What the documents look like
Short-video affiliate distribution relies on long-term content exposure and product-link conversion rather than live explanation. The working materials for this category are short-video scripts, linked-product lists, commission plans and violation feedback. Inventory procurement and supplier records form one further document family inside that same knowledge base: supplier names, purchase batches, arrival quantities, settlement terms and authorization scope, usually arriving as documents or tables. These records sit in separate collections from scripts and commission plans, but FastGPT searches the whole library. A collection only classifies data and does not affect search results. Material updates arrive daily or weekly, so the indexing question is whether each batch of procurement and supplier records produced retrievable indexes at ingestion time, and whether those indexes stay focused on procurement and supplier subject matter.
Constraints at the vector indexes step
Knowledge bases joined in one search must share the same vector model, so the actual model identifier and supported input types are checked first. One source data block can map to several vectors through its indexes field, and the index content and the original content are reviewed separately. Index content should be complete and focused; retrieval coverage and index length are traded off against sample evidence. Image retrieval depends on whether an image description or image vector index was generated at ingestion, not only on what the query submits. After any model or index change, the same question set is compared against the cited source text, and vector scores alone do not support an effectiveness claim.
Semantic recall and relevance do not perform permission or entity isolation. The minimum relevance value is a filter threshold and applies only under semantic retrieval or result re-ranking. With question optimization, hybrid retrieval, image retrieval or re-ranking enabled, final results can come from several recall paths merged together. RRF is the merge step for multiple result paths in hybrid retrieval and is not a separate switch. The citation limit is measured in tokens, not top k. Retrieval is not authorization: forms collect values, and business verification, approval, database writes and external synchronization each need their own downstream implementation.
Configuration and implementation checks
The table below covers only documented settings and implementation checks relevant to this page's feature and primary task.
| Configuration item | Setting or check | Relation to this page's samples |
|---|---|---|
| Linked knowledge bases | Select one or more knowledge bases that use the same vector model | The supplier records library and the procurement list library are checked for matching model identifiers before joint search |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Supplier names and batch numbers suit full-text or hybrid retrieval |
| Minimum relevance | A 0-1 value, effective only under semantic retrieval or result re-ranking | This hypothetical trial starting value is 0.3; the reason is that it is above an unfiltered state, and the comparison object is the number of recalled items when the value is left unadjusted |
| Citation limit | Maximum tokens cited per search | This hypothetical trial starting value is 2000 tokens; the reason is that individual supplier records are short, and the comparison object is citation completeness at 1000 tokens |
| Question optimization | Coreference removal and question expansion before retrieval | Can be enabled for multi-turn follow-up on supplier arrival status, and disabled for a single clear question |
| Index check | Verify the correspondence between a source data block and its indexes field | Confirm whether purchase batch and supplier name each have their own index |
Three common errors
- Linking knowledge bases that use different vector models. The input is the model identifiers of two knowledge bases; the comparison object is the documented requirement that linked knowledge bases use the same vector model; the observable result is that identifiers which do not match are recorded and excluded rather than searched jointly.
- Treating minimum relevance as a permission or entity filter. The input is a semantic search containing a supplier name; the comparison object is the documented statement that this parameter filters low-relevance results only; the observable result is that the relevance score carries no function of excluding unauthorized supplier records.
- Describing RRF as a separate switch that can be turned on. The input is the configuration surface for hybrid retrieval; the comparison object is the documented description of RRF as the merge process for multiple result paths; the observable result is that no corresponding switch item exists at that step.
How to verify the setup
The following uses an explicitly labelled hypothetical paper exercise. It involves no real customer or real company.
Hypothetical exercise sample: the source material reads "Supplier: Xinghai Daily Goods Trading Firm; purchase batch: XH-2024-031; arrival quantity: 1200 units; settlement method: monthly settlement, 30 days; authorization scope: limited to short-video product links on this firm's own accounts." The draft sentence to be assessed reads "Xinghai Daily Goods Trading Firm has approved authorizing this batch of goods to any influencer account for product links." The draft sentence conflicts with the value recorded in the authorization scope field of the source material and is unsupported.
- Input the source material above together with the draft sentence; compare "any influencer account" in the draft against "limited to this firm's own accounts" in the source; the observable result is that the two take mutually exclusive values in the same authorization scope field, so the draft is judged unsupported.
- Input a semantic search for "Xinghai Daily Goods Trading Firm authorization scope"; compare the returned citations against the source material; the observable result is that the citation contains the authorization scope sentence from the source and contains no authorization extension absent from the source.
- Input an empty-result downstream check; compare the downstream design against the documented behaviour that citation output is an array whose length can be 0; the observable result is that the chain completes and produces no fabricated supplier authorization conclusion.
All of the above is paper review only and does not claim that any test was run, deployed or approved. After a model or index change, the same question set is compared against the cited source text, and vector scores alone do not support an effectiveness claim.