What the documents look like
Livestream after-sales work concentrates after a single traffic session. Refund requests and reshipment orders arrive in the same window, and staff move between customer service IM, shop after-sales records, warehouse WMS, and logistics tracking to check one case against another. A host's on-stream promise and the actual order state often diverge in time. When an after-sales supervisor hands over a queue, a relayed conclusion without its source leaves the next person unable to tell which material and which version produced it. Script templates and livestream order lists are copyable documents, so an old version can stay in circulation across a handover.
These traits make citation source and tracing central to permission handover. The receiving operator needs the identifier, version, and status of the original material before judging whether the next action is within that operator's authority. The materials in scope for this page are the livestream order list, refund requests, reshipment orders, and script templates.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and the length may be zero. A zero-length output does not stop the downstream chain; the chain still runs. Retrieval output is passed to the answering step, where each answer statement is checked against the cited original text.
The citation cap controls how many tokens of content a single search may cite. The minimum relevance value is a filter threshold, and it takes effect only under semantic search or when result reranking is used. A similarity score does not perform permission or entity isolation, and a retrieval result does not constitute operational authorization. Linked knowledge bases must share the same vector model to be linked together; knowledge bases with different vector models cannot be linked together, so identifiers are checked first and non-matching ones are excluded. Full-text search suits exact queries such as numbers, models, proper nouns, and error codes. Semantic search suits natural-language questions. Hybrid search merges multiple result paths through RRF. Reranking requires a rerank model and must be enabled in the settings.
Hypothetical drill sample. The material text to be checked reads: "After-sales handover form SH-0417 states: the object is 37 refund requests from session A on 17 April, current status is pending review, and the handover scope is limited to checking refund reasons against logistics signed-receipt records, excluding approval and refund execution. Script template V3.2 took effect on 16 April and applies to session A." A draft sentence to be judged reads: "The 37 refund requests listed in SH-0417 have completed review, and the receiving operator may execute refunds." The field needed for judgment is the review status. The original states pending review, and the draft states completed review; these are mutually exclusive values of the same field, so the draft sentence is unsupported. The authorization boundary is checking, excluding approval and refund execution, and the draft extends checking into execution, which exceeds the authorized scope.
Configuration and implementation checks
| Check item | Relevant setting or implementation check on this page | Correspondence with the sample |
|---|---|---|
| Linked knowledge bases | Link only knowledge bases with the same vector model; check the knowledge base identifier and material version | Check whether SH-0417 and script template V3.2 fall in the same searchable range |
| Cited content output | Output is an array, length may be zero, and the downstream chain still runs on an empty array | Use a paper input of [] to check whether the downstream answering step handles empty citations |
| Citation input handoff | Retrieval output is passed to the answering step, and each answer statement is checked against the cited original text | Check whether the draft sentence cites the pending-review status in the SH-0417 original |
| Citation display check | Use a traceable sample to check the display of the original text | Check displayed content against the SH-0417 and V3.2 originals |
| Material timeliness check | Test the source with the current version and a historical version | Check script template V3.2 against an older version and confirm which version is adopted |
| Search mode | Exact numbers use full-text or hybrid search; natural language uses semantic search | Use full-text search to check order number SH-0417 |
Hypothetical trial starting point for this page: the citation cap is set to 2000 tokens, chosen because the sample material is short and this makes it easier to observe whether citation trimming affects the completeness of the original text; it is compared with a 500-token starting point to observe whether the original text is truncated. This value is a paper-drill setting for this page and is not an official default. The minimum relevance value is set to 0.5 and compared with 0.3 to observe whether the pending-review original text is still recalled. All of the above require manual checking and do not constitute automatic execution.
Three common errors
- Treating a retrieval hit as operational authorization. The input is the SH-0417 original, the comparison object is the draft sentence "may execute refunds," and the original authorization boundary is checking, excluding approval and refund execution. The observed result is that the draft sentence exceeds the authorized scope and is unsupported.
- Writing pending review as completed review. The input is the SH-0417 original, the compared field is review status, the original states pending review, and the draft states completed review. These are mutually exclusive values of the same field and form a contradiction.
- Linking knowledge bases that use different vector models. The input is two knowledge base identifiers, and the compared item is the vector model identifier. When they do not match, the pair is excluded, and no cross-model retrieval is run to compare missed recalls.
How to verify the setup
- Input the SH-0417 original and the draft sentence, and compare the review status field. When the original states pending review and the draft states completed review, the draft sentence is judged unsupported. The expected outcome is to record the contradiction and not execute a refund.
- Input a paper citation array of [], and compare the downstream answering step design. On an empty array the chain still runs and the answer does not fabricate a source. The condition is met when the downstream step handles the empty citation and records the gap.
- Input two knowledge base identifiers and their respective vector model identifiers, and compare whether they match. When they do not match, exclude the joint link; when they match, further check the material versions and whether SH-0417 and V3.2 fall in the same searchable range. The expected outcome is to record the check result, and the retrieval result does not constitute authorization.