What the documents look like
Ordering-event materials in this category are refreshed by season and by year. They cover ordering handbooks, ordering contracts, sample lists and credit approval records. In cross-border work the same set of materials commonly exists in several language versions, so a Chinese-language question may need to reach an English-language passage.
FastGPT builds a knowledge base from libraries, collections and data. The smallest search unit is the library. A search runs against the whole library, and collections serve only to classify and manage data; they do not change search results. For multilingual materials, which library a language version belongs to therefore matters more than how its collection is named. Embedding-based semantic retrieval supports cross-language matching, which directly shapes how multilingual material is recalled.
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 built on different vector models cannot be associated together. During a check, confirm the vector model identifier first and exclude those that do not match; do not run a cross-model search to compare missed recalls.
Retrieval is not authorization. Search results supply cited content only. Whether credit approval passes, whether a contract takes effect, and whether a dealer holds ordering permission each depend on a corresponding downstream implementation or a manual check. A pending review status does not demonstrate that an operation is unavailable. Signing, reaching cooperation, and contract effectiveness are three separate facts.
Hypothetical paper exercise sample: a brand channel department prepares an excerpt from an English ordering handbook. The source text reads: "The sample list for this ordering event follows the March 2026 version. The ordering contract may be signed only after regional finance credit approval. The credit limit approval request has been submitted and is in pending review status." A draft sentence reads: "The dealer has obtained a credit limit and can sign the ordering contract directly." The question is whether the draft sentence agrees with the source text. The source text states that the approval request is in pending review status; the draft sentence presents pending review as an obtained limit with direct signing, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that use the same vector model | If multilingual handbooks sit on different vector models, they cannot be associated together |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | Contract-number questions suit full-text retrieval; natural-language questions suit semantic retrieval |
| Citation limit | Unit is tokens; this hypothetical trial starts at 2000, chosen to cover a long handbook passage, compared with a setting of 500 for citation completeness | Tests whether sample-list clauses are truncated |
| Minimum relevance | A 0-1 value, effective only in semantic retrieval or when result reranking is used | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the credit subject is altered |
| Result reranking | Requires a ReRank model, enabled according to settings | Reranking yields a 0-1 score that can be used for filtering |
Three common errors
- Treating pending review as approved. The input is a credit approval request in pending review status; the comparison object is the draft sentence "has obtained a credit limit"; the observable result is that the two status fields disagree, so the draft sentence is unsupported.
- Treating a retrieval result as authorized execution. The input is a retrieved ordering contract clause; the comparison object is whether the dealer already holds signing permission; the observable result is that retrieval returns cited content only, and authorization status requires a separate check.
- Forcing an association across vector models. The input is two knowledge base identifiers on different vector models; the comparison object is the node requirement that the same vector model be used; the observable result is that identifiers which do not match should be excluded, and no cross-model search is run.
How to verify the setup
- Verification one: input the hypothetical sample source text and the draft sentence, compare "pending review status" in the source with "has obtained a credit limit" in the draft sentence. If the two take mutually exclusive values in the same field, the draft sentence is judged unsupported.
- Verification two: input a contract-number question and use full-text retrieval and semantic retrieval separately, then compare whether the number is matched exactly. If full-text retrieval matches and semantic retrieval does not, record that difference; the first result is not guaranteed to match.
- Verification three: input an explicit paper empty result [], and check how the downstream design handles a citation length of zero. Observe whether the output path still completes; do not assert that a fictional rule search returns [].
All of the above are paper checks. They do not constitute an actual test, deployment or approval conclusion.