What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and commission-based settlement. The document set for operating reports and exception reviews therefore comes from several sources at once: the joint-operation commission contract, mall activity letters, sales receipts, and monthly settlement statements. Daily and monthly rhythms coexist, but the retrieval task is narrower than the document set. It has to locate the material that answers a specific review question: which contract a sale belongs to, what basis a period's commission uses, whether an activity resource has been confirmed by the mall. Retrieval returns source text; authorization, approval, and settlement execution remain downstream business steps.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection is only a classification layer and does not affect search results. One retrieval combines question optimization, semantic search, full-text search, or hybrid search according to the input and parameters. Hybrid search merges results through RRF. ReRank runs when configured and returns a 0-1 score. The final output passes similarity filtering and a citation limit, and the citation array length can be 0, meaning the downstream path still runs when nothing is found. The minimum relevance value takes effect only under semantic search or when result reranking is used; it is a filter threshold and does not perform entity isolation or permission judgment. Exact queries such as numbers and model codes suit full-text or hybrid search; natural-language questions suit semantic search. Image retrieval requires image input and the corresponding model capability, and reranking requires a model and runs according to configuration. Both are optional paths.
Hypothetical paper exercise: A brand concession signs a joint-operation commission contract with a mall, contract number HYP-2024-031, valid from 1 March 2024 to 28 February 2025, with the commission rate calculated at 18% of the month's tax-inclusive sales. A mall activity letter states that a discount activity from 1 May 2024 to 5 May 2024 is subsidized by the mall, and the concession settles at the original price. The monthly settlement statement shows tax-inclusive sales of CNY 620,000 for May 2024 and commission fees of CNY 111,600. Draft sentence to be judged: "The concession's May 2024 commission fee has been settled according to the basis adjusted by the activity letter." The review scope is limited to these three materials. Signing, approval, and settlement status are not stated.
Configuration and implementation checks
The table below lists settings or implementation checks tied to this feature and primary task for this sample. The values are hypothetical starting points for this exercise and are not official defaults.
| Configuration or check item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Joint-operation commission contract base, mall activity letter base, monthly settlement statement base | The three bases must use the same vector model; check the vector model identifier first and exclude inconsistent ones. Do not associate across models. |
| Search mode | Hybrid search | The sample contains a contract number and a natural-language question; compare semantic, full-text, and hybrid paths on the same question. |
| Citation limit | Starting point 2000 tokens | The unit is tokens; compare whether citation content is complete using materials of different lengths, then adjust. |
| Minimum relevance | Starting point 0.5 | Takes effect only under semantic search or result reranking; label relevant and irrelevant samples first, then compare filtered results. |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare the original question and the optimized question to check whether subject and period were altered. |
| Result reranking | Enabled according to configuration | Requires a rerank model; compare candidate ordering and 0-1 scores before and after enabling, without presuming improvement. |
Three common errors
- Treating a retrieval hit as authorization. A retrieval returning contract text only means the material was recalled. It does not mean the contract was signed, took effect, or that settlement is complete. Signing, cooperation reached, and contract effectiveness are three distinct facts.
- Treating minimum relevance as a permission or entity filter. That value is a filter threshold and does not exclude unrelated businesses or unauthorized records. Retrieval does not perform authorization judgment.
- Treating optional paths as executed on every run. Image retrieval requires image input and model capability; reranking requires a model and runs according to configuration. The sample is a plain-text question, so it cannot be inferred that image vector retrieval ran.
How to verify the setup
- Input the three source materials above and the draft sentence. Compare the draft sentence's claim that the fee "has been settled according to the basis adjusted by the activity letter" against the contract number, validity period, commission rate, activity period, and statement amount recorded in the source text. The source text does not state settlement completion status, so the claim is unsupported.
- Input the question "What is the commission rate for HYP-2024-031 in May 2024?" Compare whether the citation content returned by hybrid search and full-text search contains the contract number and the 18% field. If the citation content lacks that field, record it as not recalled and do not infer a cause.
- Input a paper-based empty result array [] and check whether the downstream design continues to execute according to the path where citation length can be 0. This check verifies empty-citation handling and does not mean any actual query will necessarily return an empty result.