What the documents look like
Outlet discount operations work from four recurring document families: discount approvals, slow-moving inventory lists, price tags, and gross-margin calculation sheets. Their source data comes from ERP inventory age, POS sales, outlet mall campaigns, and historical discount records. Updates follow a weekly and monthly cadence, so inventory-age bands, discount tiers, and slow-moving status can differ between versions of the same material.
A single SKU may appear in an already approved discount sheet, a pending approval request, and an older price tag at the same time. Product information and selection Q&A is meant to help a user locate candidate clearance SKUs by inventory age, category, and discount tier, together with the related approval and margin materials. Vector models and indexes only carry out recall. They do not perform approval or authorization judgment.
Constraints at the vector indexes step
FastGPT builds knowledge bases with an Embedding approach, and vector distance expresses mathematical similarity. Knowledge bases joined in one retrieval must use the same vector model. The actual model and input type are checked first; different models cannot be joined together.
In MongoDB, the dataset.datas table stores source data, and an indexes field records the corresponding vector IDs. One data block can map to multiple vectors. In PostgresSQL, a vector field stores vectors, and the index is HNSW. During retrieval, vectors are recalled first, then the vector IDs are used to find source data in MongoDB. When entries correspond to the same source data, they are merged and the highest vector score is kept.
Index precision improves when index content is complete and focused, and adding more index entries can improve coverage. The two need to be weighed against the sample. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion. Uploading an image at query time does not guarantee a hit. Similarity and minimum relevance are filtering thresholds; they do not create permission or entity isolation, and retrieval does not produce authorization.
Hypothetical drill sample: material text reads "Discount approval A-07: subject is a clearance group at an outlet store, object is 12 SKUs of 2024 spring women's shoes, status is pending approval, effective scope is the July 2024 discount tier." "Slow-moving list L-03: same store, object is the same 12 SKUs of 2024 spring women's shoes, status is listed as slow-moving, scope runs through June 2024." The draft sentence to be judged is "A-07 has approved this batch of SKUs for execution at the July 2024 discount tier." The question is whether that draft sentence is supported by the materials.
Configuration and implementation checks
| Configuration or check item | Documented content | Implementation check on this page |
|---|---|---|
| Vector model | Knowledge bases joined in retrieval must use the same vector model | Record the model identifier used by each joined base; when identifiers differ, exclude the base and do not run a cross-model retrieval comparison |
| Source data and index | One data block can map to multiple vectors; indexes records vector IDs | Check whether the sample SKU's source data block and its index entries point to the same data group |
| Index content | Shorter, more accurate index content raises precision and reduces retrieval range | Build indexes separately from the sample discount approval and slow-moving list, then compare whether recalled source text stays focused |
| Image index | Image retrieval depends on whether image descriptions or image vectors were generated at ingestion | If the sample includes a price tag image, first confirm whether the corresponding index was generated, then choose the query method |
| Retrieval mode | Semantic, full-text, and hybrid retrieval can be combined; hybrid retrieval is merged through RRF | For SKU queries containing identifiers, record the actual mode and the cited source text; do not judge results from vector scores alone |
| Citation and relevance | The citation limit is controlled by tokens; minimum relevance applies only to semantic retrieval or result reranking | Record the citation limit and minimum relevance used here, and compare each cited passage against the sample materials |
The citation limit is a starting value for this hypothetical drill, chosen because the sample set is small and the comparison target is the number of cited passages needed to cover A-07 and L-03 in full. It is not an official default.
Three common errors
- Treating a pending discount sheet and an already listed slow-moving list for the same SKU as the same status. In the sample, A-07 has status pending approval and L-03 has status listed as slow-moving. Subject and object match, but the status field values differ. Mixing them produces the conclusion that the batch has already been approved.
- Joining knowledge bases that use different vector models and then comparing recall counts. When model identifiers differ, the base is excluded first. The join itself does not produce a comparable result, and inferring missed recall from a count difference lacks support.
- Using similarity or minimum relevance in place of an authorization judgment. In the sample, A-07 is pending approval. A retrieval hit on that sheet only shows the material was recalled. It does not show the discount was approved or can be executed.
How to verify the setup
- Input the sample material text and the draft sentence, then compare the status field value "pending approval" in A-07 against "has approved" in the draft sentence. Same subject, same object, same field, and mutually exclusive values. The draft sentence is not supported by the materials and is recorded as pending verification.
- Input slow-moving list L-03 and discount approval A-07 from the sample, then compare the object SKU count and the scope end month in each. If the cited text contains both materials with their fields intact, the index distinguishes the two statuses. If only one is recalled, record it as insufficient coverage and do not infer that the other does not exist.
- Input a query that contains only a price tag image, and first confirm whether an image description or image vector was generated at ingestion. If not, record it as a missing image index. If yes, compare the recalled source text against the price tag content. Do not promise a first-hit or full recall. An empty result can be checked downstream with the explicit paper input
[], without concluding that a rule-based retrieval must return empty.
A model or index change is followed by rerunning the same question set and comparing cited source text. Vector scores alone do not support a promise about results.