What the documents look like
Site selection and store opening at a chain headquarters take place before a store trades, and the core work is lease negotiation and engineering delivery. The working materials are site evaluation forms, letters of intent for lease, fit-out budgets, and opening schedules. The underlying data sources are trade-area footfall, map points of interest, historical store-type data from the point-of-sale system, and landlord documents. In a cross-border setting these materials circulate in more than one language. A single letter of intent may exist as a Chinese text and a local-language text, and the clause wording and the date formats do not match between them.
Citation tracing answers a narrow question: can a retrieval result be returned to a specific sentence in a specific material, and can each conclusion in an answer be matched to wording in the source text. Materials are updated at irregular intervals, once per store, so several versions of the same store's documents can remain in the knowledge base across negotiation stages. Retrieval must establish which version a citation points to.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model as the retrieval scope. Cited content is output as an array, and the length may be zero, which means the output path still runs even when no content is found. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. A ReRank model re-orders candidate text and produces a score from 0 to 1. The minimum relevance value takes effect only in semantic retrieval or when result re-ranking is used, and it is a filter threshold. The citation limit controls the maximum amount of content cited per search, measured in tokens.
These mechanisms mean that a retrieval hit is not an authorization confirmation, and a similarity score does not perform entity isolation or permission filtering. Knowledge bases with different vector models cannot be associated together, so identifiers should be checked first and non-matching bases excluded.
Hypothetical drill sample: a chain headquarters development staff member holds a letter of intent for lease. The Chinese version states that the lessor shall complete fire acceptance inspection before 1 March 2025, and that if this is overdue the lessee may terminate the letter of intent. The local-language version of the corresponding clause states that the lessor shall complete fire acceptance inspection before 1 April 2025. A fit-out budget table is also on file and notes that the budget was prepared on the basis of the February 2025 version of the opening schedule. The draft sentence to be assessed is: "The letter of intent has confirmed that the lessor is overdue, and the lessee has exercised the right to terminate." The request to be assessed is: "Please confirm from the original text of the letter of intent whether the right to terminate has been exercised."
Configuration and implementation checks
| Configuration or check item | Starting point for this hypothetical trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the knowledge bases holding this project's materials that share the same vector model | Check the vector model identifier of each base first, exclude bases with a different identifier, and do not run retrieval across models |
| Search mode | Hybrid retrieval | Clause numbers and dates are exact queries, and natural-language phrasing is a semantic query, so both need coverage |
| Citation limit | Set to a token amount that can hold the key passages of the letter of intent and the budget table | The starting point is estimated from the passage length in the sample materials, then compared with the length of cited content actually returned |
| Minimum relevance | Set to 0.5 as the starting point for this trial | This value takes effect only in semantic retrieval or result re-ranking, and is compared with the distribution of re-ranking scores |
| Cited content output | Accept the array result and compare each item against the source text | The array length may be zero, and when it is empty the downstream path should still be checked for continuity |
| Material version | Keep multiple versions for the same store and label each with a date | Search against the current version and a historical version separately and confirm which version a citation points to |
Three common errors
- Treating a retrieval hit as an authorization confirmation. Input a retrieval result for the letter of intent, compare the retrieval output with the original wording, and observe whether the answer turns "the lessor shall complete fire acceptance inspection before a given date" directly into an exercised right to terminate. Retrieval returns cited content only; whether the right to terminate has been exercised requires a manual check against the original text.
- Associating knowledge bases with different vector models. Input two knowledge bases with different vector model identifiers, compare whether the identifiers match, and observe whether the system permits joint association. When identifiers differ, exclude the base rather than running cross-model retrieval to compare missed recalls.
- Treating a similarity score as a permission or entity filter. Input a material that is semantically close to the question but belongs to another store, compare the score with the material's ownership, and observe whether the material is cited because of a high score. The score does not perform permission filtering, and ownership must be checked manually.
How to verify the setup
- For the draft sentence in the hypothetical drill sample, input the original text of the Chinese version and the local-language version of the letter of intent, compare the fire acceptance inspection date field in the two versions, and observe whether the answer points out that the dates differ and keeps the conclusion pending verification. If the answer directly confirms that overdue performance has occurred, the condition is not met.
- For the request to be assessed, input the original text of the letter of intent and the request text, compare whether the original text contains wording that the right to terminate has been exercised, and observe whether the answer marks the status of the right to terminate as unsupported. The original text contains only the authorization wording that the lessee may terminate, and does not state that it has been exercised; missing evidence should remain missing.
- Input an explicit empty array as the cited content output, compare whether the downstream node still runs, and observe whether the workflow stops because of the empty array. This check verifies the design for accepting an empty result and does not claim that a fictional rule retrieval will necessarily return an empty array.
These checks address the citation tracing step only. Retrieval and form collection do not constitute authorized execution. Signing, reaching cooperation, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period, and a pending review does not demonstrate a lack of operating permission.