What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so store materials sit inside the brand's own systems rather than under a franchise contract. For product information and selection questions, four document types carry most of the load: daily sales reports, shift schedules, display photographs, and inventory transfer orders. These materials describe model numbers, sizes, colours, display requirements, and the basis for a transfer.
Daily sales reports and transfer orders lean toward codes and proper nouns, which suits full-text retrieval. Display photographs and selection descriptions lean toward natural language and visual content, which suits semantic retrieval. Materials update daily and weekly, and a search reflects what has already been indexed at the time it runs. That distinction matters here: the state of a document in the source system and the state of its index are separate facts, and neither one is an approval.
Constraints at the vector indexes step
Knowledge bases joined in one search must share the same vector model. Confirm the actual model identifier and the input types it accepts before joining anything; bases built on different models cannot be associated together. A single source data block may map to several indexes, so index content and source text are checked separately rather than assumed to match.
Whether an image can be retrieved depends on what was built at ingestion time. A description index or an image vector index must exist for that image; uploading an image on the query side does not create one. Retrieval and relevance scores express similarity only. They do not perform permission checks, do not isolate one business entity from another, and do not establish that a record is authorized for a given reader.
Reference content is returned as an array and its length may be zero. A downstream path still has to handle that empty result rather than treat retrieval as guaranteed. Question optimization, image retrieval, and reranking are optional paths in the overall flow; each requires its own input and model capability, and none of them should be read as always executing.
Hypothetical paper exercise: a brand-owned store submits an inventory transfer order. The source text reads: "Transfer order number DB-2024-0317, from store A, to store B, item is spring knitwear, size M, navy, quantity 12 pieces, application date 17 March 2024, status pending regional manager review." The draft sentence under review reads: "This transfer order has been approved by the regional manager, and inventory has been synchronized to store B." The source text records only a pending-review status. Approval and inventory synchronization are not supported by it.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | How to check it and why |
|---|---|---|
| Vector model | Joined bases share one identifier | Compare each base's model identifier and accepted input types; exclude any base whose identifier differs |
| Index content | Transfer orders indexed in segments by order number, item, and quantity | Compare each index segment against the source text; weigh coverage against length using this sample |
| Image index | Display photographs generate a description index at ingestion | Inspect the ingestion record for a description index or image vector index |
| Search mode | Code-based questions use full-text or hybrid retrieval | Compare hit differences against semantic retrieval on the same question |
| Reference limit | Hypothetical starting point measured in tokens | Compare result stability across chunks of differing length |
| Minimum relevance | Hypothetical starting point between 0 and 1 | Compare against semantic or rerank scores; treat it as a filter threshold only |
These are implementation checks for this page, not documented defaults. The token-based reference limit and the 0-to-1 relevance range follow the documented behaviour of those two settings; the specific values chosen here are exercise starting points and carry no official status. RRF is the fusion step that combines results from several recall paths. It is a process inside hybrid retrieval, not a switch to be turned on.
Three common errors
- Treating an image uploaded at query time as proof the image is retrievable. The error is that ingestion indexing is overlooked. Check: submit the same display photograph as a query, compare the ingestion record for a description index or image vector index, and record the result as unsatisfied when no such index exists.
- Treating a pending-review status as approval. The error is that two distinct status values are merged. Check: input the transfer order source text above and the draft sentence, compare the status field, and note that the source reads pending regional manager review, so the draft's approved claim is unsupported.
- Joining bases built on different vector models. The error is that the same-model requirement is skipped. Check: input the model identifiers of two bases, compare them, and exclude any base whose identifier differs. Record the observed identifiers and the exclusion; do not run a cross-model search to compare what is missed.
How to verify the setup
- Input the transfer order source text and the draft sentence. Compare whether the draft sentence matches the status field in the source. When the source reads pending review, the draft sentence is not satisfied. The expected observation is a status mismatch, and the gap in evidence for approval and synchronization stays open.
- Input the same display photograph as a query. Compare the ingestion record for a description index or image vector index. When such an index exists and the query returns a hit, the condition is satisfied. When no index exists, record the gap rather than inferring a retrieval outcome.
- Input the model identifiers of two bases. Compare whether they are identical. Only identical identifiers permit association; differing identifiers lead to exclusion and a recorded observation of the actual identifiers, without predictions about runtime behaviour.
For the empty-reference path, a paper input of an empty array can be used to check downstream design. That check confirms the downstream path handles a zero-length result; it does not assert that any particular query returns an empty array. Retrieval remains separate from authorization: a similarity score, a relevance threshold, or a successful search does not grant permission, does not verify a business fact, and does not write to a database or synchronize an external system. Those actions require their own downstream implementation, and none of them is documented as automatic here.