What the documents look like
FBA preparation moves inventory into a platform warehouse ahead of demand, so the working file is built from an inbound plan, carton labels, inventory reports and compensation claims. The operational risk sits with inbound discrepancies and inventory age. Contract and transaction material comparison asks for the same field of the same entity to be checked inside one period, which means retrieval has to return the specific clause or record rather than a broad topical passage.
In FastGPT v4.17.2, the knowledge base search node can be linked to one or more knowledge bases that share the same vector model, and search operates at the level of the knowledge base as the smallest search unit. A collection is a classification layer for data and does not change search behaviour. The node emits cited content as an array whose length can be 0, so the downstream path still executes when nothing is retrieved. Semantic search works from vector distance, full-text search works from keyword hits, hybrid search merges both paths through RRF, and ReRank performs a second ordering of candidate text with a 0-1 score. The minimum relevance value is a filter threshold that takes effect only in semantic search or when result reordering is used, and it does not act as a permission or entity isolation control. The citation ceiling is measured in tokens, so materials of very different length need their completeness checked rather than assumed.
Constraints at the retrieval step
Retrieval precision depends on both the quality of the search terms and the quality of the stored data. Question optimization performs coreference resolution and question expansion before retrieval, which adds a model call and therefore adds latency; it assists multi-turn and short questions but can be skipped when the question is already explicit. Hybrid search produces a wide candidate set that cannot be similarity filtered directly, which is why a reordering model is normally used and its score is used for filtering. Reordering depends on a complete question with subject and predicate present, so question optimization typically runs before search and reordering. Image retrieval is an additional path, not a separate capability: it requires a suitable image input and the corresponding model capability, and whether an image can be retrieved depends on the index created at ingestion time as well as on the query.
Hypothetical paper exercise sample. The source text to be checked reads: "Entity: Company A; Object: SKU-001; Stage: inbound plan submitted, awaiting platform confirmation; Date range: 1 March 2025 to 31 March 2025; Authorization boundary: this replenishment batch only." A draft sentence awaiting judgment reads: "Company A's inbound plan for SKU-001 has received platform confirmation." The source records only an awaiting-confirmation state, so the draft sentence converts a pending state into a confirmed state and is unsupported by the source.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only material libraries that share the same vector model | Compare the vector model identifier of each library; exclude those whose identifier differs and do not run cross-model retrieval |
| Search mode | Start with hybrid search | Run the same question under semantic, full-text and hybrid search, then compare whether candidates cover both the SKU number and the natural-language phrasing |
| Citation ceiling | Set to a token amount that covers one inbound plan | Compare citation content against materials of different length to see whether content is truncated, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set the value | Takes effect only in semantic search or result reordering; compare filtering behaviour against the labelled samples |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the entity has been altered |
| Result reordering | Enable when candidates are numerous and a reordering model is available | Enable according to the setting and compare the ordering produced by the reordering score against the vector score |
Three common errors
- Treating a pending state as granted authorization. The draft sentence "has received platform confirmation" and the source value "awaiting platform confirmation" are mutually exclusive values on the same field of the same entity in the same period. The draft sentence is unsupported, and the evidence gap is recorded rather than filled by inference.
- Treating the minimum relevance value as a permission filter. That value is a filter threshold and does not exclude unrelated business lines or unauthorized records. Retrieval does not constitute authorization, and a relevance score is not a category, entity or permission filter.
- Linking knowledge bases that use different vector models. The correct action is to check the vector model identifier and exclude libraries that do not match, then record what is observed. No claim is made about retrieval behaviour after an unsupported cross-model link.
How to verify the setup
- Input the inbound plan source text above together with the draft sentence. Compare the two values "awaiting platform confirmation" and "has received platform confirmation" on the same field. The condition is satisfied when the draft sentence is judged unsupported and the evidence gap is recorded.
- Input the same question under semantic search and under full-text search. Compare whether the candidates cover the SKU number. The condition is satisfied when the number-style query is hit under full-text or hybrid search, and any failure branch is recorded as an observation rather than a guarantee.
- Input an explicit empty citation array and compare whether the downstream design accepts a length of 0. The condition is satisfied when the path executes and no retrieval hit is fabricated. This checks the documented output behaviour and does not assert that any particular query returns an empty array.
A request that is pending is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the source does not state a field, that field remains unknown.