What the documents look like
Group-buy leader operations manage community pickup nodes and commission arrangements. The work does not involve cold-chain quality inspection, which belongs to fresh-produce acceptance. The materials handled day to day include leader agreements, commission tables, pickup lists, and community announcements. These come from the community group-buy backend, leader community groups, order details, and after-sales records. In store service and internal training, business development staff and fulfillment sites need to retrieve clauses, commission terms, and pickup requirements from these materials.
Material update frequency is daily or per group. A retrieval configuration must therefore accommodate same-day materials from multiple groups existing side by side. The minimum search unit in FastGPT is the library. A search runs across the whole library, while collections only classify data and do not affect search results.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. Libraries built on different vector models cannot be associated together. The vector model identifier must be checked first, non-matching libraries excluded, and the actual observation recorded. A cross-model retrieval must not be run to compare missed recalls.
Semantic retrieval relies on vector distance. Full-text retrieval suits numbers, proper nouns, and exact terms. Hybrid retrieval merges semantic and full-text results through RRF, then applies ReRank for secondary ordering. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. The reference limit is measured in tokens and controls the length of content returned in a single search. Question optimization handles coreference resolution and question expansion, and adds one model call before retrieval.
The output reference is an array whose length can be zero. Even when no content is found, the output path still completes. The retrieval step does not judge authorization status. Pending review, signed, and effective are distinct facts.
Configuration and implementation checks
Hypothetical drill sample: the source material reads "Leader A submitted an onboarding application on March 1, 2025, with status pending review. The commission table states the fresh category commission rate is 5%, applicable from March 1, 2025. The pickup list shows 20 boxes arriving on March 2." The draft sentence to be judged is "Leader A has passed review and can settle the March 2 pickup commission at 5%." This sentence treats pending review as passed and merges commission applicability with an executed pickup settlement.
| Configuration item | Hypothetical trial starting point | Comparison and check method |
|---|---|---|
| Associated libraries | Associate only libraries containing the leader agreement, commission table, and pickup list that share one vector model | Check each library's vector model identifier; exclude libraries with a different model |
| Search mode | Start with hybrid retrieval | Compare semantic, full-text, and hybrid returns using the sample question |
| Reference limit | Start at a token amount that can hold a full commission-table clause | Compare against actual materials of different lengths to confirm clause completeness |
| Minimum relevance | Mark relevant and irrelevant samples first, then set the threshold | Compare filtering effect only under semantic retrieval or reranking |
| Question optimization | Enable for short follow-ups such as "what about that commission" | Compare the original and optimized question; check the subject was not altered |
The starting values are hypothetical trial settings for this exercise, not official defaults.
Three common errors
- Associating libraries built on different vector models while assuming retrieval still works. The check input is the vector model identifiers of two libraries. Compare whether the two identifiers match. When they do not match, the observable result is exclusion; do not run a cross-model retrieval.
- Treating the minimum relevance score as a permission or entity filter, assuming a low score excludes unrelated business. The check input is the sample question and the filter setting. Compare whether returned content still contains unrelated fragments. The observable result is that the threshold affects relevance ordering only and does not isolate entities.
- Accepting a pending-review draft sentence as already passed. The check input is the source material and the draft sentence. Compare the value of the review status field. The observable result is that the source states pending review while the draft states passed, which is a contradiction.
How to verify the setup
- Verification one: input the sample source material and the draft sentence "Leader A has passed review and can settle the March 2 pickup commission at 5%." Compare the source review status with the draft wording. If the source states pending review while the draft states passed, the sentence is unsupported and must be returned for confirmation. The expectation is that the status contradiction is identified; this has not been tested in practice.
- Verification two: input the identifiers of two libraries that use the same vector model. Compare whether the identifiers match. If they match, they can be associated together; if they do not match, exclude them. Record the actual observation and do not infer a runtime result.
- Verification three: input an explicit paper empty-result array []. Check whether the downstream design accommodates the empty-reference path. Do not claim that a fictional rule search necessarily returns [].