What the documents look like
Brand dealer ordering events run on a quarterly or annual cycle. The working materials come from the ordering event system, historical sales, dealer inventory, and financial credit approval. Day-to-day documents include the ordering handbook, ordering contracts, sample lists, and credit approval records. These materials update on the same rhythm as the ordering event itself, and retrieval demand concentrates in two scenarios: checking operating reports and reviewing anomalies.
The ordering handbook and sample lists carry precise identifiers such as model numbers and codes, so full-text retrieval or hybrid retrieval suits queries against them. Ordering contracts and credit approvals carry fields for subject, amount, and validity scope; semantic retrieval can match close paraphrases, but a similarity score reflects textual relatedness only. It does not determine dealer identity, credit authority, or contract status. A 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. Retrieved results output citations as an array, and the length may be 0, meaning the downstream chain still runs when nothing is found.
Hypothetical paper exercise sample: a brand channel department prepares a set of ordering event materials for cross-checking. The ordering handbook states: "For the A-type products listed in this ordering event's sample list, dealers may place reservations within their credit limit; the reservation deadline is March 20." A credit approval record states: "East China dealer A, credit limit 800,000 yuan, valid until June 30, status approved." A draft sentence reads: "Dealer A has completed the reservation for A-type products within the credit limit." The cross-check target: whether the draft sentence's claim of a completed reservation has supporting records.
Constraints at the retrieval step
The similarity score from semantic retrieval ranks textual closeness and does not act as a permission or entity filter. A high-scoring fragment does not confirm that dealer A holds the credit scope, that a reservation occurred, or that a contract is effective. Retrieval is not authorization. The credit approval record shows an approval status; an approval status is not the same fact as an executed reservation, and a pending request is not an approval.
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Before association, check the vector model identifier of each base and exclude any base whose identifier does not match. Do not run a cross-model retrieval to compare missed recalls; record only the identifier check and the exclusion.
The citation output is an array whose length may be 0, and the chain still runs on an empty result. A paper input of an empty array can be used to check how a downstream node handles a length-0 citation output; this does not claim that a fictional rule always returns an empty array.
The documented overall retrieval flow lists optional paths, and not every request executes all of them. Image retrieval requires an actual image input and the corresponding model capability. Rerank requires a rerank model and must be enabled according to settings. RRF is the fusion step for combining multiple retrieval channels; it is not an on/off switch. The minimum relevance threshold takes effect only under semantic retrieval or when result reranking is used.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Ordering handbook base and credit approval base | Both bases use the same vector model; check the vector model identifier before associating |
| Search mode | Hybrid retrieval | Model numbers and codes need full-text hits; policy wording needs semantic recall; compare recall differences between the two modes on the sample question |
| Citation limit | Hypothetical starting value 2000 tokens | The ordering handbook has long passages; test content completeness with actual materials of different lengths, then adjust |
| Minimum relevance | Hypothetical starting value 0.5 | Takes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples first, then tune the threshold |
| Question optimization | Enabled | Used for coreference resolution and question expansion; compare whether the subject is altered between the original and optimized question |
| Result reranking | Enabled per settings | The rerank model reorders candidate text and returns a 0-1 score usable for filtering; compare ordering before and after enabling |
Three common errors
- Treating a retrieval hit as authorization confirmation. Input: the "approved" status in the credit approval record. Comparison object: the "completed reservation" wording in the draft sentence. Observable result: the approval record does not state a reservation action, so the sentence is unsupported.
- Treating a similarity score as a permission filter. Input: a high-scoring fragment returned by semantic retrieval. Comparison object: dealer A's credit scope. Observable result: the score does not determine subject authority; the credit record requires manual cross-checking.
- Forcing association of bases with different vector models. Input: the vector model identifiers of the two knowledge bases. Comparison object: the node's association requirement. Observable result: when identifiers differ, exclude that base and record the actual observation without inferring runtime behavior.
How to verify the setup
- Input the ordering handbook text and the draft sentence from the hypothetical sample. Compare the draft sentence's "completed reservation" against the handbook's "may place reservations within the credit limit." Observable result: the handbook states only that a reservation is permitted, the draft sentence exceeds the handbook's scope, and the sentence is judged unsupported.
- Input the credit approval record text. Compare the "approved" status against the "completed reservation" action. Observable result: an approval status does not demonstrate that a reservation was executed; preserve the evidence gap.
- Input an empty citation array. Compare how the downstream node handles a length-0 output. Observable result: the chain runs and produces no fabricated citation content; record the actual observation.