What the data for this category looks like
Daily documents for this category include transfer announcements, transaction notices, bid application documents, margin vouchers, transaction confirmation letters, listing result announcements, property right transfer specifications, and similar materials. Data sources are the national public resource trading platform, natural resource trading systems, property right exchange systems, and state-owned asset supervision systems. The update frequency is daily. Common data fields include transaction subject ownership information, listing price, bid requirements, transaction confirmation time, and similar items. Units involve area, amount, date, and similar categories. Most document structures combine structured forms and announcement text, with certain format differences across sources.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Daily updated transaction documents require retrieval to support high-frequency incremental synchronization, to avoid data lag affecting ledger timeliness. The accuracy of core fields such as ownership and listing price directly determines ledger compliance, so retrieved results must accurately match subject attributes. Heterogeneous data from multiple systems needs unified vector space alignment to avoid result deviations in cross-system retrieval. Ledger summary requires associating multiple related documents, such as bid applications and transaction confirmation letters, so retrieval needs to support configuration for multi-document associated recall, to reduce manual verification time.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precise retrieval requirements for core fields such as ownership and price, avoiding low-similarity irrelevant transaction documents from being included in results |
limit | Top 10 entries | Covers the number of related documents required for a single ledger summary, while controlling retrieval time |
searchMode | vector | For transaction documents combining structured content and text, vector retrieval can accurately match subject attributes, adapting to retrieval needs for various types of transaction documents |
usingReRank | Enabled | Re-ranks retrieved results by relevance, improving the ranking priority of core fields such as ownership and transaction information |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms and related expressions for user queries, covering wording differences across documents such as bid notices and property right transfer specifications |
collectionIds | Bind corresponding knowledge base collections by transaction type | Accurately limits the retrieval scope, avoiding interference from cross-category documents, and adapting to classified retrieval needs for land, mining rights, and state-owned property rights |
Three common mistakes
- Phenomenon: The knowledge base retrieval returns empty results, and no preset fallback prompt is generated. Cause: The trigger logic for empty results is not configured, and
searchModeis not adjusted to adapt to empty recall scenarios. - Phenomenon: Some conversations do not trigger knowledge base retrieval, and responses only rely on large model generation. Cause: The knowledge base retrieval node is not set as a mandatory execution step in the workflow, or the global forced retrieval configuration is not enabled.
- Phenomenon: After multiple consecutive rounds of questioning, knowledge base retrieval time increases significantly, with a single retrieval duration exceeding 90 seconds. Cause: The conversation context accumulates a large amount of historical vector data, and no context clipping rules are configured, causing retrieval calculation volume to rise with conversation rounds.
How to confirm the configuration is complete
- Execute a single subject ownership query, check whether the retrieval results include documents such as transfer announcements and transaction confirmation letters for the corresponding transaction subject.
- Trigger an empty query scenario, verify whether a preset empty result prompt or fallback recall rule is returned.
- Simulate multiple consecutive rounds of questioning, check whether retrieval time remains stable without significant increase.
- Switch queries for different transaction types, verify that only knowledge base collections of the corresponding category are retrieved.
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.