What the documents look like
Complaints and official replies are directed at a platform or a regulator, so the material carries formal evidence and remediation wording rather than routine after-sales chat. The data sources named for this category are the 12315 platform, the platform complaint centre, customer service chat, and the order system. Inventory procurement and supplier records enter this category as material that gets called up: a complaint ticket may point at the procurement origin of a batch of goods, a regulatory enquiry letter may ask for supplier qualifications and purchase vouchers, and a remediation report may need to quote order evidence and supplier documents verbatim. Updates are irregular and per case; each complaint or enquiry corresponds to its own set of materials.
For this task, citation tracing means that when a reply is drafted or an evidence pack is assembled, the quoted inventory procurement records and supplier document text can be located, compared, and displayed, and the quoted content matches the actual material. The retrieval step only recalls candidate content. It does not decide whether material is authorised for use, whether it is within its validity period, or whether it falls inside the range that may be submitted externally.
Constraints at the citation tracing step
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. An empty array means the search path completed without recalling content, and downstream steps need to handle that case. Once citations are passed to the answering step, the answer and the cited source text must be checked line by line.
Hypothetical drill sample: a shop receives a complaint ticket transferred from a platform, ticket number HYP-2026-0417, concerning the procurement origin of a batch of vacuum flasks. One piece of source text available for comparison reads: "The supply confirmation issued by Supplier A on 10 March 2026 states that this batch of vacuum flasks numbers 600 units, specification 500 ml stainless steel model, with supply valid until 30 June 2026." A second piece reads: "The procurement contract between the shop and Supplier A was signed on 28 February 2026, and the contract status is signed." The draft sentence to be judged reads: "This batch of vacuum flasks was supplied by Supplier A in March 2026, and the supply relationship remained continuously valid during the complaint period." The source text records only the date of the supply confirmation and the signed status of the contract. It does not state that the supply relationship remained continuously valid during the complaint period, so the draft sentence is unsupported.
Minimum relevance is a filter threshold that takes effect only under semantic retrieval or result re-ranking, and it does not perform permission or entity isolation. The same-vector-model requirement applies among linked retrieval knowledge bases and does not extend to business registration sites or other external systems. RRF is the result fusion process after hybrid retrieval, an internal system step rather than a switch that can be turned on. Image retrieval requires corresponding image input and model capability, and re-ranking requires an available model enabled by configuration; the paths drawn in the overall document flow diagram do not mean every request executes all of them.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | One inventory procurement and supplier records base | This base belongs to the same case scope as the complaint ticket material; compare the vector model identifier of each candidate base for consistency |
| Search mode | Hybrid retrieval | The material contains identifiers and natural language description; compare what semantic retrieval and full-text retrieval each recall |
| Citation limit | Set to 3000 tokens as a hypothetical trial starting point | Estimated from material paragraph length; compare against the combined length of a single supply confirmation and contract summary |
| Minimum relevance | Set to 0.3 as a hypothetical trial starting point | Compare against the re-ranking score distribution and observe whether candidates below the value are filtered |
| Result re-ranking | Enabled by configuration | When candidates are numerous, compare ordering before and after re-ranking |
| Citation output | Accept the array output | Compare whether the source text quoted in the answer corresponds to the citation entries returned by retrieval |
These values are hypothetical trial starting points for this page and are not official defaults. The citation limit uses tokens and is distinct from the thinking budget; the two cannot be used interchangeably.
Three common errors
- Treating a retrieval recall as an authorisation confirmation. Retrieval returning a supplier record only shows the content was recalled; it does not show the record is authorised for external submission. Check by entering the citation entry returned by retrieval, comparing it against the authorisation approval record, and observing whether an approval record is missing.
- Treating contract signed status as contract effectiveness or continuous supply validity. In the sample the contract status is signed, and no effective date or supply period is recorded. Check by entering the draft sentence and the contract source text, comparing how each describes the supply period, and observing whether the draft sentence exceeds what the source text states.
- Treating an empty citation array as a workflow error. The search node output array may have length zero while the path still completes. Check by entering a paper empty array, comparing the design of the downstream answering step, and observing whether downstream handles the no-citation case.
How to verify the setup
- Enter the supply confirmation source text and the draft sentence from the sample, compare whether the draft sentence states that the supply relationship remained continuously valid, and observe whether the source text supports that wording. The source text does not state continuous validity, so the draft sentence is unsupported and the evidence gap must be recorded.
- Enter the procurement contract source text from the sample, compare whether contract signed status and contract effective status are the same fact, and observe whether the source text records an effective date. The source text records only signed status, so contract effectiveness is unsupported and requires separate verification.
- Enter a paper empty array, compare how the downstream answering step is designed to handle empty citations, and observe whether downstream can accept a no-citation case. The expectation is that downstream does not break on an empty array; actual behaviour is a matter of observation, and it is not assumed that retrieval will return an empty array.
Knowledge bases with different vector models cannot be linked together. When checking, first inspect the vector model identifier of each base, exclude those that do not match, and do not run cross-model retrieval to compare recall differences. Retrieval and forms only collect and return content; business verification, approval, database writes, and external synchronisation each need their own downstream implementation. Authorisation status, effective clauses, and supply periods that are not recorded remain missing and are not filled in by inference.