What the Data for This Category Looks Like
Daily documents for this category include purchase applications, price comparison records, contracts, and acceptance sheets. Most data comes from procurement systems, supplier quotation documents, budget sheets, and contract systems. Data updates follow cycles of individual orders or monthly periods. A single document typically contains fields such as purchasing entity, category specifications, quotation details, contract clauses, and acceptance nodes. There is no unified fixed unit, with adjustments made based on the procurement category and internal rules. Internal work orders are mostly associated with specific procurement projects, and require matching the historical document content of the corresponding order.
Constraints Imposed by These Characteristics on the Knowledge Base Retrieval and Recall Link
Procurement documents cover wide content ranges, spanning multiple business links from application to acceptance. After being split into chunks, the chunks have strong contextual relevance, so irrelevant chunks across different links must be avoided during recall. Data updates are performed per individual order, with uneven incremental update frequencies. Adapt dynamic incremental recall logic to avoid redundant overhead from full retrieval. Different document fields have clear differences: purchase applications focus on specifications and budgets, while contracts focus on clauses and responsibilities and rights. Match semantic weights for corresponding fields during retrieval to avoid retrieval bias across different document types. Internal work orders mostly point to specific purchase orders. Accurately recall historical documents for the corresponding order to avoid interference from irrelevant suppliers or content from other projects in Q&A results.
Configuration Setup
This configuration is based on FastGPT v4.17.0, and is tailored to the business characteristics of IT procurement and supplier management.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Procurement documents have significant semantic differences. This threshold filters low-relevance chunks and avoids confusion between contract clauses and purchase applications during retrieval |
limit | Top 8 entries | The number of historical documents associated with a single purchase order is moderate. This setting retains sufficient context while controlling recall redundancy |
searchMode | Semantic Retrieval | Adapts to semantic association of unstructured text such as procurement requirements and contract clauses, covering multiple types of business documents |
usingReRank | Enabled | Secondary filtering of semantic retrieval results to accurately recall content related to specific procurement links associated with work orders |
datasetSearchUsingExtensionQuery | Enabled | Supplements implicit procurement project details in work orders, expanding the effective recall range without reducing relevance |
collectionIds | Specify the collection ID of the internal IT procurement knowledge base | Limits the retrieval scope to only procurement-related documents, avoiding interference from other category data on retrieval results |
Three Common Misconfigurations
- Enabling
usingReRankresults in empty retrieval results. The interface displays "No matching chunks", and the backend logs show no valid recall entries. The cause is that thesimilaritythreshold is set too high. After combining with re-ranking filtering, all initial semantic retrieval results are eliminated. - The locally deployed knowledge base cannot upload new documents. When uploading purchase contracts or acceptance sheets, a prompt "Knowledge base count limit exceeded" appears. The cause is that the default knowledge base count limit parameter for the local instance has not been adjusted, and it does not adapt to the number of documents uploaded in a single batch for this category.
- Retrieval results mix non-procurement-related documents. Irrelevant business content appears in work order Q&A. The cause is that the retrieval scope is not limited via
collectionIds, and knowledge base content from other categories is mistakenly included in the recall pool.
How to Verify Proper Configuration
- Upload a single purchase application document, initiate a test work order containing keywords for this procurement category, and verify that the retrieved chunks only include relevant content from this document.
- After enabling
usingReRank, compare the number of retrieval results before and after enabling to confirm that low-relevance redundant chunks are filtered out after re-ranking. - Adjust the
similaritythreshold to 0.75, initiate multiple rounds of testing, and confirm that the relevance of returned results meets business expectations. - Check the
collectionIdsconfiguration to confirm that the retrieval scope only covers the internal IT procurement knowledge base, with no cross-category data mixed in.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.