000 tokens as a starting value for this exercise because complaint forms and shipment extracts vary widely in length, and the value should be compared against the observed completeness of long and short fragments. It is a hypothetical starting point, not a documented default. The minimum relevance threshold filters low-relevance fragments and does not exclude unrelated shops or unauthorized records, so it cannot be treated as an entity or permission filter. Question optimization adds a model call before retrieval, which can improve multi-turn retrieval and also increases latency; whether to enable it depends on the actual question and response requirements.
Three common errors
- Treating a retrieval hit as authorization. Entering a query for the penalty record of goods-source code 8812-04 and receiving fragments does not mean the operator is permitted to view those records or to execute a penalty. The signing and delivery of a penalty notice are downstream business actions that require their own implementation.
- Treating minimum relevance as an entity filter. This threshold only removes low-relevance fragments. It does not isolate unrelated shops or unauthorized records, and it cannot be used to conclude that other parties' materials have been separated out.
- Treating a pending state as approved. When a draft sentence or a handover request is still awaiting review, this does not show that the operator lacks permission, and it does not mean that approval has been granted. A pending request is not an approval, and an offered option is not execution.
How to verify the setup
- Enter the goods-source code and shop description from the hypothetical sample. Compare whether the recalled fragments include complaint form A-07 and the shipment record, and observe whether the flow-direction field in the shipment record agrees with the draft sentence. If the record does not state entry into East China, the draft sentence is not supported and this is recorded as an evidence gap. The expected observation is that the recalled fragments contain the two named materials; a missing fragment is recorded as an observed gap rather than assigned a cause.
- Enter the exact code 8812-04 and compare the recall of full-text retrieval against semantic retrieval. Observe whether the code-type query reaches the corresponding fragment. If it does not, record the actual observation without inferring a failure cause. The expected observation is a difference in which fragments each mode returns for an exact identifier.
- Use a paper input of an empty array to check the downstream design, and observe whether the output link still runs. This does not claim that a search under a fictional rule returns an empty array. For libraries using different vector models, check the identifiers first and exclude those that do not match; do not run a cross-model retrieval to compare missed recalls.