What the documents look like
In retail checkout and loyalty promotion work, the material centers on price and member entitlement rather than short-shelf-life stock. The main sources are the POS system, the member CRM, payment platforms and the price-tag system. Daily documents include promotional price tags, member points rules, checkout daily reports and exception refund slips. Updates arrive daily or once per shift, so the same product may carry different price-tag versions across shifts. Any retrieved reference therefore needs to keep the subject, the object and the effective period traceable to a specific shift and scope. Errors in this area lead to price complaints, cash shortfalls or overages, and member entitlement disputes.
A knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The retrieval chain includes semantic search, full-text search, hybrid search, question optimization and result re-ranking; these are optional paths, not steps that every request executes. Image retrieval requires image input and matching model capability, and re-ranking requires a re-rank model enabled by configuration. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result re-ranking is used. RRF is a fusion process for combining multiple result channels, not a switch. Retrieval and similarity scores do not perform permission or entity isolation, so relevance cannot be treated as an authorization filter. An empty result is output as an array whose length may be 0.
Hypothetical paper exercise sample. A store price-tag system holds the archived text: "From 1 May 2024 to 7 May 2024, Brand A 500 ml drinking water has a promotional price of 2.8 yuan; member points accrue at 1 point per yuan of actual payment." A draft sentence to be judged reads: "Brand A 500 ml drinking water is on promotion at 2.8 yuan for the whole of May, with member points at 2 points per yuan." Against the source, the promotion period is 1 to 7 May and the points basis is 1 point per yuan. The draft differs from the source on both the period and the points fields, giving mutually exclusive values in the same fields, which constitutes a contradiction. This sample only checks whether a retrieved reference preserves the effective scope; it does not indicate that any approval or execution has occurred.
Constraints at the retrieval step
The search node accepts one or more knowledge bases sharing the same vector model. Checking the model identifier of each base comes first; bases whose identifiers differ are excluded and no cross-model retrieval is performed. The retrieval chain is assembled from the input content and the search parameters, combining text, images, semantic recall, full-text recall, question optimization and re-ranking, then fusing the multiple channels into reference content. Question optimization performs coreference resolution and question expansion before retrieval and adds one model call, which usually improves multi-turn retrieval but also adds latency. Semantic search relies on vector distance and suits natural-language questions and semantically close content. Full-text search relies on keyword hits and suits numbers, models, proper nouns and error codes. Hybrid search uses both and merges them through RRF. Re-ranking produces a 0-1 score representing relevance between the searched content and the question, and that score can be used for filtering.
The reference limit caps how many tokens of content each search may cite, chosen instead of top k because chunk lengths vary widely in mixed knowledge bases. The minimum relevance value filters out low-relevance results and applies only under semantic search or when re-ranking is used; it is a filtering threshold, not a final ordering rule. When question optimization, hybrid search, image retrieval or re-ranking is enabled, the final result may come from multi-channel fusion and does not necessarily follow a single vector similarity order. Image retrieval depends on image clarity, whether the content is easy for the model to interpret, whether a vision model is configured, and whether the vector model supports image vectors. Whether an image can be retrieved also depends on whether a corresponding index was built at ingestion time. Retrieval and similarity scores do not isolate permissions or entities.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the promotion and member bases that share the same vector model | Check each base's vector model identifier first, exclude differing ones, and do not run cross-model retrieval |
| Search mode | Run semantic search and full-text search separately on the same question | Compare how the two modes recall "Brand A 500 ml drinking water" |
| Reference limit | Start at a token amount that covers one complete price-tag clause | Test with actual material of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Start at a middle value, effective only under semantic search or re-ranking | Label relevant and irrelevant samples first, then compare threshold changes |
| Question optimization | Enable for follow-up questions containing coreference, disable for explicit questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result re-ranking | Enable by configuration when a re-rank model is available | Compare candidate ordering before and after re-ranking and confirm the question is complete |
Three common errors
- Forcing knowledge bases with different vector models into one association. Input: two bases whose vector model identifiers differ. Compare their model identifiers. Observable result: when identifiers do not match, exclude them, do not run cross-model retrieval, and do not infer recall gaps or redundancy as runtime outcomes.
- Treating minimum relevance as a permission or entity-isolation mechanism. Input: a low-relevance refund slip belonging to another store. Compare the retrieved return with the authorized scope. Observable result: a relevance score does not perform authorization filtering; cross-scope records still require a downstream permission implementation to block them.
- Treating a promotional price tag awaiting review as an effective price. Input: a price-tag draft marked as pending review. Compare its status field with its effective field. Observable result: pending review does not demonstrate that approval was granted, and the evidence gap must remain when citing it.
How to verify the setup
- Using the hypothetical sample source text and the draft sentence, enter the retrieval question "Brand A 500 ml drinking water promotion period and points basis" and compare whether the period and points fields in the returned reference match the source. Condition met: the reference keeps 1 to 7 May and 1 point per yuan, and the draft sentence is marked as contradicting the source.
- Enter a follow-up question containing only a coreference and compare the search terms before and after question optimization. Condition met: the subject is correctly completed and no other product or store is mixed in.
- Enter an explicit paper empty result [] and compare how the downstream node is designed to handle a zero-length reference. Condition met: the downstream follows the empty-reference branch and does not claim that a fabricated rule search necessarily returns []. All three are paper checks and do not indicate that testing or deployment has occurred.