What the documents look like
On-demand retail fulfillment handles online orders minute by minute. The documents in this category are generated per order: instant orders, picking lists, out-of-stock reports, and delivery exception records. Under cross-border and multilingual retrieval, the same batch of documents may exist in Chinese, English, and other languages at the same time. A retrieval query and the stored knowledge base content frequently differ in language.
The semantic retrieval mode in FastGPT supports cross-language understanding. A Chinese question can match an English knowledge point. Multilingual material therefore does not need to be translated into a single language before it enters the knowledge base. Documents are organized into collections for classification, while the minimum search unit is the knowledge base itself. A collection is unrelated to search results. If multilingual versions are spread across several knowledge bases, each knowledge base must be associated separately.
Update frequency is minute-level and per order. This page does not depend on that frequency, so no batch synchronization, full synchronization, real-time trigger, or transaction trigger is inferred from it.
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. When candidate knowledge bases are found, check the vector model identifier first, exclude those that do not match, and do not run cross-model retrieval to compare missed recalls. Referenced content is output as an array, and its length can be zero. When no content is found, the output path still completes.
Semantic retrieval, full-text retrieval, and hybrid retrieval have different suitable uses. Queries for numbers, models, and proper nouns suit full-text retrieval or hybrid retrieval. Natural-language questions suit semantic retrieval or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. This is a retrieval process, not a separate toggle that can be switched on. Result reranking requires a ReRank model and must be enabled by configuration. The paths listed in the documented flow diagram do not mean every request executes all of them. Image retrieval requires corresponding image input and model capability.
Minimum relevance is a filter threshold. It takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases is limited to the knowledge base association. It does not extend to external systems.
Hypothetical drill sample (paper material, not a real record): A store fulfillment knowledge base contains an English out-of-stock report template. The original text reads: "Store A reports SKU-4471 out of stock at 14:20 on 2026-03-11; substitute SKU-4472 offered to customer; customer response pending." The sentence to be judged reads: "The store completed the out-of-stock substitution on 2026-03-11 and obtained customer consent." This sample is used to check consistency between the retrieval result and the drafted sentence. The material state "pending" and the drafted sentence "obtained consent" do not agree.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the Chinese and English fulfillment knowledge bases that use the same vector model | Check each knowledge base vector model identifier, exclude those with different identifiers, and do not run cross-model retrieval |
| Search mode | First run semantic retrieval and full-text retrieval separately on the same test questions | Compare recall differences between number-type and natural-language questions under the two modes, then decide whether to use hybrid retrieval |
| Reference upper limit | Hypothetical starting point 3000 tokens | This value is a trial starting point, not an official default; compare with actual materials of different lengths to see whether content is complete, then adjust |
| Minimum relevance | Hypothetical starting point 0.4 | This value is a trial starting point, not an official default; label relevant and irrelevant samples first, and compare filter effects only in semantic retrieval or with reranking enabled |
| Question optimization | Enable for multi-turn follow-up and short-question scenarios | Compare the original question with the optimized question, and check whether the subject was changed incorrectly; a single clear question can be left off to save time |
| Result reranking | Enable when candidates are many and the text question is clear | A ReRank model is required and must be enabled by configuration; compare ordering before and after reranking without assuming improvement |
Three common errors
- Treating a retrieval result as an authorization credential. In the hypothetical sample, the out-of-stock report is at "customer response pending." A retrieval hit on that document only shows the material exists. It does not show the substitution received consent, and it does not show the store obtained an operating permission. A pending review state does not prove the absence of permission, and it does not prove approval was granted.
- Forcing knowledge bases with different vector models into one association. The documents state that associated knowledge bases must use the same vector model. When a knowledge base with a different identifier is found, check the identifier and exclude it first, record the actual observation, and do not assert the runtime result of a joint association.
- Treating minimum relevance as a business filter. This value takes effect only in semantic retrieval or when result reranking is used, and it filters low-relevance results. It does not exclude unrelated businesses or unauthorized records. Retrieval itself is not authorization.
How to verify the setup
- Input the hypothetical sample original text and the sentence to be judged, and compare the drafted sentence "obtained customer consent" with the original "customer response pending." The original state is awaiting response, and the drafted sentence is completed consent. The two values in the same field are mutually exclusive, so the judgment is unsupported and the evidence gap must remain.
- Input the same batch of test questions, run them under semantic retrieval and full-text retrieval separately, and compare the recalled content for number-type questions and natural-language questions. Observe whether number-type questions show insufficient recall under one mode. If so, record it as pending adjustment, and do not assume hybrid retrieval will necessarily resolve it.
- Input a clear paper-based empty-result continuation case: assume the retrieval returns referenced content as
[], and check whether the downstream design continues the path with an empty array. Do not claim a fictional rule makes retrieval return[], and do not guarantee a first hit or full recall.