What the documents look like
Brand outlet operations exist to clear out-of-season inventory. The working documents are discount approval records, slow-moving inventory lists, price tags, and gross margin calculation sheets. Their source data comes from ERP inventory age, POS sales, outlet mall promotions, and historical discount levels, with updates arriving weekly or monthly. This shapes retrieval in a specific way: the questions that matter here name exact SKUs, discount rates, approval states, and margin bands. Precise matching of an identifier or a status field matters more than broad semantic generalization. A retrieval setup for this category has to hold both natural-language questions and identifier-style questions, and those two question types place different demands on the search mode.
Constraints at the retrieval step
A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together. At setup time, verify the vector model identifier of each library, exclude those that do not match, and record what was actually observed. Do not run a cross-model search to compare missed recalls.
One retrieval pass combines semantic search, full-text search, hybrid search, question optimization, and reranking according to the input content and search parameters. Hybrid search merges multiple result paths through RRF; RRF is a fusion process, not a switch that can be turned on. Reranking requires a model and must be enabled in settings. Image retrieval requires image input and the corresponding model capability. These optional paths are not all executed on every request. The reference limit is measured in tokens. The minimum relevance score takes effect only in semantic search or when result reranking is used. A similarity score does not perform permission or entity isolation, and a retrieval result does not constitute any authorization.
Hypothetical paper exercise sample: the source material reads "SKU A-1024, item 7 on the March 2025 slow-moving inventory list, current discount approval status is pending review. Applicant: discount approval form D-2025-0312, submission date March 10, 2025." A handover request to be judged reads: "A-1024 has completed discount approval and price tag replacement may proceed." Checking against the material, the material shows only pending review and contains no record of completed approval. The judgment is unsupported, and the evidence gap is preserved. No approval or effective clause is supplied to fill it.
Configuration and implementation checks
| Configuration item | Relation to this page's documents | Assumed starting point and comparison method for this exercise |
|---|---|---|
| Associated knowledge bases | Discount approvals, slow-moving lists, price tags, and margin sheets are built as separate libraries | Verify the vector model identifier of each library first, exclude non-matching ones, and do not run cross-model search |
| Search mode | Identifier-style questions suit full-text search; natural-language questions suit semantic search | Run the same test questions through all three modes and compare whether the returned content contains the required fields |
| Reference limit | Measured in tokens; affects completeness when long and short documents are mixed | Start at a token amount that can hold one complete approval record, then compare truncation across long and short documents |
| Minimum relevance score | Takes effect in semantic search or result reranking | Label relevant and irrelevant samples first, then adjust the threshold and compare whether irrelevant samples are filtered out |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Performs a second ordering of candidate text | After enabling per settings, compare ordering before and after reranking and confirm whether candidates are sufficient |
Three common errors
- Treating a retrieval result as an authorization basis. A retrieval hit on the pending-review record for A-1024 does not mean approval is complete. Price tag replacement and discount execution are business actions that require separate system implementation. The check is to compare the status field in the material with the status the request claims, and record the gap when the two do not agree.
- Treating a similarity score as a permission or entity filter. The minimum relevance score is a filtering threshold and does not exclude unrelated business lines or unauthorized records. The check is to confirm whether the retrieval scope contains the target material, record the gap when it is missing, and avoid inferring that a score can replace a permission decision.
- Treating an optional path as executed on every request. Image retrieval requires image input and model capability; reranking requires a model and must be enabled per settings. The check is to compare the actual input with the configuration, and not include an unenabled path in the explanation of results.
How to verify the setup
- Input the hypothetical sample text and the handover request, compare the approval status field in the material with the request's claim of "completed approval," and observe whether a status mismatch appears. If the material shows only pending review, the judgment is unsupported and the gap is preserved.
- Input one identifier-style question and one natural-language question, using full-text search and semantic search respectively, and compare whether the returned content contains the required fields. Observe the failure branch: when a mode returns no hit, record that mode and question type, without guaranteeing a first-position hit or full recall.
- Input an explicit paper-level empty result [] and check whether the downstream design accommodates a reference output whose length can be zero. This check addresses downstream handling logic and does not claim that a fictional rule search necessarily returns [].