What the documents look like
Livestream after-sales work concentrates after a single traffic session. Refund applications, reshipment orders and livestream order lists arrive in the same window, and support staff must decide which procurement or supplier document a given after-sales case should cite. The procurement and supplier materials for this category include purchase orders, arrival records, supplier quotations and contract text. In livestream after-sales handling these materials are used to check the source of a reshipment, the procurement batch and the supplier's scope of responsibility. Order and after-sales data update after each session and daily, so tracing a citation requires stating the version date and validity scope of the document being cited, and avoiding an old quotation as the basis for a current reshipment.
The knowledge base search node can be linked to one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. Whether a retrieval result is adopted still requires a person to compare it against the source text and its valid version; retrieval itself does not constitute authorization. The minimum relevance value takes effect only under semantic retrieval or result reranking, and it is a filter threshold rather than a permission or entity isolation mechanism. Hybrid retrieval merges multiple result paths through RRF, reranking requires a model and is enabled according to settings, and image retrieval requires image input and model capability. These paths do not all execute on every request.
Hypothetical drill sample: a livestream after-sales supervisor processes a reshipment order on 10 June 2025 and intends to cite a supplier quotation. The quotation text states: "This quotation is valid until 31 May 2025, applies to batch A care products, and the supplier is a named supplier." The draft sentence to be judged is: "According to this quotation, this reshipment is supplied by that supplier at the original price." This sentence involves three points to check: the quotation validity has passed, the batch object, and the supply commitment. The sample does not state signing or effectiveness clauses.
Constraints at the citation tracing step
The search node returns citation content as an array with length possibly zero, so an empty citation chain can still pass through the workflow. A zero-length output does not mean the workflow has errored or terminated. What the downstream answer node does with that array is a separate design question.
The minimum relevance value is a 0-1 number that filters low-relevance results and applies only under semantic retrieval or result reranking. It is a filter threshold, not a final ordering rule, and it does not exclude unrelated businesses or unauthorized records. Reranking produces a 0-1 score representing relevance between the searched content and the question, and that score can be used for filtering. RRF merges reranked results, vector search results and full-text search results into a final result set; it is a fusion process, not a switch that is turned on.
Full-text retrieval suits exact queries such as numbers, models, proper nouns and error codes. Semantic retrieval depends on vector similarity and suits natural-language questions and semantically close content. Hybrid retrieval uses both and merges them through RRF. Image retrieval adds image description or image vector paths and requires corresponding image input and model capability. These optional paths are listed in the overall flow and do not all run on every request.
Knowledge bases linked together must share the same vector model. This requirement applies to the knowledge base association itself, and it does not extend to external systems such as business registration sites. When vector model identifiers differ, the correct action is to check the identifiers, exclude the non-matching base and record the actual observation, rather than running a cross-model retrieval to compare missed recalls.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | A base containing purchase orders and supplier materials | Compare vector model identifiers across bases; exclude any that differ |
| Search mode | Hybrid retrieval | Compare recall of exact terms such as batch numbers against pure semantic retrieval |
| Minimum relevance | Hypothetical starting value 0.5 | Compare candidate counts after filtering at 0.3 and 0.7, and observe whether valid materials remain |
| Citation limit | Hypothetical starting value 2000 tokens | Compare citation completeness at 1000 and 4000 tokens to avoid truncating key clauses |
| Result reranking | Enabled according to settings | Compare candidate ordering before and after enabling, and confirm the rerank model is available |
| Question optimization | Enabled for multi-turn follow-ups | Compare coreference resolution against the disabled state, and observe added latency on short questions |
These values are hypothetical starting points for this paper drill, used for comparison with adjacent values. They are not official defaults.
Three common errors
- Treating a retrieval hit as authorized execution. Input is the quotation retrieval result above; compare the retrieval output against the authorization boundary and observe whether the cited content is written directly as a supply commitment. Retrieval does not produce authorization, and the supply conclusion in the draft sentence requires separate verification.
- Conflating quotation validity with the supply period. Input is the quotation text "valid until 31 May 2025"; compare the draft date against the validity date and observe whether an expired quotation is treated as a current valid basis. Validity does not express a supply period, and the two are different fields.
- Treating a pending-review state as approved. Input is the pending-review marker on the reshipment order; compare the review state against the execution state and observe whether the pending review is written as an arranged reshipment. Pending review does not show a lack of operating permission and does not constitute approval.
How to verify the setup
- Input the quotation text and the draft sentence from the hypothetical sample; compare the supply conclusion in the draft sentence against the validity date and batch object stated in the source text, and observe whether the draft sentence exceeds what the source supports. The expected result is that it exceeds the source. If the draft sentence is changed to cite only the quotation facts and note that validity has passed, the condition is met.
- Input the pending-review state of the reshipment order from the hypothetical sample; compare the review state against the execution action, and observe whether approved or reshipped wording appears. The expected result is that it is unsupported. If the output retains a pending-verification marker, the condition is met.
- Input the paper empty result array
[]; compare whether the downstream answer step accepts it normally, and observe whether the workflow breaks because of the empty citation. The expected result is that an empty array can pass through. If the downstream output states that no material was retrieved, the condition is met. This check does not claim that a fictional rule retrieval will necessarily return an empty array.
Knowledge bases with different vector models cannot be linked together. Check the identifiers first and exclude non-matching bases, then record the actual observation. Do not run a cross-model retrieval to compare missed recalls. Cases that are not stated, or that have only related-stage evidence, are recorded as unsupported or pending verification, and the evidence gap is preserved.