What the documents look like
Multi-store matrix operations revolve around four document types: store authorization lists, price control tables, cross-territory complaint forms, and monthly sales reports. An authorization list records the store entity, platform, authorization start and end dates, and authorization scope. A price control table records channel prices and minimum price floors. A complaint form records the complained entity, batch, and processing stage. A monthly sales report records per-store volume and settlement basis. These documents commonly exist in Chinese, English, or bilingual versions, and the same fact may be worded differently across language versions.
The retrieval step handles name variations of the same entity across documents, differences in the basis of the same price field across tables, and the correspondence between authorization status in a complaint form and in an authorization list. Semantic recall can match across languages, but a match does not carry any authorization determination. Retrieval output is an array whose length may be 0; the empty-result path still completes, and downstream design must accept an empty array.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries with different vector models cannot be associated together. The vector model identifier is checked first, and non-matching libraries are excluded; cross-model retrieval is not run to compare missed recalls.
Search modes include semantic search, full-text search, and hybrid search. Semantic search relies on vector distance and suits natural-language questions. Full-text search relies on keyword hits and suits numbers, models, and proper nouns. Hybrid search merges the two paths through RRF. RRF is a fusion process, not a switch that can be turned on. The citation limit is measured in tokens and controls how much content each search may cite. The minimum relevance score is a 0-1 filter threshold that takes effect only under semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call.
Retrieval does not perform permission or entity isolation. The minimum relevance score is a filter threshold and is not an authorization filter. The optional paths in the overall retrieval flow do not all execute on every request: image retrieval requires image input and matching model capability, and reranking requires a rerank model enabled in settings.
Hypothetical exercise sample
Source text one (authorization list, Chinese): "Store A is operated by Distributor Jia, authorization period 1 January 2025 to 31 December 2025, authorized platform Platform P."
Source text two (cross-territory complaint form, English): "Complaint received on 2025-06-10 regarding Store A, batch B-0610, status is pending review."
Draft sentence to assess: "Store A was within its authorization period in June 2025, and the complaint form has confirmed cross-territory diversion."
Assessment points: the authorization list and the complaint form share the same entity, and the periods overlap, but the complaint form status is pending review and carries no confirmed conclusion. The first half of the draft sentence can be supported by the authorization list; the second half is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Authorization list library and complaint form library, with matching vector model identifiers | Check both vector model identifiers first and exclude non-matching ones; do not run cross-model retrieval |
| Search mode | Hybrid search | Entity names and complaint numbers need keyword hits, while status descriptions need semantic matching; compare recalled content against semantic search alone and full-text search alone |
| Citation limit | A starting value based on actual chunk lengths in the material | Test with authorization lists and complaint forms of different lengths to see whether cited content is truncated, then adjust |
| Minimum relevance score | Label relevant and irrelevant samples first, then set a threshold | Takes effect only under semantic search or result reranking; compare whether irrelevant fragments are filtered at different thresholds |
| Question optimization | Enable for multi-turn follow-ups or short questions | Compare the original question with the optimized question and check whether the entity was altered |
| Result reranking | Enable when candidates are numerous and the question is complete | Compare ordering before and after reranking and confirm whether the text question is complete |
Three common errors
- Treating a retrieval hit as authorization confirmation. An authorization list records an authorization period; retrieving that list only shows the document exists and does not show the store is currently authorized. The authorization start and end dates need to be checked against the query date, and the authorization scope field needs to be confirmed.
- Reading pending review on a complaint form as confirmed cross-territory diversion. The complaint form status field states pending review, and writing it as confirmed in the draft sentence treats an unsupported conclusion as fact. The status field value needs to be checked, and the evidence gap retained.
- Treating the minimum relevance score as a permission filter. The minimum relevance score is a similarity threshold and does not distinguish store entities or authorization scope. Retrieval scope and access permission need to be checked separately, and similarity settings are not relied on to exclude unauthorized records.
How to verify the setup
- Input the authorization list source text and the complaint form source text from the hypothetical sample, and compare the draft sentence "Store A was within its authorization period in June 2025, and the complaint form has confirmed cross-territory diversion" against both source texts. The authorization period covers June 2025, so that part is satisfied; the complaint form status is pending review, so that part is not supported. The observed result is that the first half of the draft sentence can be checked and the second half is marked as pending verification.
- Input a short question containing only a store name, and compare the retrieval terms before and after question optimization is enabled. Observe whether the optimized question fills in the entity and time range, and whether it introduces an authorization conclusion absent from the source text. If the optimized question contains a conclusion not stated in the source text, that optimization result is not used.
- Input a query unrelated to both the authorization list and the complaint form, and compare the returned citation array. Observe whether the return is an array of length 0 and whether downstream handles it as an empty array. A claim cannot be made that an unrelated query always returns an empty array; only the citation content and length actually observed in this instance are recorded.
The retrieval scope needs to cover the target material to which access is held; when it is missing, the evidence gap is recorded. A pending review status does not demonstrate a lack of operating permission, and signing, reaching cooperation, and contract effectiveness are three distinct facts. Fields not stated in the authorization list remain unknown, and authorization or effectiveness clauses are not supplied to fill them in.