What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand. Store operations are not administered through a franchise contract, so the material set differs from franchise supervision work. Daily sales reports, shift schedules, display photographs and stock transfer orders update on a daily or weekly cadence, drawn from the POS system, member CRM, ERP inventory and store inspections.
Contract and transaction material comparison requires placing materials that describe the same transaction, the same period and the same field side by side. A daily sales report records sales quantity and amount. A stock transfer order records transfer quantity and a document number. The product code and date range on the two materials need to agree before a field-level comparison is meaningful. Display photographs and shift schedules supply store-state background; they do not by themselves constitute evidence of a transaction amount.
Because these materials update frequently, fixed-length chunking can split a document number, a product code or a date range across separate data blocks. A retrieval pass then returns only half of the qualifying condition, and the comparison loses its complete basis. Chunking therefore has to keep a condition, its explanation and the content the condition constrains inside one block.
Constraints at the document chunking step
In FastGPT a knowledge base is composed of libraries, collections and data. The smallest search unit is the library; a search runs across the whole library, and collections serve classification management. In the MongoDB dataset.datas table, the original vector data is stored, and an indexes field records the corresponding vector IDs as an array, so one group of data can map to multiple vectors. In PostgresSQL a vector field stores the vectors, with HNSW as the index. During retrieval the vectors are recalled first, then the vector IDs are used to find the original data in MongoDB. Records corresponding to the same original data are merged, and the highest score among them is taken.
Chunking needs semantic completeness and a single meaning. Full-text retrieval suits exact queries such as numbers, models and proper nouns. Semantic retrieval suits natural-language phrasing. Hybrid retrieval combines multiple result channels through RRF, and Rerank requires a model and must be enabled by configuration. Retrieval is a recall process; it does not carry permission or entity-isolation duties. A relevance score is not a category, entity or permission filter.
The following is a hypothetical drill sample. It does not involve a real customer or a legitimate identity number.
Material A original text: "1 March 2026 to 31 March 2026, store D07, product code SP-2201, transfer quantity 120 units, transfer order number DB-20260315-07, status is pending regional manager review."
Material B original text: "1 March 2026 to 31 March 2026, store D07, product code SP-2201, daily sales report cumulative sales quantity 95 units, report status is submitted."
Draft sentence under review: "Store D07 confirmed both the transfer quantity of 120 units and the sales quantity of 95 units for this product in March 2026."
Review points: the two materials share the same subject, the same period and the same product code, while the quantity fields hold different values. Material A carries the status pending regional manager review; Material B carries the status submitted. The draft sentence states that both items are confirmed, which does not match the statuses recorded.
Configuration and implementation checks
| Configuration or check item | Setting applicable to this page | Basis and method of checking |
|---|---|---|
| Search mode | Hybrid retrieval | Transfer order number DB-20260315-07 and product code SP-2201 are exact fields and need full-text recall to participate |
| Reference limit | Hypothetical trial starting point of 800 tokens | Material A and Material B each contain period, store, code, quantity and status; two complete passages must fit; compare with a 400-token starting point to observe whether the status field is truncated |
| Minimum relevance | Hypothetical trial starting point of 0.35 | This value takes effect only under semantic retrieval or result reranking; compare with a 0.5 starting point to observe whether the pending-review status passage is still recalled |
| Result reranking | Enabled according to configuration | Performs secondary ordering of text when candidate results are numerous; when not enabled, record that this path was not executed |
| Question optimization | Decided by observed effect | A single clear question can leave it off; enabling it adds one model call and additional latency |
| Chunk boundary check | Manual review of passage boundaries | Check whether period, store, code, quantity and status remain in the same data block |
Three common errors
- Reading "pending regional manager review" in Material A as confirmed. A pending review is not an approval. The draft sentence states that both items are confirmed, and approval completion evidence is absent. During checking, look for an approval-completion field in the original text of Material A; when the text does not record it, log the item as unsupported.
- Treating transfer quantity 120 units and sales quantity 95 units as a contradiction in the same field. The two materials use different field names and different quantity scopes. A contradiction exists only when the same subject, the same period and the same field hold mutually exclusive values. During checking, confirm the field name and the statistical scope first; when the scopes differ, log the pair as not directly comparable.
- Treating retrieval recall as an authorization or effectiveness determination. Vector recall and relevance scores do not perform permission filtering, and recalling Material A does not mean the review has passed. During checking, keep retrieval results separate from business status fields, and take the status from the original text of the material.
How to verify the setup
- Input the original text of Material A and the original text of Material B, then compare the draft sentence "the transfer quantity of 120 units and the sales quantity of 95 units are both confirmed" against the status fields of the two materials. Material A records pending regional manager review, and Material B records submitted; neither records a completed confirmation. The observable result is that the draft sentence does not match the recorded statuses and should be changed to pending verification.
- Input transfer order number DB-20260315-07 for full-text retrieval, then compare whether the recalled passage contains that number together with its corresponding status field. The observable result needs to show the number and the status inside the same data block. If the status field falls into an adjacent block, record a chunk boundary problem and do not determine that the review has passed.
- Input the period "1 March 2026 to 31 March 2026" and store "D07" for hybrid retrieval, then compare whether the recalled results cover both Material A and Material B. The observable result needs to show both materials recalled with complete status fields. If only the quantity field is recalled and the status is missing, record a reference-limit or chunking problem and preserve the evidence gap. For the empty-result case, a paper input of
[]can be used to check the downstream design; no claim is made that a fictional rule retrieval necessarily returns an empty result.
This page addresses only brand-owned store operations, contract and transaction material comparison, and document parsing and chunking. During operational handover, the material, action or request status and its authorization state must be distinguished; retrieval and forms collect values, and business verification, approval, database writes and external synchronization require their stated downstream implementation.