What the documents look like
Chain-store site selection happens before a store opens. The work centres on lease negotiation and engineering handover, and the everyday documents are site evaluation forms, letters of intent to lease, renovation budgets and opening schedules. These materials frequently exist in bilingual or multilingual versions. One letter of intent may be drafted by different regional teams in different languages.
Materials are updated irregularly, one set per store, so retrieval has to locate content by specific store and specific stage. Multilingual mixing makes semantic retrieval's cross-language matching the main reliance, while rent-to-sales ratio calculations, clause numbers and engineering milestone dates need full-text retrieval as a fallback.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. Check the vector model identifier first and exclude those that do not match; do not run cross-model retrieval to compare missed recalls.
Semantic retrieval relies on vector distance and supports cross-language understanding, which suits natural-language questions. Full-text retrieval relies on keyword hits and suits numbers, proper nouns and dates. Hybrid retrieval uses both recall paths and merges them through RRF. RRF is a result-fusion process, not a switch that can be turned on.
The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. It is a filtering threshold, not a permission or entity-isolation mechanism, and retrieval itself does not constitute authorization. The citation limit is measured in tokens and controls how much content each search can cite at most.
Question optimisation adds one model call before retrieval, used for coreference resolution and question expansion, and increases overall latency. Image retrieval requires the corresponding image input and model capability. Reranking requires a model and must be enabled according to configuration. The optional paths shown in the overall flow diagram do not mean every request executes all of them.
Configuration and implementation checks
Assumed paper exercise sample: a chain headquarters development team processes an English letter of intent to lease in March 2025. The source text contains three sentences: "This letter of intent takes effect from the date of signature by both parties. The lessee shall pay a deposit equal to two months' rent within 10 working days after signature. The lessor reserves the right to terminate this letter of intent if the lessee pays late." The same batch contains a Chinese site evaluation form stating: "The preliminary rent-to-sales ratio for this location is 18%, above the internal reference line." The draft sentence to be judged is: "The letter of intent has taken effect, the deposit has been paid, and the lessor may not terminate." This sentence concerns three facts: signature status, payment status and the right to terminate. The source text states only the conditions for taking effect and the reserved right to terminate. It does not state that signature has been completed or that the deposit has been paid.
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the Chinese and English site-selection bases that use the same vector model | Check the vector model identifier first, exclude mismatches, do not associate across models |
| Search mode | Hybrid retrieval | The draft sentence contains natural language and a precise field such as "10 working days"; compare the semantic and full-text recall paths |
| Citation limit | Set to a token amount that covers the full text of one letter of intent | Test with actual materials of different lengths to see whether content is complete, then adjust |
| Minimum relevance | Annotate relevant and irrelevant samples first, then set | This value takes effect only in semantic retrieval or reranking; compare against the annotated results |
| Question optimisation | Enable for multi-turn follow-ups or short questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Use when a reranking model is available and enabled per configuration | Reranking scores are 0-1; compare scores against annotated samples |
Three common errors
- Treating a retrieval hit as authorization or status confirmation. The draft sentence says the letter of intent has taken effect. Retrieving the source text only shows the material exists. Signature and effect status must be checked separately against signature records; retrieval does not confirm that it has taken effect.
- Treating minimum relevance as an entity or permission filter. That value is a filtering threshold. It does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization, and business verification requires downstream implementation.
- Writing RRF as an on/off switch, or moving thinking budget into the citation token limit. RRF is the result-fusion process of hybrid retrieval. The citation limit is measured in tokens. The two cannot be substituted for each other.
How to verify the setup
- Input the English letter of intent source text and the draft sentence from the assumed sample. Compare the draft sentence's claims of "has taken effect, has been paid, may not terminate" against the source text's stated conditions for taking effect, payment obligation and reserved right to terminate. Observe whether the retrieval result returns only the source text without attaching a status conclusion. If the returned content is treated as status confirmation, the condition is not met.
- Input the rent-to-sales ratio sentence from the Chinese site evaluation form. Compare whether the retrieval result hits that sentence and preserves the "preliminary" qualifier. If the result omits the qualifier or rewrites it as a definitive conclusion, the condition is not met.
- Input a paper summary of a letter of intent that does not state a signature date. Compare whether the retrieval output is a clear empty result array [], and check how the downstream design handles an empty result. If the downstream treats the empty result as a negative conclusion, the condition is not met. All of the above are paper checks; no retrieval or deployment was actually executed.