What the documents look like
Presale group-buy collection takes orders first and organizes purchasing and fulfillment afterward. The main data sources are mini-program orders, supplier inventory, group-leader feedback, and warehouse-and-delivery route planning. The documents handled day to day are the presale product table, the collection result, the purchasing plan, and the cut-off notice. Specifications, quantities, times, and shipping commitments that appear in product copy need to correspond item by item with these materials. The brief describes the update frequency as daily and per session. That frequency does not change how retrieval behaves on this page, so it is not developed further here.
For copy drafting, the evidence needed is spread across the presale product table, the collection result, the purchasing plan, and the cut-off notice. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. The library is the smallest search unit, and a collection is only a classification layer for managing data. Semantic retrieval relies on vector similarity. Full-text retrieval relies on keyword hits. Hybrid retrieval merges the two paths through RRF. ReRank performs a second ordering of candidate text and returns a score from 0 to 1. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. It is a filtering threshold and does not act as a permission or entity-isolation control. The reference limit controls how much cited content is returned, measured in tokens. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires corresponding image input and model capability, and re-ranking requires an available model and that the setting is enabled. These paths are not all executed on every request.
Hypothetical paper exercise sample: the source material reads "Presale product table: March 10, 2026 session, Gannan navel orange 5-jin pack, presale price 29.9 yuan, order cut-off March 9 at 18:00, expected arrival at warehouse March 12." The draft sentence to be judged reads "This session's Gannan navel orange 5-jin pack is priced at 29.9 yuan presale, orders close March 9 at 18:00, and the goods arrive at the warehouse March 12." The judging task is to check whether each field in the draft sentence matches the source material in the same subject, the same session, and the same field.
Constraints at the retrieval step
The minimum relevance value is a threshold that filters low-relevance search results. It does not exclude unrelated businesses or unauthorized records from the returned set. A relevance score is not a category, entity, or permission filter. Retrieval is not authorization. If a returned fragment belongs to a different session, that observation is recorded as an observation, and no conclusion is drawn that permission failed.
The search node requires that associated knowledge bases use the same vector model. When two libraries use different vector models, they cannot be associated together. The check is to inspect each library's vector-model identifier and exclude those that do not match. A cross-model retrieval is not executed in order to compare missed recalls, and no runtime outcome is asserted for an association that the documentation does not permit.
Question optimization performs coreference resolution and question expansion before retrieval. The optimized question is not the source material. When the subject, session, or field in the optimized question has been altered, that difference is recorded, and no conclusion is drawn that the material itself is wrong.
The reference limit is measured in tokens. It is not the same as a top-k item count. The minimum relevance value takes effect only under semantic retrieval or when re-ranking is used. RRF is a merging process for hybrid results, not a switch that is turned on. Image retrieval runs only when image input is present and the model capability is available. Re-ranking runs only when a model is available and the setting is enabled. The optional paths shown in the overall retrieval flow are not executed on every request.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Presale product table and cut-off notice, two libraries | The two libraries can be associated only if they use the same vector model; inspect the vector-model identifier first and exclude non-matching libraries |
| Search mode | Hybrid retrieval | The copy contains exact fields such as specification and time alongside natural-language phrasing; compare against semantic-only and full-text-only results |
| Reference limit | 800 tokens | A starting value for this exercise only; compare against 400 and 1200 tokens and observe whether key fields are truncated in the cited content |
| Minimum relevance | 0.35 | A starting value for this exercise only; label relevant and irrelevant samples first, then compare whether irrelevant fragments are introduced at different thresholds |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject has been altered |
| Result re-ranking | Enabled | The candidate range after hybrid retrieval is wide; compare candidate ordering before and after re-ranking and observe whether exact fields move forward |
These values are hypothetical starting points for this exercise and are not official default values. The following distinctions are kept. Returning cited content is existing node behavior. Associating knowledge bases, search mode, reference limit, minimum relevance, question optimization, and result re-ranking are actual configurable fields. Checking the fields of the draft sentence against the source material is a manual verification action. Publishing copy and synchronizing orders require separate downstream system implementation.
Three common errors
- Treating the minimum relevance value as a permission or entity filter. The minimum relevance value is a filtering threshold and does not exclude unrelated businesses or unauthorized records. Check: input the same question, compare the returned content at different thresholds, and observe whether fragments from an unrelated session still appear. If they appear, record that observation and do not infer that permission failed.
- Associating libraries that use different vector models. The node requires that associated knowledge bases use the same vector model. Check: inspect each library's vector-model identifier and exclude those that do not match. Do not execute a cross-model retrieval in order to compare missed recalls.
- Treating the optimized question as the source material. Question optimization performs coreference resolution and question expansion. Check: compare the original question with the optimized question, and check whether the subject, session, or field has been altered. If the subject has been altered, record that difference and do not conclude that the material itself is wrong.
How to verify the setup
- Input the hypothetical sample source material and the draft sentence, and compare the presale price, order cut-off time, and warehouse arrival time in the draft sentence against the corresponding fields in the source. The condition is met when all three fields share the same subject, the same session, and the same field, with matching values. If any field is not stated, or only related-stage evidence exists, record it as unsupported or pending verification.
- Input a short question containing only the session and the product name, and compare whether the returned cited content includes the presale product table fragment for that session. The condition is met when the returned content contains the corresponding fragment and it is not truncated. When the return is empty, check the downstream design against an explicit paper input of
[], and do not claim that a fictional rule necessarily returns an empty array. - Input a query containing an image, and compare whether the returned result comes from the image-description path or the image-vector path. The condition is met only when the vector model supports image input or a visual model is configured, and a corresponding index was built at ingestion time. When those conditions are absent, record that the path was not executed, and do not infer a failure cause.
The verification steps above are expected checking methods and were not actually executed. The retrieval checks record observations and failure branches only. They do not guarantee a first-hit match, full recall, or that an unrelated query necessarily returns empty.