What the documents look like
Knowledge-base maintenance for customer service supports the preparation of question-and-answer material that sits in front of the service conversation. It does not approve refunds or returns. The material comes from the customer-service system, product detail pages, after-sales question statistics and quality spot checks. Day-to-day documents include FAQ entries, reply templates, product parameter tables and escalation tickets. The retrieval object for product data and selection questions is material such as product parameters, model numbers and applicable scenarios, where a question needs to match a precise item. Customer-service supervisors, knowledge-base operators and product operators maintain these documents together. Errors surface as off-topic answers, escalated complaints and failed agent training. Updates follow a weekly or monthly rhythm, and the retrieval step responds to whatever is currently stored.
Constraints at the retrieval step
A knowledge base is made of libraries, collections and data. The smallest search unit is the library, so a search runs against an entire library; collections only classify data and do not change search results. Vector search uses the PostgresSQL PG Vector extension with an HNSW index. The MongoDB dataset.datas table stores the original vector data together with an indexes field, and one data group can map to several vectors. A knowledge-base search node can associate one or more knowledge bases that use the same vector model; libraries built on different vector models cannot be associated together. Semantic search relies on vector distance. Full-text search suits numbers, model names, proper nouns and error codes. Hybrid search merges the two paths through RRF. Result reranking uses a ReRank model for a second ordering. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. Question optimization performs coreference removal and question expansion before retrieval, which adds one model call and increases overall latency. Retrieval and similarity scores do not perform permission or entity isolation, and a retrieval result does not represent authorization.
Hypothetical paper exercise sample: one knowledge base holds a product record whose source text reads, "The Type A water-purifier filter fits B-series models produced after March 2024, with a replacement interval of 12 months." A draft sentence awaiting judgment reads, "The Type A filter fits all B-series models." The same record also carries a handover request: "Please synchronize this filter record to the after-sales ticket system." That request is pending review and has no signature or execution authorization.
Configuration and implementation checks
| Configuration or check item | Setting or action on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases that use the same vector model; verify the vector model identifier first | The library holding the sample record and any library to be associated must share one vector model |
| Search mode | Test model-number questions with full-text search or hybrid search | The sample contains model words such as "Type A" and "B-series" |
| Citation limit | For this exercise, start at 800 tokens because the record is short and the fit-range sentence must survive; compare against 600 and 1000 to observe content completeness | Used to observe whether the fit range is cut off |
| Minimum relevance | Set only under semantic search or result reranking; label relevant and irrelevant samples before adjusting | The draft sentence is close in meaning to the source text, so manual labelling is required |
| Question optimization | Under multi-turn follow-ups, compare whether the subject of the original question and the optimized question has been altered | The sample is a single clear question, so on and off can be compared |
| Empty-result handling | Check the downstream chain design against a paper input of an empty array | Used to confirm the chain still runs when nothing is matched |
Three common errors
- Treating the draft sentence "The Type A filter fits all B-series models" as a conclusion supported by the source text. The source text limits the fit to B-series models produced after March 2024, and the draft sentence widens the fit range, so it is an unsupported rewrite. Check: input the source text and the draft sentence, compare the production-time range field for fitted models, and mark the draft as unsupported if it omits the "produced after March 2024" limit.
- Treating the pending handover request as already executed. The request is pending review with no signature or execution authorization. Check: input the request, compare its status field, and if the status is pending review, record it as not executed without writing any synchronized conclusion.
- Treating a similarity score as a permission or entity filter. A retrieval score reflects content relevance only and does not exclude unauthorized records. Check: input a query that includes an unrelated business, compare the returned content against the authorized-scope list, and if the returned content falls outside that scope, record that retrieval performed no permission isolation and that a separate system implementation is required.
How to verify the setup
- Input the sample source text and the draft sentence, then compare the production-time range field for fitted models. If the draft sentence keeps the "produced after March 2024" limit and matches the source text, the condition is met; if the limit is missing, record it as unsupported and keep the evidence gap open.
- Input the pending handover request and compare its status field against the authorization record. If the status remains pending review with no signature record, record it as not executed; if a signature record appears, verify the signing party and scope, and do not infer authorization from a pending state.
- Input an empty array and observe whether the downstream chain runs. If the output length is 0 and the chain continues, the empty-result handling condition is met; if the chain stops, record the actual observation without asserting that a retrieval rule must return an empty array.
These checks cover the named task of retrieval and recall for product data and selection questions. They do not test identity certification, report analysis or any other task, and no test, deployment, approval or synchronization is claimed to have occurred.