What the documents look like
Central chain procurement and distribution supports headquarters purchasing and the delivery chain. The documents in this category are centralised procurement contracts, replenishment plans, delivery notes and store requisitions. Their subjects are suppliers, distribution centres and stores. Their fields include product name, specification, quantity, delivery date and validity range.
Drafting product and marketing copy requires retrieving quotable source text from these documents and judging whether a proposed sentence has material support. The generic brief gives daily and weekly update frequencies. Update frequency does not change the search mode or the citation limit in this task, so it is noted here only as a property of the source material.
Search treats the whole library as the smallest search unit. A collection is a classification device and does not affect search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Semantic search relies on vector distance, full-text search relies on keyword hits, and hybrid search merges the two result sets through RRF. The minimum relevance value takes effect only in semantic search or when result reranking is used. The citation limit is measured in tokens. Question optimisation performs coreference resolution and question expansion and adds one model call. Retrieval does not perform permission or entity isolation. A similarity score does not determine supplier identity or contract ownership, so recalled content still needs manual comparison against the source text.
Hypothetical paper exercise sample:
Source text one: "Supplier A signed the centralised procurement contract on 1 March 2025, contract number JG-2025-031, agreeing on product A with a specification of 500 grams, with a supply validity period ending on 31 December 2025."
Source text two: "The distribution centre received a store requisition on 10 April 2025 for 200 units of product A, required to arrive before 15 April."
Proposed sentence: "Product A is supplied by Supplier A on a long-term and stable basis, with a 500-gram specification, and can meet routine store requisition demand."
The subject needed for judgement is Supplier A, the object is product A, the state is contract signing and supply validity, and the validity range is 1 March 2025 to 31 December 2025. The phrase "long-term and stable supply" has no corresponding wording in the source text and is unsupported.
Constraints at the retrieval step
The minimum relevance value is a filter threshold, not a final ranking rule. It applies only in semantic search or when result reranking is used. It is not a permission filter and it does not exclude unrelated suppliers or unauthorised records. A relevance score is not a category, entity or permission filter, and retrieval is not authorisation. Whether a recalled contract belongs to the drafting subject still requires manual comparison.
The citation limit is expressed in tokens. It is not a top-k item count. Because chunk lengths differ across a mixed knowledge base, token-based control is used instead of top-k. When the cited content is truncated and key fields are missing, the limit is adjusted and the comparison is repeated.
RRF is the fusion process for hybrid search results. The documents do not list a separate RRF switch, so it is not written as a toggle. Question optimisation runs before retrieval, performs coreference resolution and question expansion, and adds one model call. It is compared against the original question to check whether the subject has been altered.
Cross-model association is not allowed. Knowledge bases that use different vector models cannot be associated together. Library identifiers are checked first and non-matching libraries are excluded. Image retrieval requires an image input and the corresponding model capability. Reranking requires a rerank model and must be enabled according to the setting. Optional paths in the overall flow diagram are not executed on every request.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One library containing centralised procurement contracts and delivery notes | The library uses the same vector model; check whether the library identifiers match |
| Search mode | Hybrid search | The proposed sentence contains a product name and a specification, so semantic and keyword coverage are both needed; compare against semantic search alone |
| Citation limit | 800 tokens | Hypothetical starting point for this exercise; compare against 400 and 1200 tokens to see whether the recalled content is complete |
| Minimum relevance | 0.35 | Hypothetical starting point for this exercise; label relevant and irrelevant samples first, then compare filtered results |
| Question optimisation | Enabled | The proposed sentence contains a reference; compare the original question and the optimised question to see whether the subject was altered |
| Result reranking | Enabled | Used for secondary ranking when candidates are numerous; compare ranking before and after reranking |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input: the proposed sentence and Supplier A's contract. Comparison: whether the recalled content includes material from other suppliers. Observation: the relevance score reflects semantic similarity only and does not exclude unrelated supplier records. Check: compare the supplier identifier in the recalled results against the subject of the proposed sentence.
- Describing the RRF fusion of hybrid search as a toggle switch. Input: search mode set to hybrid search. Comparison: whether the configuration interface contains an RRF switch. Observation: RRF is a fusion process and the documents do not list a separate switch. Check: compare the configuration item names against the documented parameter table.
- Treating an unstated supply duration as a confirmed fact. Input: the proposed sentence "long-term and stable supply". Comparison: whether the source text states "long-term" or "stable". Observation: the source text states only a validity period ending on 31 December 2025. Check: compare the proposed sentence field by field against the source text and record unsupported wording.
How to verify the setup
- Input: source text one, source text two and the proposed sentence from the hypothetical exercise sample. Comparison: the phrase "long-term and stable supply" in the proposed sentence against the contract signing date and validity period fields in the source text. Observation: the source text does not contain "long-term" or "stable", so the proposed sentence is unsupported. If the source text were supplemented with "long-term" wording consistent with the contract validity period, the support condition would be met.
- Input: the same proposed sentence, with search mode set to semantic search and then to hybrid search. Comparison: whether source text one appears in the recalled results in each case. Observation: hybrid search uses semantic and full-text recall together and may recall different results. If neither mode recalls source text one, record it as a miss and do not conclude that the material does not exist.
- Input: citation limit set to 800 tokens, using source text one and source text two. Comparison: whether the recalled content completely includes the contract number, specification and requisition quantity. Observation: if the cited content is truncated and key fields are missing, adjust the limit and compare again. For the empty-result case, a paper input of [] can be used to check the downstream design, without claiming that retrieval necessarily returns [].
Knowledge bases using different vector models cannot be associated together. Check the library identifiers first and exclude non-matching ones. Image retrieval requires an image input and model capability. Reranking requires a model and must be enabled according to the setting. Retrieval does not perform authorisation checks. Business approval and database writes require a separate system implementation.