What the documents look like
Customer-service knowledge base maintenance is preparatory question-and-answer material organization. It does not approve refunds directly. The inventory procurement and supplier materials task handles supplier quotations, purchase contract clauses, arrival batch descriptions, out-of-stock substitution options, and escalation tickets. These materials come from the customer-service system, product detail pages, after-sales issue statistics, and quality inspection sampling. Update frequency is weekly or monthly. The materials contain both identifiers, model numbers, and proper nouns alongside natural-language clauses, so both retrieval paths apply. If chunking separates a quotation validity period, a supply scope, and the corresponding party, a customer-service citation will answer the wrong question, which can escalate complaints or cause customer-service training failure.
Constraints at the document chunking step
Chunking must preserve document identity and version that allow reference back to the original text. Conditions, descriptions, and the content they constrain cannot be split apart. Original data and retrieval indexes are two layers: one group of original data can correspond to multiple indexes, and one group of data can correspond to multiple vectors. During retrieval, vectors are recalled first, then the vector ID is used to find original data in MongoDB dataset.datas. Original data corresponding to the same group is merged, and the vector score takes the highest score. Vector retrieval is carried by the vector field in PostgresSQL, with HNSW as the index, while MongoDB handles other data storage and access. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, hybrid retrieval merges the two paths through RRF, and ReRank produces a 0-1 score used for filtering. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold, not a sorting rule. Retrieval itself does not constitute authorization and does not perform entity isolation.
Hypothetical paper test case: the material source text reads "Supplier A, quotation number A-001, quotation valid until June 30, 2025, covering the filter category at the East China warehouse. Arrival batch description number B-007, signing party Supplier A, signing status pending review." The draft sentence to be judged reads "Supplier A has confirmed filter supply to the East China warehouse until June 30, 2025." This draft sentence treats the quotation validity period as a supply period and treats pending review as confirmed, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this test | Reason for selection and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | Materials contain both identifiers and natural-language clauses; compared with semantic retrieval alone |
| Minimum relevance | 0.35 | Hypothetical starting point; compared with 0.5 to observe candidate count change after filtering |
| Citation limit | 3000 tokens | Hypothetical starting point; compared with 1500 tokens for citation completeness |
| Question optimization | Enabled | Multi-turn follow-up and short-question scenarios; compared with disabled for retrieval latency |
| Result reranking | Enabled | Compared with disabled for ranking position when candidates are numerous |
| Post-update review | Manual check of training and index status | Upload success and retrieval updated are two separate states |
Three common errors
- Treating the quotation validity period as the supply period. Input the draft sentence and the material source text, compare the same party and the same field. The material states only a quotation validity period and does not state a supply period, so the judgment is unsupported.
- Treating pending review as confirmed. Input the signing status field, compare the two mutually exclusive values "pending review" and "confirmed." The material shows pending review, the draft sentence states confirmed, so the judgment is contradictory.
- Treating upload success as retrieval already updated. Input the upload record and the training and index status, compare whether both completed at the same time. When training or index evidence is missing, the judgment is pending verification.
How to verify the setup
- Search with identifier A-001, observe whether the returned content contains the Supplier A party and the quotation validity field, and compare word by word with the original text. If the returned content lacks the party or the validity period, chunking has separated the constraining condition.
- Search with the natural-language question "until when does the East China warehouse filter supply last," observe whether returned candidates contain the quotation original text, then compare with the draft sentence. The candidate contains only the quotation validity period while the draft sentence states a supply period, so the judgment is unsupported.
- Use the paper input [] to check downstream design, observe what the system returns on an empty result, and compare with the expected empty citation. Do not claim that a fictional-rule search necessarily returns []. For libraries across different vector models, check the identifier first and exclude those that do not match, record the actual observation, and do not perform cross-model retrieval to compare missed recalls.