What the documents look like
Customer-service knowledge base maintenance covers FAQ entries, reply templates, product parameter tables and escalation tickets. The work is front-loaded question-and-answer preparation, and it does not approve refunds directly. Under the contract and transaction material comparison task, source material is refreshed on a weekly or monthly cadence, and one product may have several versions of a parameter table alongside historical reply templates. In FastGPT, the smallest search unit is the library. A collection only classifies data and does not affect search results. Citation tracing requires that an answer can point back to the exact source text, so material version, subject and field state must be checkable.
A knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. A zero-length array does not mean the workflow terminates with an error; downstream steps still run. Retrieval results pass through semantic retrieval, full-text retrieval, RRF fusion and ReRank reordering, then similarity filtering and citation-limit trimming. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. It is a filtering threshold and does not perform permission or entity isolation. The citation limit is measured in tokens, which is a different concept from a thinking budget.
A hypothetical paper exercise: a knowledge base holds one product parameter table whose source text reads "Model A-100, rated voltage 220V, quotation valid until 30 June 2025." A separate escalation ticket records "Customer asks whether A-100 can be used in a 380V environment, pending technical confirmation." A proposed draft sentence reads "A-100 supports 380V, and the quotation is valid long term." In the source, only 220V and the quotation validity date are stated. The 380V value is unsupported, and "valid long term" and "valid until 30 June 2025" are conflicting values in the same field. "Pending technical confirmation" is an unverified state; it is neither a denial nor an authorization.
Constraints at the citation tracing step
The material version, subject and field state must be checkable before a draft sentence is accepted. The minimum relevance value is a similarity threshold, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. A search node returning citations does not mean the downstream answer step, any business verification, any approval, any database write or any external synchronization has occurred; those require their own stated downstream implementation.
The same-field test matters here. A contradiction exists only when the same subject, the same period and the same field carry clearly mutually exclusive values. A value that is simply not stated, or that has only related-stage evidence, counts as unsupported or pending verification. It does not count as a denial or a conflict. A pending review does not show that an operating permission is absent. Signing, reaching cooperation and contract effectiveness are three separate facts. A quotation validity date does not state a supply period. Missing evidence stays missing, and no authorization or effectiveness clause may be added to obtain a definite answer.
The document flow chart lists optional paths, and these do not all execute on every request. Image retrieval requires a corresponding image input and model capability. Reordering requires a model and must be enabled by configuration. Knowledge bases with different vector models cannot be associated together. The check is to inspect the identifiers, exclude non-matching bases and record what was observed; it does not run cross-model retrieval to compare missed recalls.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases with the same vector model | The sample material must sit in one searchable library; libraries with different vector models cannot be associated together |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Model A-100 suits full-text or hybrid retrieval; natural-language questions can use semantic retrieval |
| Minimum relevance | A 0-1 value, effective only under semantic retrieval or result reordering | A hypothetical starting value of 0.5 for this exercise, chosen to observe whether low-relevance fragments are filtered, compared against a run without the threshold |
| Citation limit | Maximum tokens cited per search | A hypothetical starting value of 2000, chosen to control cited content length, compared against the source text length |
| Question optimization | Coreference removal and question expansion before retrieval | Used to complete the subject in a follow-up such as "what about 380V"; a short, clear question can leave it off |
| Citation output | Array format, length may be zero | A paper input of [] is used to check whether downstream still runs; no claim is made that a real search returns [] |
Three common errors
- Treating the minimum relevance value as a permission or entity-isolation control. Relevance is a similarity threshold and does not exclude unrelated businesses or unauthorized records. Check: input the sample source text and the proposed draft sentence, compare whether the citation array still contains other business material before and after filtering, and observe whether the relevance setting changes the visibility range of records.
- Treating "pending technical confirmation" as a denial or an authorization. The escalation ticket records only a pending state; it states neither that 380V is unsupported nor that it is approved. Check: input the ticket text and the draft sentence, compare the field values of "pending confirmation" against "unsupported" and "approved," and observe whether a pending state was rewritten as a definite conclusion.
- Treating the quotation validity date as a supply period or long-term validity. The source states only that the quotation is valid until 30 June 2025 and does not state a supply period. Check: input the parameter table text and the draft sentence, compare whether "validity" and "valid long term" are mutually exclusive values in the same field, and observe whether an unstated clause was added.
How to verify the setup
- Citation source check. Input the sample parameter table text and the draft sentence "A-100 supports 380V, and the quotation is valid long term." Compare the subject, object and field values in the draft against the source. The source states only 220V and the quotation validity date; the 380V value and long-term validity in the draft are unsupported. Condition met: the answer's citation array contains that parameter table text, and the answer contains no value absent from the source.
- Empty-result downstream check. Use a paper input of [] to simulate a citation array of length zero, and compare whether the downstream answer step still runs. Condition met: downstream does not terminate with an error because of the empty array, and no citation content is fabricated. This check does not claim that a real search returns [].
- Cross-model association check. Inspect the vector model identifier of each knowledge base to be associated and compare whether the identifiers match. Condition met: only knowledge bases with the same vector model are associated; non-matching ones are excluded and recorded. This check does not run cross-model retrieval to compare missed recalls and does not infer the runtime result of associating them.