What the documents look like
Wholesale stall trading runs on purchase notes, handwritten sales slips, stall lease contracts and payment receipts. Transactions are fragmented, negotiated on site and moved by shipping agents, and the records update daily or per order. Product reference and selection questions must pull product name, specification, quantity, unit price and the parties involved out of these materials before answering a buyer's selection question. Document form shapes index content directly: purchase notes lean on model codes and quantities, handwritten slips carry inconsistent handwriting and abbreviations, and lease contracts and receipts carry parties and periods. A single question can match several material types at once.
Constraints at the vector indexes step
Knowledge bases joined to one search must use the same vector model; libraries on different models cannot be linked together, so the model identifier is checked first and non-matching libraries are excluded. One source data block can map to several indexes, and index content is checked separately from the source text. Index content needs to be complete and focused, and retrieval coverage is weighed against index length using samples. Image retrieval also depends on whether an image description or image vector index was produced at ingestion time; the query side alone does not decide this. Semantic search relies on vector distance, full-text search suits exact lookups such as codes and model numbers, hybrid search merges results through RRF, and reranking requires a model and must be enabled in settings. The minimum relevance value is a filter threshold and does not perform permission or entity isolation. Retrieval is not authorization.
Configuration and implementation checks
| Configuration or check item | Basis and condition | Reason for the assumed starting point and comparison method |
|---|---|---|
| Vector model identifier | Linked knowledge bases must share one vector model | List the actual model identifier of each library first; when identifiers differ, exclude the non-matching library and do not run a cross-model retrieval comparison |
| Index content | Complete and focused, with length weighed against coverage | Assumed starting point: keep product name, specification and unit in each index; compare recalled source text against an index holding the full passage |
| Multi-index mapping | One data block can map to several vectors | Build a model-code index and a product-name index for the same source data, then check whether a hit merges back into the same source group |
| Search mode | Semantic, full-text or hybrid | Use full-text or hybrid for code-type questions and semantic or hybrid for natural-language wording, comparing quoted source text across one question set |
| Reference cap | Each search quotes at most n tokens | Assumed starting point: use the length of one purchase note as the reference; do not adopt a fixed number the documents do not state |
| Minimum relevance | A 0-1 value, effective only in semantic search or with reranking | Record how quoted source text changes before and after adjusting the value; do not infer a permission-filtering effect from it |
Three common errors
- Linking knowledge bases that use different vector models. Input two libraries with different model identifiers, compare the model identifier field, and exclude one library when the identifiers do not match; the observable result is that only same-model libraries take part in retrieval.
- Treating an image uploaded at query time as a searchable image. Input one product photo, compare whether the ingestion record contains an image description or image vector index, and when that index is absent the image does not enter the image retrieval path.
- Treating minimum relevance as a permission or entity filter. Input the same question against materials from different stalls, compare whether the returned references are decided by similarity alone; if returned content crosses entities, the value has not performed an isolation function.
How to verify the setup
Hypothetical paper exercise sample: the source text reads "On 12 March 2024, Stall A sold Buyer B one lot of rebar, specification HRB400, quantity 20 tonnes, unit price negotiated that day, payment receipt number A-017, balance outstanding." The proposed sentence reads "Stall A has received all payment from Buyer B." The judgment covers the same party, the same period and the same field: the source states the balance is outstanding while the proposed sentence states all payment is received, so the two take mutually exclusive values on one field and the proposed sentence is unsupported.
- Check one: input the source text above and the proposed sentence, compare the payment settlement status field, and when the source shows outstanding while the draft shows settled, judge the draft sentence as conflicting with the source.
- Check two: input the question "Which specifications exist for HRB400 rebar" and compare whether the quoted source text contains that model code; the condition is met when full-text or hybrid retrieval returns a purchase note containing the code. When no hit occurs, record the gap and do not infer that a rule must return an empty result.
- Check three: input one photo of stall goods and compare whether the ingestion record contains an image description or image vector index; when the index exists and the model supports image input, the image takes part in retrieval. When the index is absent, record it as not indexed and do not claim retrieval must fail.
Authorization boundary: retrieval or form collection supplies candidate materials only; business verification, approval, database writes and external synchronization require separate downstream implementation. A pending review status does not indicate granted authorization, and a quotation validity period does not indicate a supply period.