What the documents look like
Cold-chain receiving at a store produces records at batch granularity. A single delivery can generate a cold-chain vehicle temperature log, an arrival acceptance sheet, a rejection note and a supplier delivery note. These records enter the knowledge base for inventory procurement and supplier materials, where store receiving staff and quality inspectors search them. Every arrival adds new temperature records and acceptance results, so the index keeps growing at a daily and per-batch rhythm. The people involved are fresh warehousing and distribution staff, store receiving staff and quality inspectors. Their retrieval needs cluster around whether a temperature went out of range, whether a batch passed, and how a rejection reason was registered. Because mistakes here can lead to spoiled food, supplier disputes and regulatory penalties, the cited content returned by retrieval must be traceable to a specific batch and document. Retrieval itself does not decide authorization.
Constraints at the retrieval step
The materials contain identifiers, batch numbers, temperature values and rejection reason text. Identifier queries depend on keyword hits from full-text retrieval, while natural-language questions depend on vector distance from semantic retrieval. Hybrid retrieval fuses multiple result paths through RRF, and a rerank model reorders candidate text and produces a score between 0 and 1. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. The citation limit controls the total cited content per search in tokens. Question optimization performs coreference resolution and question expansion before retrieval and makes an extra model call. A knowledge base search node can associate one or more knowledge bases that use the same vector model; knowledge bases with different vector models cannot be associated together. Cited content is output as an array whose length can be 0, and the downstream path still runs when no content is found.
Hypothetical drill sample: a store receives a cold-chain delivery. The supplier delivery note states "delivery date 3 June 2024, item chilled pork, batch number P-20240603-07". The cold-chain temperature log states "transport temperature for this batch 2°C to 4°C throughout, recording window 3 June 05:00 to 08:00". The arrival acceptance sheet states "acceptance status is pending quality inspection confirmation, temporarily held in the receiving area". The draft handover request to be assessed is: "Please search the knowledge base to confirm whether this batch has completed acceptance and can be put into storage directly." In this request, "has completed acceptance" and the source text "pending quality inspection confirmation" point to different values of the same field, so this is a status pending verification.
Configuration and implementation checks
The table below lists, for this page's sample, the officially documented settings or implementation check items relevant to this feature. The values are hypothetical starting points for this drill only, used for comparison with another value, and are not official defaults.
| Setting or check item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the inventory procurement and supplier materials knowledge base that uses the same vector model | The node requires the same vector model; first check each knowledge base's vector model identifier and exclude inconsistent ones |
| Search mode | Run the sample question with hybrid retrieval first | The sample contains both a batch number and natural-language description; compare what semantic retrieval and full-text retrieval each recall |
| Citation limit | Hypothetical starting point 1500 tokens | Compare whether citation content stays complete using acceptance sheets and temperature logs of different lengths, then adjust |
| Minimum relevance | Hypothetical starting point 0.3 | Label relevant and irrelevant samples first, then observe whether filtered citations still contain the required batch information |
| Question optimization | Enabled for the sample request | Compare the original question with the optimized question and check whether the batch number and status word were altered |
| Result reranking | Enabled per actual configuration, observe the score | The rerank score of 0 to 1 is used for filtering; compare candidate order and citation content before and after reranking |
Three common errors
- Treating retrieval results as an authorization basis. In the sample the acceptance sheet status is "pending quality inspection confirmation", and retrieving that sheet does not mean acceptance is complete or storage is permitted. The input is the cited content returned by retrieval; the comparison is between the document's status field and "has completed acceptance" in the request; the observable result is that the two values differ and the case must be recorded as pending verification.
- Associating knowledge bases across vector models. If two knowledge bases using different vector models are associated at the same time, the node's same-vector-model requirement is violated. The input is each knowledge base's vector model identifier; the comparison is whether the identifiers match; the observable result is that inconsistent ones are excluded, and no cross-model retrieval is run to compare missed recalls.
- Treating minimum relevance as an entity or permission filter. Minimum relevance is a filter threshold between 0 and 1 and takes effect only under semantic retrieval or result reranking. The input is the sample question and the threshold setting; the comparison is whether citations before and after filtering contain the batch; the observable result is that changing the threshold affects relevance ordering only and does not exclude unauthorized records.
How to verify the setup
- Using the sample source text and draft request above, the input is "Please search the knowledge base to confirm whether this batch has completed acceptance and can be put into storage directly". Compare whether the acceptance sheet status field in the returned citations is "pending quality inspection confirmation". The condition is met when the citations contain that status and do not present a conclusion of "acceptance completed", showing that retrieval did not overstep into an authorization judgment.
- Using batch number P-20240603-07 from the sample as the query, the input is two modes, full-text retrieval and hybrid retrieval. Compare whether the citations returned by each mode contain the supplier delivery note and temperature log for that batch number. The condition is met when the identifier query hits under full-text retrieval, showing the identifier query path is usable.
- Using the explicit paper input
[]to check empty-citation handling, the input is a retrieval request that matches no content. Compare whether the output citation array length is 0 and whether the downstream path still runs. The condition is met when the path continues to execute, showing the empty-result branch is handled as documented.