What the documents look like
Warehouse picking, checking and packing handles outbound orders. It does not handle supplier arrivals. The main data sources are the WMS, the order system, the electronic waybill platform and weighing equipment. Daily documents include wave picking lists, checking records, waybills and packaging consumable sheets. Data updates per order or per hour.
In cross-border work, these documents may exist in Chinese, English and other languages at the same time. A retrieval step has to match across those languages. The role of knowledge base search and recall here is to find content in already ingested documents that relates to a current picking, checking or packing question, for later human review or process use.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. Check the vector model identifier first, exclude non-matching bases, and record what was observed. Do not run a cross-model retrieval to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Semantic retrieval relies on vector distance and supports cross-language understanding. Full-text retrieval relies on keyword hits and suits numbers, models and proper nouns. Hybrid retrieval uses both and fuses the results through RRF. RRF is a fusion process for mixed results, not an on/off switch.
The citation limit is measured in tokens and controls how much content each search may cite. The minimum relevance score is a 0-1 filter threshold, effective only in semantic retrieval or when result re-ranking is used. That threshold is a filter condition and does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external systems needing the same model.
Question optimization handles coreference resolution and question expansion, adding one model call before retrieval. Image retrieval needs image input and the corresponding model capability. Re-ranking needs a model and is enabled according to settings. The optional paths shown in the overall document flow diagram do not mean every path runs on every request.
On an empty result, the cited content is output as an array whose length can be 0, and the output link still completes.
Configuration and implementation checks
The configuration below is built around a hypothetical drill sample. Sample content: a wave picking list states "Wave number W-2024-0512, picker Zhang San, checker Li Si, order number SO-8841, item SKU-3302 quantity 2, packaging consumable carton C-40". One draft sentence to be judged is "This wave has completed checking, and checker Li Si has signed". The source text does not state a checking-completion status or a signature field, so the draft sentence is not supported.
| Configuration item | Hypothetical trial starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one base using the same vector model | Check the vector model identifier first, exclude bases with different models, record what was observed |
| Search mode | Hybrid retrieval | The sample contains number W-2024-0512 and a natural-language question; compare the semantic and full-text recall paths |
| Citation limit | Set to 500 tokens | Hypothetical starting point for this exercise, since the sample source text is about 60 characters; compare content completeness against 200 and 800 |
| Minimum relevance score | Set to 0.3 | Hypothetical starting point for this exercise, since low-relevance fragments need filtering; compare recall counts against 0.1 and 0.5 |
| Question optimization | Off | The sample question has a clear subject; compare whether the subject is altered when on versus off |
| Result re-ranking | Enabled according to settings | A re-ranking model is required; compare candidate ordering before and after enabling |
Three common errors
- Treating retrieval or form submission as authorized execution. Input: a checking record form shows "pending review". Compare: the form status against the authorization status. Observable result: pending review does not show that operating permission is missing, and it does not equal approval; record it as pending verification.
- Treating the minimum relevance score as a permission or entity filter. Input: minimum relevance score set to 0.3. Compare: filtering low-relevance fragments against excluding unrelated business records. Observable result: the threshold affects similarity filtering only and does not perform permission isolation; unrelated records may still be recalled.
- Writing RRF as an on/off switch. Input: a hybrid retrieval configuration. Compare: the role of RRF in the documentation against configurable fields. Observable result: RRF is a fusion process for mixed results; the documentation does not describe it as a standalone switch, so that switch should not appear in configuration.
How to verify the setup
- Input the hypothetical sample source text "Wave number W-2024-0512, picker Zhang San, checker Li Si, order number SO-8841, item SKU-3302 quantity 2, packaging consumable carton C-40" and the draft sentence "This wave has completed checking, and checker Li Si has signed". Compare the draft sentence against the source text: the source text does not state a checking-completion status or a signature field. Observable result: the draft sentence is not supported; record the evidence gap, and do not judge it as a contradiction or a denial.
- Input a query that clearly contains "SKU-3302". Compare the recalled content from full-text retrieval against semantic retrieval. Observable result: full-text retrieval should hit the fragment containing that number; semantic retrieval may hit semantically similar content. If full-text retrieval does not hit, record what was observed; do not guarantee a first-position hit or a complete recall.
- Input an empty query and check the cited content output. Compare the output array length against the downstream design. Observable result: the cited content length can be 0, and the output link still completes; do not claim that a fictional rule retrieval necessarily returns an empty array.