What the data for this category looks like
Daily documents for this category include property rights registration forms, asset evaluation reports, property rights transfer plans, on-site transaction announcements, and asset disposal approval slips. Data sources include state-owned assets supervision information systems, enterprise asset management systems, property rights exchange systems, evaluation agency reports, and financial statements. The data includes structured fields and unstructured text, such as asset ownership numbers, evaluation base dates, and transaction approval statuses. The data updates monthly. Service subjects include state-owned assets supervision institution property rights departments, central enterprise asset management departments, property rights trading institutions, and evaluation agencies.
Constraints for Knowledge Base Retrieval and Recall
This category’s data includes tens of thousands of rows of structured Excel tables and long document reports. The monthly update cycle requires the retrieval system to support incremental data synchronization. Multi-source, multi-format data needs unified preprocessing logic to ensure consistent retrieval results. Party building learning Q&A must cover both state-owned assets business terms and party building keywords. Retrieval must cover both business fields and topic-related content. The total data volume is large. It is necessary to strictly control the recall range and result length to avoid token overrun and reduced retrieval efficiency.
How to Set Configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Business documents contain state-owned assets professional terms and structured fields. A threshold that is too low will introduce irrelevant results. A threshold that is too high will miss valid business and party building related content. Calibrate based on actual testing. |
limit | Top 3–5 entries | Single asset or party building content has high information density. Too many recalls will exceed token limits and affect retrieval response speed. |
searchMode | Hybrid vector retrieval + keyword retrieval | Data contains structured fields and free text. Hybrid mode balances field matching and semantic relevance, and adapts to the keyword requirements of party building learning Q&A. |
usingReRank | Enabled | Long documents and multi-source data are prone to semantic deviation. Reranking improves result relevance and meets the accuracy requirements of party building learning Q&A. |
datasetSearchUsingExtensionQuery | Enabled | Can add extension words related to party building and state-owned assets business to retrieval, improving cross-topic associated recall effects. |
collectionIds | Bind property rights asset exclusive dataset and party building learning dataset | Need to isolate irrelevant business data to ensure the retrieval scope only covers documents of the current category and party building learning content. |
Three Common Misconfigurations
- Phenomenon: Single retrieved result occupies excessive tokens, or recalled content fails to match detailed property rights or party building questions. Cause: Did not adjust chunking parameters based on category document length, used general chunking configurations, and did not adapt to the structure of long documents such as property rights transfer plans and asset evaluation reports.
- Phenomenon: When the knowledge base recall is empty, the preset online search cannot be triggered, or the search is triggered incorrectly. Cause: Did not correctly configure the trigger logic for empty retrieval, and did not distinguish between non-business chat questions and business questions with no matching results.
- Phenomenon: Chat questions are incorrectly recalled as valid results. Cause: Did not set a reasonable
similaritythreshold, or did not add exclusive filtering rules, resulting in low-relevance chat content being included in retrieval results.
How to Confirm Configuration is Correct
- Upload a single property rights registration form and party building learning material, perform retrieval, check whether the number of recalled results matches the
limitconfiguration, and whether the relevance meets thesimilaritythreshold requirements. - Construct a chat question with no matching content, check whether the preset filtering logic or online search is triggered (if configured).
- Import 10,000 rows of Excel asset data, perform batch retrieval testing, check whether the system response time and token consumption meet expectations.
- View the FastGPT v4.17.0 retrieval log to confirm whether
usingReRankanddatasetSearchUsingExtensionQueryare working properly.
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.