What the documents look like
Seller ratings and experience score maintenance draws on post-fulfilment consumer feedback: negative-review handling records, experience score breakdowns, service diagnostic reports and rectification ledgers. Inventory procurement and supplier materials form one document class inside that pool. Purchase orders, correspondence with suppliers, arrival records and acceptance records are used to check whether stock shortages, delayed shipment or quality problems relate to the causes attributed in a negative review.
These materials update daily or weekly. A single record commonly carries a supplier name, a material number, an order number, an arrival date, a quantity and an acceptance conclusion. Retrieval requests usually come from customer service supervisors and quality inspection staff seeking the supply-chain facts behind one negative review. Record length varies widely: one acceptance note may run to a few dozen tokens, while a rectification ledger may span several pages. The reference cap is expressed in tokens precisely to accommodate this uneven length.
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 built on different vector models cannot be associated together. The operational step is to check the vector model identifier first, exclude those that do not match, and record what was observed. A cross-model retrieval is not run in order to compare missed recalls.
The search mode determines the recall path. Semantic retrieval relies on vector distance and suits natural-language descriptions of negative-review causes. Full-text retrieval relies on keyword hits and suits material numbers, order numbers and supplier proper nouns. Hybrid retrieval uses both paths and merges them through RRF. RRF is a fusion process and is not written as a toggle.
The minimum relevance score takes effect only in semantic retrieval or when result reranking is used. It is a filtering threshold, not a sorting rule, and it does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external systems needing the same model. Reference content is output as an array whose length may be 0, and the downstream chain still runs when nothing is retrieved.
Hypothetical paper exercise sample: the source text to be checked reads, "Supplier A, material number A-102, purchase order PO-2024-0311, arrival date 11 March 2024, acceptance conclusion: quantity short by 20 units." The draft sentence to be judged reads, "The A-102 arriving on 11 March 2024 from Supplier A passed acceptance at the ordered quantity." The draft and the source take mutually exclusive values on the acceptance conclusion field, for the same subject, the same period and the same field.
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 inventory procurement and supplier materials base | Check the vector model identifier against the ratings materials base; associate together only when the identifiers match |
| Search mode | Hybrid retrieval | Run the same test question through semantic and full-text separately, comparing hits for a material number query and a natural-language query |
| Reference cap | 2000 tokens as the starting point for this exercise | Take one short and one long actual record, observe whether content is truncated, and compare completeness against 1500 and 2500 |
| Minimum relevance score | 0.4 as the starting point for this exercise | Label several relevant and irrelevant samples first, observe whether irrelevant fragments decrease after filtering, and compare against 0.3 and 0.5 |
| Question optimisation | Enabled as the starting point for this exercise | Compare the original question with the optimised question and check whether supplier names and material numbers were altered |
| Result reranking | Enabled as the starting point for this exercise | Confirm a rerank model is available and enabled per the setting, then compare candidate ordering before and after reranking |
Three common errors
- Treating the minimum relevance score as a permission or entity filter. The input is a query aimed at Supplier A, and the comparison object is whether records for Supplier B appear in the returned content. A relevance score reflects semantic closeness only and does not judge record ownership. The observable result is that a highly similar fragment for Supplier B may still be returned even after the threshold is raised, and ownership must be checked separately.
- Forcing knowledge bases built on different vector models into one association. The input is the vector model identifier of each of two knowledge bases, and the comparison is whether the two identifiers match. When they do not match, the correct action is to exclude the non-matching base and record what was observed. No inference is then drawn about missed recalls or redundant returns from associating them.
- Treating RRF as a standalone toggle that can be switched on. The input is the configuration interface for hybrid retrieval, and the comparison object is whether an enable item named RRF exists. The documentation describes RRF as the fusion process for multiple result paths. The observable result is that no such toggle is found in the configuration items, and fusion occurs with hybrid retrieval.
How to verify the setup
- For the draft sentence in the hypothetical sample, the input is the source text "acceptance conclusion: quantity short by 20 units" together with the draft sentence "passed acceptance at the ordered quantity." Compare the values taken on the same acceptance conclusion field. When the two values are mutually exclusive, the draft is judged contradictory. If the source does not state an acceptance conclusion, the draft is recorded as unsupported or pending verification, and is not called a denial.
- For the reference cap, the input is one short acceptance note and one long rectification ledger, and the comparison is whether the returned content retains subject, object and status in full. When the long ledger is truncated and key fields are lost, the cap is low. When the short note returns a large amount of irrelevant fragments, the cap is high. An empty result is checked against the downstream design using an explicit paper input of an empty array, without asserting that a fabricated rule search must return an empty array.
- For question optimisation, the input is the original question "how did the acceptance go for that A-102 batch" and the optimised question, and the comparison covers whether the supplier name, material number and time range were changed. When the optimised question adds a supplier or date absent from the source, the subject is judged altered and optimisation must be disabled or adjusted.
This page covers only platform seller ratings and experience score maintenance, inventory procurement and supplier materials, and knowledge base retrieval and recall. Retrieval or a form does not constitute authorisation. Which materials, operations and requests sit at what authorisation state during an operational handover remains a separate matter, and a pending request is not an approval.