What the documents look like
Knowledge-base maintenance for customer-service enquiries sits before refund approval. It organises FAQ entries, script templates, product parameter tables and escalation tickets so that store staff answer from one agreed wording. Materials are refreshed weekly or monthly, and new-product parameters plus high-frequency question summaries often lag behind live enquiries. Citation tracing decides whether a training sheet can actually be checked against its source.
After-sales question statistics and quality-inspection sampling show how problems are distributed. They do not state the authorisation status of any single reply. For store service and internal training, every script line needs to point back to a specific FAQ entry or parameter-table version.
Constraints at the citation tracing step
A knowledge-base search node can be linked to one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. The retrieval output then goes to the answering step, where each answer is checked line by line against the cited source text.
Semantic recall and relevance filtering affect content ordering. They do not perform permission or entity isolation, and retrieval on its own is not authorisation. Checking the currency of materials requires testing the current version and a historical version separately as sources.
Hypothetical drill sample: the source text reads, "Store A applied the standard labour-hour quotation for X-type water-purifier filter replacement between 1 March 2025 and 31 March 2025; quotation number PX-2025-031; applicable store: Store A." The proposed sentence under review reads, "Store A applied the standard labour-hour quotation for X-type water-purifier filter replacement." That sentence omits the period and the quotation number, so it is unsupported and needs the valid scope supplied before any judgement.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Relation to the sample |
|---|---|---|
| Linked knowledge bases | Confirm the selected knowledge bases and material versions; only identical vector models can be linked together | The sample source must fall inside the selected knowledge bases |
| Citation limit | Controls the length of cited content per search in tokens | The proposed sentence needs the full source scope quoted |
| Minimum relevance | A filter threshold that applies in semantic search or with result re-ranking | Not used as a permission or entity filter |
| Search mode | Semantic, full-text or hybrid search; hybrid results are merged through RRF | Number-type lookups favour full-text or hybrid |
| Citation display check | Use a traceable sample to check the displayed source text | Compare the proposed sentence against the sample source |
| Material currency check | Test sources with the current version and a historical version | The sample period must match the version |
Hypothetical trial starting value for this page: citation limit set to 2000 tokens. The reason is that the sample source text plus the proposed sentence total roughly 200 characters, so 2000 tokens can hold several citations. Compare against a 500-token starting value to observe whether the sample source text is truncated. This figure is a paper-exercise setting, not an official default.
Three common errors
- Treating retrieval output as an already authorised reply. The search node returns a citation array whose length may be zero, and that does not indicate business approval. Check: input the sample source text, compare whether the proposed sentence carries the period and the quotation number, and record it as pending verification when they are absent.
- Replacing source checking with semantic similarity. A similarity score does not constitute an entity or permission filter. Check: compare the sample source text and the proposed sentence field by field for subject, object and period; similarity alone does not satisfy the condition.
- Linking knowledge bases that use different vector models. The documentation requires the same vector model for joint linking. Check: first verify the knowledge-base identifier and the vector-model identifier, exclude those that do not match, and record what is actually observed without running a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the sample source text and the proposed sentence, and compare the period and quotation-number fields. Expected observation: once the proposed sentence adds "1 March 2025 to 31 March 2025" and "PX-2025-031", it matches the source and counts as supported; without those additions it is recorded as unsupported.
- Input an empty citation array [] and compare the design of the downstream answering step. Expected observation: the downstream step handles the empty citation and records the gap. This does not claim that a fabricated-rule retrieval will necessarily return an empty array.
- Input the sample source text and compare the displayed citation content with the source. Expected observation: the displayed content matches the source word for word, and answer fluency is not used in place of source evidence. All three are paper expectations; no deployment or test has taken place.