What the documents look like
Regional dealer stock replenishment runs as a daily rolling order flow rather than a seasonal centralised ordering event. The four document types that matter here are replenishment requests, inventory turnover tables, outbound delivery notes and payment collection plans. These materials update on a weekly and monthly cadence, so retrieval faces several near-identical records for the same store across different weeks. In the store service and internal training setting, questions typically come from sales back-office staff or channel managers. The questions are short, they carry store abbreviations and document numbers, and they often omit the requesting subject and the period.
The knowledge base search node can link one or more knowledge bases that use the same vector model. Reference content is returned as an array and its length can be zero, which means the downstream output path still runs even when nothing is retrieved. The retrieval step returns references only. It does not perform authorisation checks.
Constraints at the retrieval step
Knowledge bases built on different vector models cannot be linked together. Before configuration, check the model identifier and exclude any knowledge base that does not match, then record the actual observation. Do not run cross-model retrieval to compare missed recalls.
Semantic search, full-text search and hybrid search are compared separately for each test question. No single mode is assumed to suit every document type. Queries for numbers, model codes and proper nouns are better served by full-text or hybrid search. The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filter threshold, not a permission or entity isolation mechanism. The reference upper limit is measured in tokens. Question optimisation handles coreference resolution and question expansion, and it adds one model call. Image retrieval requires the corresponding image input and model capability. Reranking requires a model and must be enabled according to the setting. The optional paths listed in the documentation do not mean every request executes all of them.
Hypothetical paper exercise sample. The source material reads: "Jia County Yi Trading Firm, August 2025 replenishment request, requesting SKU-3302 for a total of 120 units, status is pending review." "Jia County Yi Trading Firm, August 2025 outbound delivery note, SKU-3302 actually shipped 100 units, status: shipped." The draft sentence to be judged reads: "The replenishment request of Jia County Yi Trading Firm for SKU-3302 has been approved." This sentence conflicts with the source. The request status is pending review, and pending review does not constitute approval. The delivery note states only the shipped quantity and records no approval conclusion.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the base containing the replenishment request and the delivery note | Check the vector model identifier first and exclude any base using a different model |
| Search mode | Hybrid search | Run the same test question against semantic, full-text and hybrid modes and compare the returned references |
| Reference upper limit | Starting point of 2000 tokens | Hypothetical starting point; compare content completeness using actual materials of different lengths |
| Minimum relevance | Starting point of 0.4 | Hypothetical starting point; label relevant and irrelevant samples first, then compare |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled | Where a reranking model is available, compare candidate ordering before and after reranking |
The values in this table are hypothetical starting points for this exercise, not official defaults. RRF is the fusion process for hybrid search results and is not a toggle configuration.
Three common errors
- Treating a retrieval hit as authorisation. Input "Jia County Yi Trading Firm, August 2025 replenishment request" and retrieve a reference, then conclude the request has been approved. Check: input the source text above, compare the request status field, and observe that when the status is pending review no approval conclusion may be output.
- Linking knowledge bases built on different vector models. Input two bases with different model identifiers and expect merged retrieval. Check: compare the vector model identifiers of the bases, observe that a mismatched base is excluded and the observation is recorded, and do not infer missed recalls or redundant results.
- Using minimum relevance to exclude unrelated business. Input a question containing another store name and expect the threshold to filter out that store's records. Check: compare the returned references against the question subject, and observe that the relevance score does not constitute entity or permission filtering. Records outside the authorised scope must be handled by a downstream implementation.
How to verify the setup
- Input the two source sentences and the draft sentence from the hypothetical sample, compare the draft sentence against the request status field, and observe that the system output is consistent with "pending review does not constitute approval" and contains no approval conclusion. This counts as satisfied.
- Input an explicit paper empty result
[], compare how the downstream node handles empty references, and observe that the output path runs through and the downstream design has corresponding handling. This counts as satisfied. Do not claim that retrieval under a fictional rule necessarily returns[]. - Input a short question containing a store abbreviation and a document number, compare the question text before and after optimisation, and observe that the subject and period were not rewritten and no authorisation information was added. This counts as satisfied. If the subject is still missing after optimisation, record it as pending verification rather than as a negative finding.
All of the above are paper checks. No retrieval, approval or synchronisation action was actually executed.