What the documents look like
Creator-led sales campaigns rest on external creator performance and commission settlement, so store service and internal training materials must trace four document types: creator cooperation agreements, commission setup sheets, sample shipment forms, and livestream performance reports. These record the cooperating parties, the commission basis, sample destination, and per-session transaction results. Missing any one removes the source a training explanation depends on. Creator historical data and store transaction details commonly come from selected alliance and Juliang Xingtu sources, updated per session or weekly, and any citation in training must identify the document version and the session it belongs to.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model and outputs cited content as an array whose length may be zero; an empty array still completes the output path. Retrieved results are candidate materials only. They do not constitute authorization, a completed cooperation, or confirmed commission. Similarity and minimum relevance belong to recall ranking; they do not perform entity isolation or permission filtering. Hybrid retrieval merges results through RRF, reranking requires a model and must be enabled by configuration, and image retrieval requires image input and model capability. These paths do not run on every request.
Hypothetical drill sample: a brand business representative and an MCN agency sign a cooperation agreement for one creator session. The agreement states: "The cooperation session is June 18, 2025; the commission ratio follows the commission setup sheet; signature on the sample shipment form marks completion of material handover." A store drafts a training handover sentence: "This creator cooperation is now effective, commission settles per the setup sheet, and the store may ship first." The request to be judged: "Please confirm this session's cooperation is effective and authorize the store to ship." The agreement states only the session, the commission basis, and the meaning of sample signature. It states no effectiveness condition and no shipping authorization.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Only the creator cooperation material base with the same vector model | Compare vector model identifiers with another base; exclude when identifiers differ |
| Search mode | Hybrid retrieval | Compare keyword-type hits against a semantic-only path |
| Minimum relevance | Set to 0.4 for this trial | Compare recalled entries at 0.3 and 0.5 to observe unrelated fragment changes |
| Citation cap | Set to 3000 tokens for this trial | Compare with 2000 and 4000 to observe truncation of long performance reports |
| Question optimization | Enabled | Compare recalled entries against the disabled state on multi-turn follow-ups |
| Citation intake | Compare each answer sentence against cited source text | Observe whether the answer states authorization absent from the citation |
Three common errors
- Treating a retrieval hit as effective cooperation or shipping authorization. Input is the request above; retrieval returns the agreement fragment. Outputting "shipping authorized" treats recalled material as an authorization fact.
- Forcing knowledge bases with different vector models into one link. Input is two bases with different model identifiers. After comparing identifiers, one must be excluded. Continuing to link them and asserting missed recall or redundancy invents a runtime result.
- Treating an empty citation array as a workflow error. Input is a paper
[]. The downstream path still completes. Declaring retrieval failure or automatic retry lacks documentary basis.
How to verify the setup
- Input the agreement text above and the drafted sentence. Compare whether the draft contains "effective" or "may ship," which the source does not state. If present, the condition is unmet; change to pending verification and record the evidence gap.
- Input two knowledge bases with different vector model identifiers. Compare whether identifiers match; only matching identifiers permit joint linking. When they differ, record the exclusion result and do not run a cross-model retrieval comparison.
- Input a paper
[]as citation output. Observe whether the downstream answer step still completes and truthfully states that no material was retrieved. Completion without fabricated sources satisfies the condition.