What the documents look like
Group-buy leader operations revolve around community nodes and commission settlement. The working materials are leader agreements, commission tables, pickup lists and community announcements. Copy drafting for products and marketing draws on the rules and wording recorded in these materials, for example which product class a commission rate applies to, or the participation scope stated in an activity announcement. The materials update daily and per group, and the same commission table can differ between groups, so retrieval is bounded by a specific group and a specific date. Whether a promotional sentence in a draft holds depends on whether the corresponding material states that subject, that object and that effective range. The retrieval step only surfaces candidate source text; the judgement still returns to the source text.
A hypothetical paper exercise sample: in one platform's leader operations materials, Article 3 of the leader agreement states "the commission rate for fresh products under this agreement is 8%, effective from 1 March 2025 to 31 March 2025"; the commission table for the same period states "March fresh tier commission 8%, household tier 5%"; a community announcement states "on 15 March the pickup point self-collection time is adjusted to 18:00 to 20:00". The draft sentence to be assessed is "fresh products in this group are settled at 8% commission, valid long term". When checking that sentence, the comparison objects are the effective period in Article 3 and the tier wording in the commission table. The phrase "valid long term" has no support in the source text.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded. Retrieval output is an array of cited content whose length can be 0, and the empty-result path still runs through, so downstream design must accept an empty array. Semantic retrieval, full-text retrieval and hybrid retrieval are three selectable paths. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process, not a standalone switch. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Image retrieval requires the corresponding image input and model capability, and reranking requires a model and must be enabled in settings. The optional paths shown in the overall flow diagram do not mean every request executes all of them.
For copy drafting, the retrieval step supplies candidate source text and nothing more. A high similarity score does not mean the current leader, group or product is authorised to cite that passage. An unsigned termination is not a completed termination, a pending request is not an approval, and an offered option is not an execution. Where a material field is not provided, the value remains unknown rather than negative.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Leader agreement base, commission table base, community announcement base | All three use the same vector model; check the vector model identifier one by one and exclude any that differ |
| Search mode | Hybrid retrieval | The draft sentence contains terms such as "fresh" and "commission" and also needs semantic matching; compare candidate coverage against semantic-only and full-text-only retrieval |
| Citation limit | 800 tokens | Starting point chosen to cover agreement clauses and commission table rows; compare against 400 and 1200 to see whether cited content is truncated |
| Minimum relevance | 0.35 | Starting point chosen to filter low-relevance fragments; compare against 0.2 and 0.5 for retention of relevant and irrelevant samples |
| Question optimisation | Enabled | The draft sentence contains references and omissions; compare the original question and the optimised question to check whether the subject was altered |
| Result reranking | Enabled | Second-pass ordering when candidates are numerous; compare the ranking position of the same source text before and after reranking |
These values are hypothetical starting points for this paper exercise, not official defaults. Each needs a stated reason and a comparison method before any adjustment is considered.
Three common errors
- Treating "valid long term" as already stated in the agreement. Article 3 states an effective period from 1 March 2025 to 31 March 2025, and the draft phrase "valid long term" does not match that period. The phrase is unsupported, and it cannot be written as though the agreement confirmed a long-term validity.
- Treating minimum relevance as a permission filter. Minimum relevance is a similarity threshold and does not distinguish leader, group or product ownership. A fragment scoring high does not mean the current subject is authorised to cite that fragment.
- Treating an empty result as an error. The cited content array can have length 0, and the empty-result path still runs through. Downstream must accept an empty array, and it cannot be asserted that a given retrieval necessarily returns content.
How to verify the setup
- Input the Article 3 source text from the hypothetical sample together with the draft sentence "fresh products in this group are settled at 8% commission, valid long term", and compare the period wording in the draft against the effective period in the source. If retrieval returns Article 3 and the draft still says "valid long term", the phrase is unsupported. If the returned content does not include that clause, record an evidence gap.
- Input the draft sentence and compare the candidate source sets returned by hybrid retrieval against semantic-only and full-text-only retrieval. If hybrid retrieval does not cover Article 3, check whether the vector model identifiers of the associated knowledge bases match. Where an identifier differs, exclude that base and record the actual observation, without inferring a cause of missed recall.
- Input a paper question for which no corresponding material exists, and check whether the cited content is an array whose length can be 0, and whether downstream accepts an empty array. If downstream treats the empty array as an error and halts, record a downstream design gap. Do not claim that this question necessarily returns an empty array.
Across all three checks, the judgement must use the same subject, the same period and the same field. Only mutually exclusive values in the same field constitute a contradiction. A field that is not stated, or that carries only related-stage evidence, is unsupported or pending verification, not a negation or a conflict. Missing evidence remains missing.