What the data for this category looks like
Daily documents for this category include transfer announcements, transaction notices, bid application materials, margin vouchers, transaction confirmation letters, listing result announcements, property transfer manuals, and similar items. Data mainly comes from national public resource trading platforms, natural resource trading systems, property exchange systems, and state-owned asset supervision systems. The update frequency is daily. Common data fields include transaction subject ownership information, restriction conditions, margin amount, bidding rules, transaction confirmation clauses, and similar details. Most document structures combine structured forms and announcement texts. Some data must be obtained via cross-system synchronization.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Daily updated data synced across multiple systems requires the retrieval and recall link to support incremental synchronization and real-time update verification. This prevents expired listing or transaction information from being recalled. The mixed structure of structured forms and unstructured announcement texts requires retrieval to adapt to both keyword matching and semantic recall. This accurately locates core fields such as ownership and restriction conditions. Information tied to transaction subject ownership, margin, and other details directly relates to business compliance. This requires recall results to include clear source identifiers and data versions. This ensures traceability during verification processes. Cross-system data format differences also require unified field mapping rules. This avoids field mismatch issues during retrieval.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Core fields for this category (such as ownership and restriction conditions) require high matching precision to avoid recalling irrelevant transaction announcements. The 0.75–0.85 range balances recall volume and accuracy, and can be fine-tuned based on actual testing. |
limit | Top 10 results | Transaction-related documents have large individual information volumes. Too many recall results will exceed the context window. The top 10 results cover associated announcements and vouchers for core subjects, meeting verification needs for internal office work. |
searchMode | hybrid | Combines semantic recall and keyword matching to adapt to mixed document types of structured forms and unstructured announcements. This accurately targets key information such as ownership and margin amount. |
usingReRank | Enabled | Reranks recall results to prioritize content strongly related to transaction subject ownership and bidding rules. This resolves result ranking bias that may occur with single semantic recall. |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms to add category-specific terminology such as "ownership", "margin", and "bid application". This improves recall effectiveness for specialized fields. |
collectionIds | Configured by business domain | Separates dedicated knowledge base collections for land use rights, mining rights, and state-owned property transactions. This prevents cross-category data from being mixed in, ensuring retrieval results accurately correspond to business scenarios. |
Three common configuration errors
- Reference fields in retrieval results are empty or show incorrect IDs. This occurs when the
collectionIdsparameter is not configured correctly, so the system cannot associate the file collection ID of the corresponding knowledge base. Or when the reference template does not correctly bind the queried single data ID. - XLSX-format transaction notices and bid application documents cannot recognize question-and-answer pairs. This happens when question-and-answer pairs in the document are not formatted according to FastGPT-supported standards, such as separate lines for questions and answers and use of fixed separators. Or when cells contain merged formatting or hidden content, leading to parsing failure.
- A single knowledge base retrieval takes 6 to 7 seconds. This occurs when
usingReRankis not enabled and the recalllimitis set too high. Or when the knowledge base contains a large number of uncompressed large-volume documents, causing excessive latency in vector retrieval and reranking steps.
How to confirm the configuration is complete
- Initiate a query that includes specialized terms such as "land use right ownership" and "mining right transfer conditions". Check whether retrieval results include documents from this category, and no cross-category procurement data is mixed in.
- View the reference identifiers of retrieval results. Confirm that each result includes a unique data ID, and can be associated with the source system and update time of the corresponding document.
- Test uploading an XLSX-format transaction notice document. Check whether the question-and-answer content in it can be parsed normally, with no parsing failure prompts.
- Count the time taken for a single retrieval. Confirm that it falls within a reasonable range. If latency is too high, adjust the
limitparameter or compress large-volume documents.
The values listed here are common starting points. They should be evaluated against the relevant samples of the target deployment environment.
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.