What the data for this category looks like
Data for this category mainly comes from official policy documents, industry standards, local infrastructure approval guidelines released by the Ministry of Housing and Urban-Rural Development, the Ministry of Transport, provincial housing and urban-rural development departments, and compliance requirement documents from project locations. Updates are triggered by the release of new policies, with no fixed cycle. Most documents follow the official red-headed document format, and include fields such as policy document number, issuing authority, effective date, scope of application, and core clauses. These fields mainly consist of text numbers, dates, geographic scopes, and clause content, with no complex numerical units.
What constraints these characteristics impose on knowledge base retrieval and recall
Policy documents are long texts with hierarchical clause numbering. Retrieval must match both keywords and clause identifiers to avoid vague recall of irrelevant content. Data sources are scattered across national, provincial, and municipal official platforms, with minor format differences. Normalization of document numbers and effective dates is required before importing into the knowledge base. Policies have clear effective and expiration dates. Retrieval must include time filtering rules to only recall currently valid policy clauses. Policies also include geographic scope fields, so matching policies must be filtered based on project location to reduce invalid recall results.
How to set configurations
| Configuration Item | Suggested Value | Basis for This Value |
|---|---|---|
similarity | 0.75–0.85 | Infrastructure policies are rigorous written documents. Low-match irrelevant content must be filtered to avoid recalling non-corresponding clauses |
limit | Top 8–12 results | The number of valid clauses in a single infrastructure policy is concentrated. Excessive recall will exceed the context window limit |
searchMode | Precise matching mode | Policy retrieval requires strict matching of clause content and keywords to avoid generalized recall of documents from unrelated fields |
usingReRank | Enabled | Local key clauses in long-text policies are easily overlooked by global matching. Reranking improves the ranking weight of core clauses |
datasetSearchUsingExtensionQuery | Disabled | Policy expressions are rigorous. Extended queries easily introduce irrelevant keywords, leading to recall bias |
collectionIds | Specify the dedicated dataset ID for infrastructure policies | Isolate infrastructure policy data from knowledge base data of other categories to avoid cross-category recall interference |
Three common configuration mistakes
- Single retrieval returns empty results. The cause is failure to specify the dedicated infrastructure policy
collectionIds, so the retrieval scope covers other datasets with no matching content. - Retrieval results include expired policy clauses. The cause is failure to set filtering rules based on policy effective date fields, resulting in recall of expired document content.
- Retrieval results include non-infrastructure policy content. The cause is enabling
datasetSearchUsingExtensionQuery. Extended queries introduce cross-category keywords that interfere with matching accuracy.
How to confirm configuration is properly applied
- Access the dataset management interface of FastGPT v4.17.0, confirm the dataset corresponding to the dedicated infrastructure policy
collectionIdsis selected. - Enter a known infrastructure policy keyword, check if returned results include the original text of the corresponding clause.
- Verify the effective date field of retrieval results, confirm only currently valid policy content is recalled.
- Disable
datasetSearchUsingExtensionQueryand run retrieval again, check if the matching accuracy of results meets expectations.
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.