What the documents look like
Central chain procurement and distribution handles product master data that arrives from several operational systems at once. ERP procurement records carry supplier and purchasing terms, WMS holds warehouse movement, store POS shows what stores actually sold, and supplier delivery data confirms what was shipped. The documents that matter for this page are central procurement contracts, replenishment plans, distribution orders, and store requisition forms. Product code, specification, supplying entity, and distribution scope frequently sit in different documents rather than one consolidated record. A selection question has to align the same product across a contract, a requisition, and a distribution order before any answer is useful. The vector model and index step maps those varied expressions into comparable vectors so retrieval can return candidate source text for downstream business review. Update frequency affects when an index is rebuilt; it does not change the index structure itself.
Two features of this category shape the work. First, the searchable unit is the whole knowledge base, and collections inside it are only for classification, so they do not influence search results. Second, product identity is split across documents, which means a retrieval result is a candidate for human comparison, not a confirmed match.
Constraints at the vector indexes step
Knowledge base search uses the knowledge base as its smallest search unit. A collection groups data for management and does not change search behavior. Vector retrieval runs on the PostgresSQL PG Vector extension with an HNSW index. MongoDB stores the original data in the dataset.datas table, and its indexes field records the corresponding vector IDs as an array, so one data record can map to multiple vectors. When several knowledge bases are linked for retrieval, they must use the same vector model. The correct action is to check the model identifier and input type first and exclude any knowledge base that does not match, rather than running a cross-model retrieval to compare recall.
One original data block can carry several indexes, and the index content is checked separately from the original text content. Image retrieval depends on whether an image description or image vector index was generated at ingestion time; the query side alone does not determine it. Hybrid search merges multiple retrieval paths through RRF, which the documentation describes as a fusion process rather than a configurable switch. ReRank requires a model and must be enabled in settings. The minimum relevance value applies only during semantic retrieval or when result reranking is used, and it acts as a filter threshold. It does not perform permission or entity isolation.
Configuration and implementation checks
The table below covers settings and implementation checks relevant to product data and selection questions. The starting values are hypothetical test starting points for this page only, chosen so they can be compared against the actual cited source text of the same question group. They are not official defaults.
| Configuration item | Hypothetical starting point | Check method and basis |
|---|---|---|
| Vector model | Same model identifier | Linked contract and requisition knowledge bases must share one vector model; verify the identifier before linking |
| Index content | Product code plus specification phrase | Build indexes from complete, focused information; weigh retrieval coverage against index length using samples |
| Citation limit | Hypothetical value counted in tokens | The documentation controls citation limit by tokens; compare the starting value against cited source length for the same question group |
| Minimum relevance | Hypothetical value between 0 and 1 | Applies only during semantic retrieval or result reranking; observe filtering alongside rerank scores |
| Search mode | Hybrid search | Code and model number queries suit full-text or hybrid search; natural language questions suit semantic search |
| Image index | Generated at ingestion | Whether an image is retrievable depends on whether an image description or image vector index was built at ingestion |
Three common errors
- Linking knowledge bases that use different vector models. The input is two knowledge bases with different model identifiers. Compare the model identifier field; if they differ, exclude the mismatched one and do not run a cross-model retrieval. The observable result is a linked list that keeps only knowledge bases with matching model identifiers.
- Treating minimum relevance as a permission or entity filter. The input is a product record that is semantically similar to the question but belongs to a different supplying entity. Compare the entity named in the returned cited source text; if the entity does not match, record it as pending verification rather than treating the permission filter as effective.
- Describing RRF as a toggle switch. The input is the hybrid search parameter interface. Compare whether an RRF switch field exists; the documentation describes RRF as a fusion process and does not document an independent switch, so this item is recorded as having no corresponding configuration field.
How to verify the setup
Hypothetical paper exercise sample. Source text A: "Central procurement contract HT-2024-031 specifies the supplying entity as Distribution Center A, and product code SP-8821 with a specification of 500-gram bags, valid until 31 December 2024." Source text B: "Store requisition YHD-2024-1107 requests product code SP-8821, quantity 200 bags, for the Chengdong store." Proposed draft sentence for review: "SP-8821 is supplied by Distribution Center A to the Chengdong store, with a supply term ending 31 December 2024." When reviewing this sentence, compare it against the contract text, which states only a validity date and does not state a supply term. Record the supply term as unsupported and preserve the evidence gap.
- Input source text A and source text B above. Compare the product code and specification fields. If both documents use the same code and the specification wording matches, the index can cover that product; if the specification wording differs, record it as pending verification.
- Input the proposed draft sentence. Compare its "supply term" against the contract's "valid until" field. If the source text does not state a supply term, judge the sentence as unsupported rather than contradictory.
- Input a query containing only the product code. Compare whether the returned cited source text contains that code. If the returned array is empty, check the downstream design against the documented node behavior that output length can be 0. Do not claim that a retrieval rule always returns empty.
For operational handover, note that retrieval and form submission are not authorization. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a document field is not provided, it remains unknown.