What the documents look like
In refund-only disputes, the returned goods often never reach a warehouse. The evidence therefore centers on dispatch and signed receipt, plus the assessed value of the product. After-sales supervisors and platform appeal specialists pull four document types into product-information and selection questions: the refund-only application, the dispatch voucher, the signed-receipt proof, and the appeal materials. These come from the platform after-sales console, carrier tracking, customer-service chat, and warehouse outbound records. The material is updated per order and daily.
Timing directly affects citation tracing. If the application version, dispatch voucher, and signed-receipt record for one order change across days, a citation must state which version and which date range were used. Without that, the cited source may point at an older version of the material.
A second property matters here. Knowledge-base search can select one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero. When nothing is found, the output link still runs. Citation tracing requires checking each answer statement against the cited source text; a thinking process or a fluent answer does not substitute for source evidence. Semantic search, full-text search, hybrid search, and Rerank are retrieval paths. A similarity score is a relevance measure and does not perform permission or entity isolation. The retrieval scope is set by the selected knowledge bases and the material versions. Knowledge bases built on different vector models cannot be associated together; the model identifier must be checked first and non-matching bases excluded.
Constraints at the citation tracing step
The retrieval chain is not a single path. Depending on the input and the search parameters, FastGPT combines text, images, semantic recall, full-text recall, question optimization, and reordering, then fuses the multiple result sets into citation content. RRF is a fusion step for combined results, not a switch that is turned on. Image retrieval requires an actual image input and the corresponding model capability; reordering requires a reordering model and must be enabled by configuration. The optional paths in the overall flow diagram are not all executed on every request.
The reference limit is expressed in tokens per search, not in a top-k item count. Minimum relevance is a filter threshold, not a final sort rule, and it applies only under semantic search or when result reordering is used. After question optimization, hybrid search, image retrieval, or reordering, the final result may come from fused multi-path recall and does not necessarily follow a single vector similarity order.
A similarity setting does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization require their own downstream implementation. No scheduled trigger or hidden control is assumed.
Assumed paper exercise sample. Material text for checking: "Order A-2024-0317. The warehouse outbound record dated 18 March 2024 shows the goods were dispatched. Carrier tracking dated 20 March 2024 shows signed receipt." Draft sentence to be judged: "The order has been signed for; approve the refund-only request." The draft cites the signed-receipt fact but does not cite the refund-only application content or the appeal materials, and does not state the authorization boundary. The request is pending review. Pending review does not demonstrate a lack of operating permission, and it is not an approval.
Configuration and implementation checks
| Configuration item | Assumed starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the base holding the refund-only application and the dispatch voucher | Check that the selected bases share the same vector-model identifier; exclude bases on a different model |
| Search mode | Full-text search | Order numbers and voucher numbers are exact queries; compare hit results against semantic search |
| Reference limit | 4000 tokens | Assumed starting value; compare citation completeness against a 2000-token result |
| Minimum relevance | 0.3 | Assumed starting value; compare the number of unrelated fragments against a 0.5 result |
| Citation content output | Pass the array to the answering step | Check that each answer statement corresponds to cited source text |
| Material timeliness check | Use the March 2024 version | Compare against historical versions to confirm the citation points at the currently valid version |
Three common errors
- Treating a similarity score as a permission filter. Input: the draft sentence and the retrieval results. Comparison: the score against the authorization status. Observable result: a high-scoring item is cited even though the material is not authorized. Check: confirm whether each retrieval result lies within the authorized scope; authorization status must be confirmed separately.
- Associating knowledge bases across different vector models. Input: two bases on different vector models. Comparison: the model identifiers. Observable result: bases with non-matching identifiers are still selected. Check: verify the identifiers and exclude non-matching bases; do not run a cross-model retrieval to compare missed recall.
- Treating an empty citation array as a workflow error. Input: the paper input []. Comparison: the downstream receiving design. Observable result: the downstream step does not handle the empty array. Check: confirm whether the downstream step is designed to receive a citation array of length zero.
How to verify the setup
- Verification one: input the assumed sample material text and the draft sentence. Compare whether the draft cites the refund-only application content and the authorization status. The condition is met when the draft cites only the signed-receipt fact and does not cite the application content, which indicates an incomplete source.
- Verification two: input a full-text search request for order A-2024-0317. Compare whether the returned citations include the dispatch voucher and the signed-receipt proof. The condition is met when the citation array contains both voucher texts, which indicates the retrieval path matches the sources.
- Verification three: input the paper empty array []. Compare whether the downstream answering step receives it normally. The condition is met when the downstream step does not error and outputs an empty-citation notice, which indicates the empty-result handling matches the design.
All three are paper checks. No live test is claimed.