What the documents look like
Livestream sample selection and quality inspection happen before a broadcast. The work centers on whether a product can be sold and whether it is suitable for on-camera display. The daily materials are sample evaluation forms, quality inspection reports, supplier qualifications, and prohibited-item lists. Sources include supplier files, testing-agency reports, platform negative catalogs, and historical return rates. These materials update weekly and per item. They commonly carry Chinese and English product names, model numbers, report numbers, and test items. In cross-border work, one archive may combine several languages, so retrieval has to match across languages rather than only within one language.
Constraints at the vector indexes step
Vector search measures similarity through vector distance, and semantic retrieval has cross-language capability: a Chinese question can match an English knowledge point. A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. Knowledge bases associated for retrieval must use the same vector model, so the actual model and input type are checked first. One original data block can be linked to multiple indexes, and index content and original content are checked separately. Index content should be complete and focused; retrieval coverage and index length are weighed against the sample. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time, not only on the query side. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a model and must be enabled in settings. These optional paths do not all execute on every request. Semantic recall and relevance do not perform permission or entity isolation, and a retrieval result does not constitute authorization.
Hypothetical drill sample: the original text of a supplier qualification record for one batch reads "Supplier: Company A (example); qualification document: copy of business license; valid until: March 31, 2026; status is submitted, pending review." A draft handover sentence reads "Company A's qualification has passed review and can be scheduled for broadcast." The judgment requires the same subject, the same period, and the same field: the original status field is "submitted, pending review," while the draft sentence states "passed review." The two values are mutually exclusive on the same field, and the draft sentence is unsupported. Pending review does not show a lack of operating permission, and it does not equal passing review.
Configuration and implementation checks
| Check item | Assumed starting point for this drill | Comparison method |
|---|---|---|
| Vector model consistency | Associated libraries all carry the same vector model identifier | Check the identifier library by library, exclude inconsistent ones, and do not run cross-model retrieval |
| Index content scope | Each index contains supplier name, qualification type, validity period, and status | Compare field by field against the original text and record missing fields as gaps |
| Image index generation | A description or vector is generated for sample images at ingestion | Check the ingestion record for an image index; without one, the query side cannot match |
| Retrieval mode | Full-text or hybrid retrieval for questions containing model numbers or report numbers | Compare cited original text across the same question set and observe which fields are matched |
| Citation limit | Set to a token amount covering the main fields of one quality inspection report | Compare against the actual report length; a value that is too small truncates fields |
| Minimum relevance | Set to a level that retains the qualification status field | Lower it and observe whether unrelated fragments enter the citations; record the actual result |
The values in the table are assumed starting points for this drill, not official defaults. The citation limit is controlled in tokens, and minimum relevance takes effect only in semantic retrieval or with result reranking.
Three common errors
- Treating "submitted, pending review" as passed review. The input is the qualification original text above and the draft sentence. Compare the status field values and observe whether the draft sentence matches the original on the same field. A mismatch is recorded as unsupported.
- Associating libraries that use different vector models. The input is two libraries marked with different vector model identifiers. Compare the identifiers and observe whether they match. Exclude inconsistent ones and do not run retrieval to compare missed recalls.
- Treating a retrieval hit as authorized execution. The input is one retrieved qualification record. Compare the record status with the action in the handover request and observe whether the status supports that action. A pending-review status does not support scheduling for broadcast.
How to verify the setup
- Input the hypothetical qualification original text and draft sentence above, compare the status field "submitted, pending review" with "passed review," and observe whether the draft sentence is judged unsupported. The judgment is correct when that condition is met.
- Input two libraries with different vector model identifiers, compare whether the identifiers match, and observe whether the system excludes the inconsistent one. The model consistency check works when that condition is met.
- Input an empty citation result, compare whether the downstream design accepts an array of length 0, and observe whether the path completes. Empty-result handling is normal when that condition is met. Retrieval checks record actual observations and failure branches and do not guarantee a first hit or full recall.