What the documents look like
Site selection and store opening happen before a store starts trading. The work centers on lease and engineering delivery, and the documents involved are the site evaluation form, the lease letter of intent, the renovation budget, and the opening schedule. These documents arrive from regional development staff who may submit in different languages. A single lease letter of intent can carry Chinese clauses together with English attachments, so vector retrieval needs to match semantically similar passages across languages.
The site evaluation form mixes footfall estimates and rent-to-sales ratio derivations, with numbers and conclusions in the same block. The lease letter of intent records parties, object, term, deposit, and conditions for taking effect. The renovation budget is itemized by engineering milestone. The opening schedule links dates to responsible persons. Materials update irregularly, per store, and one store produces several document versions across stages, so retrieval has to distinguish version and status.
Multilingual retrieval in this category mainly lets headquarters use a Chinese question to find landlord materials or engineering notes in other languages. The cross-language understanding of semantic retrieval serves that need. Full-text retrieval serves the matching of identifiers, model numbers, and proper nouns.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must use the same vector model. Check the actual model and input type first; knowledge bases with different vector models cannot be linked together. Verify the identifier and exclude those that do not match. One source data block can map to multiple indexes. Check the index and the original content separately. Index content needs to be complete and focused, and retrieval coverage against index length is weighed with samples.
Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion, not only on what happens at query time. After a model or index change, compare cited original text using the same set of questions. Do not promise an outcome from a vector score alone. Semantic recall and relevance scores do not carry permission or entity isolation. The same-vector-model requirement for linked knowledge bases does not extend to business-registration sites or other external systems.
RRF is a fusion step for mixed results, not a toggle. Image retrieval requires an image input and matching model capability. Re-ranking requires a model and must be enabled per its settings. The citation limit is measured in tokens. The minimum relevance score takes effect only in semantic retrieval or when result re-ranking is used.
Configuration and implementation checks
The following is a hypothetical drill sample used to explain configuration and checking. It does not represent a real record or an executed operation.
Sample material text: A chain headquarters development staff member submitted a lease letter of intent on 10 March 2025. The lessor is a local property owner; the lessee is a regional company under headquarters; the object is a ground-floor unit in a shopping mall; the term is five years; the deposit is three months' rent; the letter states "this letter of intent takes effect on the date both parties sign, valid for thirty days." The site evaluation form submitted the same day states a monthly rent-to-sales ratio estimate of 18% for the unit, with footfall data sourced from weekend counts provided by the mall.
Draft sentence to be judged: "The lease letter of intent has taken effect and the deposit clause has been confirmed."
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Linked knowledge bases | Select one or more knowledge bases with the same vector model | The lease letter and site evaluation form must share a model |
| Search mode | Semantic retrieval, full-text retrieval, or hybrid retrieval | Semantic for multilingual clauses, full-text for unit identifiers |
| Citation limit | Maximum tokens cited per search | Set against the sample passage length |
| Minimum relevance | A 0-1 value, effective only in semantic retrieval or result re-ranking | Used to filter low-relevance passages |
| Question optimization | Coreference resolution and question expansion before retrieval | Assess whether to enable for follow-up questions on unit terms |
| Result re-ranking | Requires a re-ranking model and must be enabled per settings | Secondary ordering when candidates are numerous |
Three common errors
- Treating a lease letter of intent awaiting signature as an effective contract. The sample text says it takes effect on the date both parties sign. The draft sentence says it has taken effect. Evidence of completed signing is absent, so that judgment is unsupported.
- Using semantic similarity in place of permission or entity isolation. A high retrieval score does not show that the lessor's material is visible to the current account. Access scope has to be checked separately.
- Treating "valid for thirty days" as a supply or delivery deadline. The letter's validity period constrains the letter itself. Renovation scheduling and delivery milestones are different fields.
How to verify the setup
- For the sample draft sentence, the input is the lease letter text above and the draft sentence. Compare "has taken effect" against the original "takes effect on the date both parties sign," and observe whether any record of completed signing exists. With only the letter text and no signing evidence, the result is pending verification.
- For the deposit clause, the input is the original "the deposit is three months' rent" and the draft "the deposit clause has been confirmed." Compare whether "confirmed" has a confirming action or record. The original states only the calculation method and no confirmation status, so the result is unsupported.
- For empty-result handling, the input is an explicit paper input of []. Compare whether the downstream node proceeds along the path where the citation content length is 0. Observe whether the output is an array of length 0. Do not claim that a fictional rule retrieval necessarily returns [].
All three checks rest on the sample original text and draft sentence and do not indicate an actual test. When material is missing, record the gap rather than supplying an authorization or effectiveness clause.