What the documents look like
Presale group-buy order collection takes orders first and then organizes procurement and fulfillment. The working materials for this category are the presale product list, the order-collection result, the procurement plan, and the cut-off notice. These materials arrive from mini-program orders, supplier inventory, group-leader feedback, and warehouse-routing schedules. Platform operations, procurement planning, and supply-chain scheduling all use them. Update frequency is daily or per session.
For document comparison, the same session's product list, order-collection result, and procurement plan need a checkable relationship across subject, object, quantity, and status. The task at the chunking step is to parse the material text into retrievable data chunks so that field differences among the cut-off notice, the procurement plan, and the order-collection result can be surfaced and passed to a human reviewer.
Constraints at the document chunking step
Chunking must preserve document identity and version so a chunk can be traced back to its source text. Conditions, notes, and the content they constrain cannot be split apart. Original data and retrieval indexes are distinct, and one group of original data can correspond to multiple indexes. Retrieval runs as semantic search, full-text search, or hybrid search; hybrid search merges multiple result paths through RRF. ReRank requires a model and must be enabled according to settings. The minimum relevance value takes effect only in semantic search or when result reranking is used, and it is a filter threshold that does not carry permission or entity isolation. The reference cap controls the amount of referenced content in tokens. The same-vector-model requirement for a retrieval library applies only to associations inside that library and does not extend to external systems.
Hypothetical paper test case: the source text for comparison is "March 10, 2025 cut-off notice: this session's presale product is spring bamboo shoots, specification 5 jin per box, cut-off time March 11 at 18:00, procurement placed after cut-off according to the order-collection result." and "March 11, 2025 order-collection result: spring bamboo shoots, specification 5 jin per box, group-formed quantity 120 boxes, status cut off." The draft sentence to be judged is "This session's spring bamboo shoots have completed procurement at the group-formed quantity of 120 boxes, and the supplier has confirmed supply." Judging requires the same subject, the same period, and the same field; the procurement completion and supplier supply confirmation in the draft sentence have no corresponding field or status record in the given source text.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this trial | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Materials contain product names and quantities, so natural-language and field matching both matter; compare hit differences against pure semantic search |
| Reference cap | Hypothetical starting value 3000 tokens for this trial | Material paragraphs are short; used to observe whether referenced content covers complete conditions; compare against a lower value to see whether condition sentences are truncated |
| Minimum relevance | Hypothetical starting value 0.3 for this trial | Takes effect only in semantic search or reranking; compare candidate result counts against a higher value |
| Question optimization | Decide whether to enable based on actual effect | Used for short questions and multi-turn follow-ups; compare retrieval time and hits against the disabled state |
| Chunk integrity check | Manually check whether condition sentences and constraint sentences stay in the same chunk | Check whether cut-off time and procurement conditions are split apart |
| Post-update review | Manually check actual training and index status | A successful upload does not mean retrieval has updated; the actual status needs review |
Three common errors
- Treating a retrieval hit as authorized execution. The input is the draft sentence "This session's spring bamboo shoots have completed procurement at the group-formed quantity of 120 boxes." The comparison object is the group-formed quantity and status field in the order-collection result. The expected result is that the source text has no procurement-completion or supplier-confirmation field, so the sentence is unsupported; a retrieved related material does not mean procurement has been executed.
- Treating a pending-review status as missing operation permission. The input is the cut-off time in the cut-off notice and the cut-off status in the order-collection result. The comparison is whether both belong to the same session and the same product. The expected result is that cut-off status and procurement execution are different facts, and the unverified part keeps its evidence gap.
- Treating minimum relevance as an entity or permission filter. The input is a query about spring bamboo shoots. The comparison is whether returned content includes other products or unauthorized records. The expected result is that a relevance score does not carry isolation; data scope and access permission need separate review.
How to verify the setup
- Input the cut-off notice source text and the draft sentence from the hypothetical sample. Compare whether the cut-off time, product, specification, and group-formed quantity fields match. Observe whether the procurement completion and supplier confirmation in the draft sentence have corresponding records in the source text; when there are no corresponding records, record the sentence as unsupported.
- Input the order-collection result source text from the hypothetical sample. Compare whether the group-formed quantity of 120 boxes and the cut-off status belong to the same session. Observe whether mutually exclusive values exist for the same field; fields not stated are recorded as pending verification.
- Input an explicit paper-based empty result []. Compare how the downstream design handles an empty reference. Observe whether an empty result is conflated with unauthorized or nonexistent status; this check verifies downstream design only and does not claim that a fictional rule retrieval necessarily returns empty.