What the documents look like
Group-buy presale orders collect customer orders first and organize procurement and fulfillment afterward. The daily materials center on presale product tables, order aggregation results, procurement plans, and cut-off notices. In cross-border and multilingual retrieval, one batch of materials may contain Chinese product names, foreign supplier names, and foreign specification descriptions at the same time. Retrieval questions often carry references and cross-language expressions. Materials update daily or per session, so retrieval configuration needs to be checked against the actual language and field structure of each session's materials. Uniform formatting across all sessions cannot be assumed.
A hypothetical drill sample is used below. On 12 March 2025, after the session closed, a platform operator holds three materials. Material one, presale product table, original text: "Product code SG-0312-A, Chinese name 'frozen shrimp', foreign specification 'Frozen Shrimp 500g', supplier name 'Ocean Foods', cut-off time 2025-03-12 18:00." Material two, order aggregation result, original text: "SG-0312-A aggregated quantity 120 units, status: cut-off completed." Material three, procurement plan, original text: "SG-0312-A procurement quantity 120 units, status is pending review." Draft sentence to be judged: "SG-0312-A has completed procurement approval." This draft sentence concerns the procurement plan status. Material three records only "pending review" and does not record approval completion.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. The vector model identifier must be checked first, non-matching bases excluded, and actual observations recorded. Cross-model retrieval is not performed to compare missed recalls.
Reference content is output as an array, and its length can be 0. Even when no content is found, this output path still completes. Retrieval results do not carry permission or entity isolation duties. The minimum relevance score is a filter threshold, not a rule for excluding unrelated businesses or unauthorized records. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled according to settings. Image retrieval requires the corresponding image input and model capability. The overall retrieval flow lists optional paths; not every request executes all of them.
The minimum relevance score takes effect only in semantic retrieval or when result reranking is used. It is a filter threshold, not a final ranking rule. After question optimization, hybrid retrieval, image retrieval, or result reranking is enabled, final results may pass through multi-path recall fusion and do not necessarily follow a single vector similarity order. The reference limit is measured in tokens per search. This unit is documented and is not a guessed item count. Question optimization performs reference resolution and question expansion before formal retrieval, adding one model call; it can improve multi-turn retrieval and also increases overall latency.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Associate one presale material base using the same vector model | Check the vector model identifier first, exclude bases with different models, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Materials contain codes and multilingual specifications; compare semantic and full-text recall results |
| Reference limit | Hypothetical starting point 2000 tokens | Compare content completeness using actual materials of different lengths, then adjust the limit |
| Minimum relevance score | Hypothetical starting point 0.3 | Label relevant and irrelevant samples first, compare filtered results, then adjust the threshold |
| Question optimization | Enabled | Multilingual reference questions are common; compare the original question and the optimized question to check whether the subject was changed incorrectly |
| Result reranking | Enabled according to settings | Reranking requires a model; compare rerank scores and vector scores before deciding on filtering |
These values are hypothetical trial starting points for this page. They are not official defaults. Each starting point needs a stated reason and a comparison method, and each must be adjusted against observed results.
Three common errors
- Treating the minimum relevance score as a permission or entity filter. Input a cross-language question, compare whether returned references contain unauthorized records, and observe whether the reference content changes before and after the minimum relevance score setting. This threshold filters only low-relevance results and does not exclude unauthorized records.
- Treating a pending-review status as completed approval. Input the original text of material three, "pending review," and the draft sentence "has completed procurement approval." Compare the status values in the same field and observe whether the two are mutually exclusive. Pending review does not demonstrate completed approval, and it does not demonstrate a lack of operating permission.
- Treating RRF as a switch that can be turned on. Input the hybrid retrieval configuration, compare whether an RRF switch item exists among node parameters, and observe whether the documentation describes RRF as a fusion process. RRF is a hybrid result fusion process, not an independent switch.
How to verify the setup
- Input the original text of material one and the question "What is the foreign specification of SG-0312-A?" Compare whether the returned reference contains "Frozen Shrimp 500g," and observe whether the reference content matches the original text. If an empty array is returned, record it as not recalled. Do not claim that a rule-based search must return empty.
- Input the original text of material three and the draft sentence "SG-0312-A has completed procurement approval." Compare the values of the same field "status," and observe whether "pending review" and "approval completed" are mutually exclusive. This draft sentence is not supported, and the evidence gap remains.
- Input the original text of material two and the question "What is the aggregated quantity of SG-0312-A?" Compare whether the returned reference contains "120 units," and observe whether the reference content matches the original text. If the reference contains data from another session, record it as a scope mismatch and do not infer a permission problem.
For the empty-reference path, a paper input of an empty array can be used to check the downstream design. The claim that a fictional rule-based search must return an empty array is not made. Retrieval is not authorization. Forms collect values, while business verification, approval, database writes, and external synchronization require their stated downstream implementation. Material three records only "pending review"; approval completion, signing, cooperation reached, and contract effectiveness are separate facts. Where a document field is not provided, it remains unknown.