What the documents look like
Seller ratings maintenance handles consumer feedback after post-sale fulfilment, and it does not directly set campaign prices. The daily materials are negative-review handling records, experience score details, service diagnostic reports and rectification ledgers. Data sources include Douyin Store experience scores, Tmall service indicators, JD after-sales reviews and customer service systems, with daily or weekly update frequency.
Product information and selection Q&A requires that an answer can be traced back to the exact wording of a specific material. Document version and retrieval scope therefore need to be confirmed before answer generation. The materials that matter most for citation tracing are the negative-review handling records and the experience score details, because those two carry the subject, the object, the status and the validity period that a citation comparison needs.
Constraints at the citation tracing step
The knowledge base search node selects 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 still lets the output path run through. Whether a retrieved result is actually used by the answering step must be checked item by item against the cited source text.
A similarity score and the minimum relevance value affect filtering and ordering only. They do not carry permission or entity isolation. Knowledge bases built on different vector models cannot be associated together; the vector model identifier is checked first and non-matching bases are excluded.
Hypothetical paper exercise sample. Source text one: "On 3 June 2025, store A negative-review handling record number P-0603, status pending customer service supervisor review, handler signature absent." Source text two: "On the same day the experience score details show this store service indicator at 4.6 points, data source Douyin Store experience score daily update." The draft request to be judged: "Please confirm that P-0603 has completed rectification and taken effect." The completion and effectiveness asked about are not stated in either material. A pending-review status does not mean approved, and an absent signature does not mean effective.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the negative-review handling record base and the experience score detail base whose vector model identifiers match | Compared with a base whose vector model identifier differs; a different identifier is excluded from association |
| Search mode | Hybrid search | Compared with pure semantic search; number P-0603 is an exact query, so full-text recall can add a hit |
| Question optimization | Off | The current question has a clear subject; compared with the case where turning it on adds latency |
| Citation upper limit | Hypothetical starting point 3000 tokens for this exercise | Compared with the default starting point, to observe whether the citations cover both source texts P-0603 and the experience score details |
| Minimum relevance | Hypothetical starting point 0.3 for this exercise | Compared with a higher threshold, to observe whether those two records are still recalled |
| Reply format | Keep default | Compared with json_schema; this task needs no fixed-structure output |
The citation upper limit uses tokens as its unit, while the thinking budget belongs to model thinking configuration; the two are not interchangeable. Rerank requires a model and must be enabled by setting. Image retrieval requires image input and matching model capability. The overall retrieval flow lists optional paths, and those paths are not all executed on every request.
Three common errors
- Treating a pending-review status as completed rectification that has taken effect. Check: input source text one, compare "pending customer service supervisor review" with "completed rectification", and observe that the status is not supported by the material.
- Treating a similarity score as a permission filter. Check: input the draft request, compare the retrieval output with the authorization scope, and observe that the score does not exclude unauthorized records.
- Treating an empty citation array as a workflow error. Check: input the paper citation array [], compare the downstream node design, and observe that the path continues to execute.
How to verify the setup
- Input source text one and the draft request, and compare "pending customer service supervisor review" with "completed rectification and taken effect". The judgment is met only when the material contains a rectification completion and effectiveness record with the same subject, the same period and the same field; otherwise record it as not supported.
- Input source text two, and compare the 4.6 points in the experience score details with the value cited in the answer. Citation tracing is met only when the answer value matches the source text and states the source as Douyin Store experience score daily update.
- Input the paper citation array [], and compare the downstream answering node design. The receiving design is met only when the downstream still executes under an empty array and does not fabricate a source. All of the above are expectations and were not actually executed.
Retrieval checks record observations and failure branches only; they do not guarantee a first-item hit or full recall.