What the documents look like
Producer export tax processes combine domestic-sale tax offsets with export refunds in a single calculation, so the inventory procurement and supplier materials must line up with the export tax declaration detail sheet, export contracts, the sales sub-ledger and customs export declarations. Purchase receipt records, supplier invoices and payment vouchers commonly diverge from the goods name, quantity and free-on-board value shown on the customs declaration. Materials update monthly, and collecting self-produced export items and handling free-on-board differences consume the most working time. This page addresses one narrow step: after these materials are built into a searchable knowledge base, whether the associated knowledge bases can be linked under the same vector model, and whether each index can be checked against its source content separately.
A knowledge base search operates at the knowledge base level as the smallest search unit; a collection only classifies data and does not affect search results. Knowledge bases selected for a single search must use the same vector model, so the actual model and its input types are checked first. One source data block may be associated with multiple indexes, and each index is verified against the original text separately. Index content should be complete and focused, and retrieval coverage is weighed against index length using actual samples. Image retrieval also depends on whether an image description or image vector index was generated at ingestion time. Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search merges results through RRF, and reranking runs when configured. The minimum relevance value takes effect only under semantic search or when result reranking is used, and it is a filter threshold rather than a permission or entity-isolation mechanism. Retrieval does not constitute authorization, and a form collecting a value does not equal business verification, approval or a database write.
Constraints at the vector indexes step
The inventory procurement and supplier knowledge base and the export tax declaration knowledge base must be linked only when their vector model identifiers match. If the identifiers differ, the knowledge bases are excluded from joint association and the mismatch is recorded; no cross-model search is executed to compare recall. The index content for purchase receipts, supplier invoices and customs declaration blocks is checked against the original text field by field, because one source block can carry several indexes and a single index does not represent the whole block. Image retrieval requires both a corresponding image input and model capability; without either, the image path does not execute and only the text path returns references. The reference output is an array whose length may be zero, so downstream design must accept an empty result. The minimum relevance value is a 0-1 number that filters low-relevance results only under semantic search or result reranking, and it never isolates unauthorized records or unrelated businesses.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to the inventory procurement and supplier materials task. The two numeric starting values are hypothetical trial values for this page only, chosen so that filtered and unfiltered result counts can be compared against the documented example values; they are not official defaults.
| Configuration or check item | Value or object for this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Inventory procurement and supplier materials knowledge base | Must use the same vector model as the export tax declaration knowledge base; verify the model identifier first |
| Search mode | Semantic search, full-text search or hybrid search | Exact queries such as goods codes and customs declaration numbers suit full-text or hybrid search |
| Index content | Purchase receipt, supplier invoice and customs declaration blocks | Index and source text are checked separately; one source block may carry multiple indexes |
| Reference limit | Counted in tokens; hypothetical trial starting value 3000 | Trial value chosen for this page to compare the number of cited items against the documented example value |
| Minimum relevance | 0-1 number; hypothetical trial starting value 0.4 | Takes effect only under semantic search or result reranking; compared against the documented example value for filtered counts |
| Result reranking | ReRank enabled when configured | Rerank score is 0-1 and can be used for filtering; a rerank model must be available |
Three common errors
- Linking knowledge bases that use different vector models. The input is the model identifiers of two knowledge bases; the comparison is whether the two identifiers match; the observable result is that a mismatch excludes the pair from joint association and no cross-model search is executed.
- Treating a semantic recall score as a permission or entity filter. The input is a query that reaches beyond the authorized scope; the comparison is between the returned references and the authorized scope; the observable result is that the score level carries no authorization meaning and an out-of-scope record may still be recalled.
- Treating supplier materials pending review as already approved for ingestion. The input is a purchase record marked pending review; the comparison is between its status field and an approved record; the observable result is that pending review does not show approval and does not show that an operation permission is absent.
How to verify the setup
Hypothetical paper exercise sample. The source text reads: "Supplier A supplied batch A in March 2026, purchase receipt number RK-2026-03-018, corresponding customs export declaration number BG-2026-03-066, amounts and quantities in the attachment, material status pending review." The draft sentence to be judged reads: "Supplier A's batch has completed ingestion verification and can be used for the export tax declaration." The comparison is between the draft sentence and the status field in the source text. The source text states only pending review and does not state that verification is complete, so the draft sentence is unsupported and the evidence gap is recorded.
- Verification one: input the source text above and the draft sentence; compare the two values "pending review" and "verification complete"; the condition is met when the status fields differ, which marks the draft sentence as unsupported.
- Verification two: input the model identifiers of two knowledge bases; compare whether they are identical; the condition is met when the identifiers match, which permits association, while a mismatch excludes the pair and records the observation.
- Verification three: input a query with no image input; compare the retrieval path against the image retrieval conditions; the condition is met when no image input and no model capability mean the image path does not execute and only the text path returns references; the reference output may be an array of length zero, and the downstream design accepts the empty result.
All of the above are paper checks. No test, deployment, approval, refund or other business action is claimed to have occurred.