What the documents look like
Customer-service knowledge base maintenance covers FAQ entries, script templates, product parameter tables and escalation tickets. In the product and marketing copy drafting task, retrieval and recall determine whether a proposed sentence can find material that supports it. The brief describes weekly and monthly update rhythms, with product parameter tables changing alongside new items and FAQ entries and script templates revised after quality sampling. Retrieval does not approve refunds and does not publish copy externally. It only delivers source fragments relevant to a draft sentence into later judgment.
The knowledge base is organised into library, collection and data. The smallest search unit is the library; collections serve classification management and do not affect search results. One group of data can map to multiple vectors, and a hit on any one vector recalls that group of source data. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together.
Constraints at the retrieval step
During retrieval, vectors are recalled first, then the vector ID is used to locate source data in the MongoDB dataset.datas table. Records corresponding to the same group of source data are merged, and the vector score takes the highest value. Semantic search, full-text search and hybrid search suit different material. Numbers, models and proper nouns suit full-text or hybrid search; natural-language questions suit semantic or hybrid search. Hybrid search merges multiple result paths through RRF, then applies ReRank for a second ordering when configured.
The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. Question optimisation performs coreference resolution and question expansion and adds one model call. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled by configuration. Optional paths do not mean every request executes all of them. The citation limit is measured in tokens. The output citation content is an array and its length can be 0, meaning the downstream link still runs even when nothing is retrieved.
Configuration and implementation checks
The settings below centre on this page's sample. The numeric values are hypothetical starting points for this trial, used to compare against adjacent values. They are not official defaults.
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the base containing the product parameter table and FAQ | First check each base's vector model identifier and exclude those whose identifiers do not match |
| Search mode | Hybrid search | Compare recall of the draft sentence against semantic search alone and full-text search alone |
| Citation limit | 600 tokens | Compare against 300 and 1200 to observe whether the parameter table fragment stays complete |
| Minimum relevance | 0.35 | Compare against 0.2 and 0.5 to observe whether unrelated fragments are filtered |
| Question optimisation | Enabled | Compare against disabled to check whether the subject is altered incorrectly |
| Result reranking | Enabled | Compare candidate ordering against disabled; the rerank score is used for filtering |
Hypothetical exercise sample. Source material for checking: "The inner liner of the A-type insulated cup is made of 304 stainless steel. This model has a capacity of 500 millilitres. This batch of products has followed the above parameters since 1 March 2025 and applies to the online retail channel." A draft sentence to be judged: "The A-type insulated cup has a 316 stainless steel inner liner, a capacity of 500 millilitres, and suits all sales channels." The material states only 304 stainless steel and the online retail channel. The draft sentence's 316 stainless steel and all sales channels are unsupported. The 500 millilitre capacity matches the material.
Three common errors
- Treating the minimum relevance value as a permission or entity filter. Observable error: assuming that raising this value excludes other business lines or unauthorised records. The input to check is the draft sentence and the parameter table source text in the hypothetical sample. Compare whether the returned citation content is determined only by similarity. Observe whether the result contains the 316 stainless steel wording that the material does not state.
- Associating knowledge bases that use different vector models. Observable error: associating a base with a non-matching identifier and then inferring missed recall or redundancy. The input to check is each base's vector model identifier. Compare whether the identifiers match. Observe whether non-matching bases are excluded, and record the actual observation without inventing a runtime result.
- Treating a pending review or a quotation validity period as an effective fact. Observable error: concluding that copy is usable from "submitted for review", or concluding a supply period from a quotation validity period. The input to check is the status field in the source text. Compare the value of the same subject, the same period and the same field. Observe whether a clearly mutually exclusive value exists. When only stage-related evidence is present, record it as pending verification.
How to verify the setup
- For the hypothetical sample draft sentence "The A-type insulated cup has a 316 stainless steel inner liner, a capacity of 500 millilitres, and suits all sales channels", enter that sentence to run retrieval. Compare the material, capacity and channel fields in the recalled fragments. Observe whether the recalled content contains the source text's 304 stainless steel and online retail channel. If the recalled fragments do not cover 316 stainless steel, record it as unsupported rather than as a negation.
- For the citation limit, enter the same draft sentence. Compare whether the parameter table fragment stays complete at 600 tokens against 300 and 1200 tokens. Observe whether the capacity and material fields are truncated. The expectation is that the three fields appear together at 600 tokens. This expectation has not been measured.
- For empty-result handling, use an explicit paper input of an empty array to check the downstream link design. Compare whether the downstream continues to execute on citation content whose length can be 0. Observe whether the link runs through. Do not claim that a fabricated rule retrieval necessarily returns an empty array.
When knowledge bases with different vector models cannot be associated, first check the identifiers and exclude the non-matching ones. Do not run cross-model retrieval to compare missed recall. Retrieval checks record only observations and failure branches. They do not guarantee a first-hit result, full recall, or an empty result for an unrelated query.