What the documents look like
This page covers knowledge-base retrieval and recall for the inventory purchasing and supplier records used in supermarket supplier entry settlement. In this category, settlement usually runs through purchase-and-sale or consignment payables, and a resident in-store operation is not required. The records involved are purchase contracts, delivery acceptance notes, rebate agreements, and reconciliation statements. These records are updated weekly and monthly, and retrieval must separate values by the same supplier, the same period, and the same field. Update frequency only affects how current a record version is; this page does not derive batch synchronization or real-time triggers from it. Retrieval output is reference content only and does not constitute authorization, approval, or confirmation of payables.
A hypothetical paper exercise sample is used throughout. A supplier portal export of a delivery acceptance note summary reads: "Supplier: Huayue Daily Distribution; receiving store: Chengbei Store; delivery date: 10 March 2025; accepted quantity: 120 cases; acceptance status: accepted." A reconciliation statement summary from the same batch reads: "Supplier: Huayue Daily Distribution; settlement period: March 2025; accepted quantity: 120 cases; reconciliation status is pending supplier confirmation." A draft sentence to be checked reads: "Huayue Daily Distribution Chengbei Store March 2025 delivery accepted quantity is 120 cases, and reconciliation is complete." This sentence turns a pending confirmation into a completed one and is unsupported.
Constraints at the retrieval step
Retrieval uses the library as the smallest search unit; collections serve only as classification management. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Libraries with inconsistent vector model identifiers cannot be associated together: verify the identifiers first and exclude those that do not match, and do not run cross-model retrieval to compare missed recalls. When the vector store is Milvus (≥ 2.5.16), full-text retrieval uses BM25; other vector stores use MongoDB $text with jieba tokenization. Semantic retrieval, full-text retrieval, hybrid retrieval, result reranking, and image retrieval are optional paths and occur only when the corresponding input and model capability exist. The reference limit is measured in tokens, its length can be 0, and the output chain still runs. The minimum relevance score takes effect only in semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. RRF is a fusion process for mixed results and is not a switch. Question optimization performs coreference resolution and question expansion and adds one model call.
Configuration and implementation checks
The table lists only documented settings or implementation checks relevant to this page's sample. Where a numeric value is needed, it is marked as a hypothetical starting point for this exercise, with the reason for choosing it and the comparison object.
| Configuration item | Hypothetical value for this exercise | Verification method and basis |
|---|---|---|
| Associated knowledge bases | Supplier records library, settlement records library | The vector model identifiers of the two libraries must match; exclude any that do not, and do not associate them together |
| Search mode | Hybrid retrieval | The sample contains terms such as "Huayue Daily Distribution" and "120 cases"; compare semantic recall and full-text recall together |
| Reference limit | Hypothetical starting point 2000 tokens | A starting point for this exercise only; adjust after comparing the length difference between reconciliation statements and acceptance notes |
| Minimum relevance score | Hypothetical starting point 0.5 | A starting point for this exercise only; label relevant and irrelevant samples first, then compare filtering effects |
| Question optimization | Enabled | In multi-turn follow-ups such as "what about that batch", compare whether the subject in the original question and the optimized question is altered incorrectly |
| Result reranking | Enabled according to settings | When candidates are numerous, compare the 0-1 reranking score with the vector score and confirm whether ordering changes |
Three common errors
- Reading "reconciliation status is pending supplier confirmation" as completed reconciliation. The input is the reconciliation statement summary above; compare the value of the "reconciliation status" field and observe whether the draft sentence turns pending confirmation into completed. That sentence is unsupported.
- Associating knowledge bases with different vector models directly. The input is the model identifiers of the two libraries; compare whether the identifiers match and observe whether cross-model association occurs. When identifiers do not match, exclude them and do not infer missed recalls or redundant results.
- Treating the reference limit as top k. The input is an acceptance note and a reconciliation statement with a clear length difference; compare whether the returned content is truncated and observe that the reference limit is measured in tokens. When truncation appears, adjust that value and do not change it to a fixed item count.
How to verify the setup
- Input the delivery acceptance note summary and the reconciliation statement summary above; compare the value of the same field "accepted quantity" for the same supplier and the same period, and observe that both are 120 cases. The phrase "reconciliation is complete" in the draft sentence is inconsistent with "pending supplier confirmation" and is judged unsupported.
- Input the model identifiers of the two libraries above; compare whether the identifiers match and observe whether joint association is allowed. If they do not match, record the gap and exclude them, and do not run cross-model retrieval to compare missed recalls.
- Input a paper-based empty result []; compare whether the downstream design accepts a reference length of 0 and observe whether the output chain runs. Do not claim that a fictional rule retrieval necessarily returns [].
This page addresses only supermarket supplier entry settlement / inventory purchasing and supplier records / knowledge-base retrieval and recall. Whether materials, operations, or requests are in a given authorization state during operational handover cannot be treated as authorized execution through retrieval or a form.