What the documents look like
Complaints and official replies in this category rest on four document families: complaint work orders, regulatory inquiry letters, order evidence, and rectification reports. They arrive from the 12315 platform, platform complaint centres, customer service chats, and order systems. Each document is generated around a specific case, and updates are irregular, per case. The people handling them are customer service supervisors, legal and compliance staff, and store owners. Errors carry consequences that include administrative investigation, platform penalties, and brand exposure.
For product information and selection questions, this material does a specific job: it organises product details, order evidence, and rectification positions into retrievable entries that can be cited when drafting a reply. Retrieval quality depends on how that material is written and indexed. A passage with complete structure, a clear subject, and a single meaning retrieves more accurately through vector search. Exact queries such as a case number, order number, or model code depend on full-text search. One group of source data can map to several vectors, and a hit on any one of them brings back that group.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. The library is the smallest search unit; a collection only classifies data and does not affect search results. Vector retrieval is handled by the PG Vector plugin in PostgresSQL with an HNSW index. The dataset.datas table in MongoDB stores the source data of vectors, and its indexes field records the corresponding vector IDs, so one group of source data can correspond to multiple vectors. When results merge, the vector score taken is the highest one.
The retrieval chain combines semantic search, full-text search, hybrid search, question optimisation, reranking, and image retrieval according to the input and the search parameters. Hybrid search merges multiple result paths through RRF. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. Cited content is returned as an array whose length can be zero, which means the downstream path still runs even when nothing is retrieved.
Retrieval itself does not perform authorisation checks. A similarity score does not separate case ownership, subject permissions, or business categories. Knowledge bases built on different vector models cannot be linked together; the model identifier must be checked first and non-matching bases excluded.
Configuration and implementation checks
The following is a hypothetical exercise sample used to show how checking works. It does not represent a real record or a completed test. Source text: a store received a platform-transferred work order on the hypothetical date 10 March 2025. The work order states order number A-0001, records that the complaint concerns a mismatch between the specification shown on the product page and the delivered item, and shows the handling status as awaiting the store's explanation. Draft sentence to be assessed: "This store confirms that the order has been refunded and that no specification problem exists." Checking point: the source text records only that an explanation is pending. It records no refund and no specification verification conclusion, so neither fact in the draft sentence is supported.
| Configuration item | Relation to this page's documents | Checking method |
|---|---|---|
| Linked knowledge bases | Select the bases holding complaint work orders, regulatory inquiry letters, order evidence, and rectification reports | Compare vector model identifiers across bases and exclude those that do not match |
| Search mode | Use full-text or hybrid search for order numbers and model codes; use semantic or hybrid search for natural-language questions | Run the sample source text and the draft sentence separately and observe the candidate content |
| Citation limit | Measured in tokens; work orders and rectification reports differ greatly in length | Set a hypothetical starting value of 2000 tokens for this exercise, then compare against what shorter work orders actually need and adjust |
| Minimum relevance | Takes effect only under semantic search or result reranking | Label relevant and irrelevant items in the sample first, then compare the threshold |
| Question optimisation | Used for coreference resolution and question expansion | Compare the original question with the optimised question and check whether the subject and order number were altered |
| Result reranking | Requires a rerank model enabled under the settings | Observe rerank scores and candidate order, and confirm whether they are used for filtering |
Three common errors
- Treating a retrieval hit as authorisation or as confirmation of fact. Input the sample source text and the draft sentence, compare the retrieval result against what the source text actually states, and note that a draft sentence asserting a refund the source text does not record goes beyond the evidence.
- Treating minimum relevance as an entity or permission filter. Input unrelated items from the same document set, compare their scores against the filter condition, and note that expecting the threshold to exclude other subjects or unauthorised records misuses the setting, because retrieval does not perform authorisation checks.
- Linking knowledge bases that use different vector models. Input two bases whose vector model identifiers differ, compare the identifier fields, and note that linking them directly and judging recall from the result exceeds the documented condition. The identifier should be checked and non-matching bases excluded.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the refund fact in the draft sentence against the pending-explanation status recorded in the source text. The condition is met when the draft sentence does not assert an unrecorded fact and keeps only a pending-verification formulation.
- Input order number A-0001 from the sample. Compare the candidate content returned by full-text search and by semantic search. The condition is met when an exact-number query brings back the source data entry containing that number. When it does not match, record that as an observation and do not infer a cause.
- Input a paper input explicitly marked as empty, []. Compare whether the downstream node accepts a citation of length zero. The condition is met when the path continues to execute and produces no fabricated citation. No claim is made that any rule-based search necessarily returns an empty array.