What the data for this category looks like
Data sources for this category include the Chinese Government Procurement Network, local government procurement websites, the fiscal integration system, the purchaser’s budget system, and supplier credit records. Update frequency is daily. Routine documents cover procurement requirements, procurement plans, bidding documents, tender documents, bid evaluation reports, winning bid announcements, procurement contracts, response to queries, and more. Each document has dedicated business fields, with no universal shared fields across documents. Individual document sizes range from thousands to tens of thousands of characters. The data covers all stages of the full procurement process. Retrieval scope must be restricted by job roles, to ensure different employees can only access business documents they are authorized to view.
What constraints these characteristics impose on the "knowledge base retrieval and recall" link
Decentralized data sources with daily updates require the retrieval system to support multi-source incremental synchronization, to avoid data lag that affects the timeliness of procurement processes. Diverse document types with no unified fields require configuring dedicated recall matching rules for different document types such as procurement requirements and bid evaluation reports, to improve retrieval accuracy. Large individual document sizes require adapting long-text chunking strategies, to avoid truncating core business clauses such as procurement methods and scoring standards. Errors in the procurement process have serious consequences, so the relevance threshold for retrieval and recall results must strictly match the business scenario, to reduce procurement risks caused by invalid recall.
How to set the configuration
This adapts the configuration logic of FastGPT v4.17.0, with specific configuration items and recommendations as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the business relevance requirements for government procurement documents, to avoid recalling irrelevant non-procurement documents |
limit | Top 8 entries | Covers the number of core documents for a single procurement project, while controlling context length to stay within processing limits |
searchMode | vector | Adapts to the semantic matching needs of government procurement documents, prioritizes returning business documents semantically related to the query terms |
usingReRank | Enabled | Performs secondary sorting on recall results, corrects deviations from single vector matching, and ensures that core clauses such as bid evaluation methods and procurement requirements are ranked higher |
datasetSearchUsingExtensionQuery | Disabled | Avoids automatically completing irrelevant information, which meets the strict information source requirements in government procurement scenarios |
collectionIds | Configured by business group | Distinguishes dedicated datasets for purchasers, regulatory authorities, and procurement agencies, to ensure the retrieval scope only covers authorized documents for the corresponding role |
Three common errors
- Symptom: Irrelevant information automatically completed by AI appears even after
datasetSearchUsingExtensionQueryis disabled. Cause: The global question optimization switch was not also disabled, or the configuration was not synchronized to the deployment instance. - Symptom: Knowledge base index cannot be created after deploying the PG database in Docker, on an 8c16G no-GPU virtual machine host. Cause: Parameters were not adjusted to adapt to vector computing resources in a no-GPU environment, or the database connection timeout configuration did not match the indexing pressure from large document chunks.
- Symptom: No results are returned after calling the search plugin. Cause:
collectionIdsis configured incorrectly, the ID of the corresponding government procurement dataset was not specified, or thesimilaritythreshold was set too high, resulting in no matching results.
How to confirm the configuration is correct
- Manually enter query terms related to procurement requirements, check whether the recalled results only include government procurement documents with no irrelevant content.
- View the index construction logs, confirm that daily incremental synchronization tasks are executed normally with no data pull failure errors.
- Test whether query results only return uploaded government procurement document content with no automatically completed information, after disabling
datasetSearchUsingExtensionQuery. - Adjust the
similaritythreshold to 0.8, verify whether the relevance of the recalled results meets the business scenario requirements.
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 2026-09-28.