What data in this category looks like
Data for this category originates from state-owned assets supervision information systems, enterprise asset management systems, property rights exchange systems, evaluation agency reports and financial statements. The update frequency is monthly. Daily documents include property rights registration forms, asset evaluation reports, property rights transfer plans, on-site transaction announcements and asset disposal approval forms. Each document contains structured fields such as ownership subject, asset number, evaluation value, transaction status and approval process node. Some reports include scanned attachments or structured tables. Most fields cover ownership, value measurement and transaction process information related to state-owned assets.
What constraints do these characteristics impose on the knowledge base retrieval and recall workflow
Monthly updated data requires regular full index rebuilds to prevent expired data from appearing in search results, which would affect the timeliness of ledger summary. The mixed nature of structured and semi-structured documents requires retrieval logic to adapt to chunking rules for different document types, to avoid losing key ownership information due to long text truncation. There is a high demand for exact matching of fields such as asset number and evaluation value, which imposes clear constraints on the retrieval similarity threshold. Low-relevance non-asset content must be filtered out. Transaction documents have strong timeliness requirements. Updated content from the past 30 days must be prioritized for recall, to avoid compliance risks caused by calling outdated transaction announcements or approval forms.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Exact matching of key fields such as asset ownership and number is required to filter out low-relevance non-asset content and ensure the accuracy of ledger summary |
limit | Top 8–12 entries | Ledger summary needs to cover associated information for multiple assets. Too many entries will exceed the context length limit, while too few will miss key transaction records |
searchMode | Vector retrieval + keyword hybrid | Can match both structured fields and natural language descriptions, adapting to multiple document formats such as property rights registration forms and asset evaluation reports |
usingReRank | Enabled | Key information for asset documents is scattered across different paragraphs. Reranking can improve the ranking priority of highly relevant content, avoiding core ownership information being pushed down |
datasetSearchUsingExtensionQuery | Enabled | Queries for ledger summary often include batch requests for multiple assets. Optimization expands the retrieval scope to cover associated documents such as transfer plans and transaction announcements |
collectionIds | Unique dataset ID for this category | Isolates property rights and asset data from other business data, avoiding cross-category retrieval interference and ensuring search results only cover the target business scope |
Three Common Misconfigurations
- Symptom: After multiple consecutive knowledge base retrieval calls, the matching accuracy of returned results continues to decline, and normalcy is restored after restarting the conversation. Cause: The context cache of multi-turn conversations is not cleared. Residual old retrieval vectors cause vector space drift for subsequent queries.
- Symptom: After replacing the embedding model, the imported knowledge base cannot recall correct content, returning empty results or low-match content. Cause: Document vector indexes were not regenerated for the new model. The old index still uses the vector space of the original model, which cannot adapt to the vector dimensions and spatial distribution of the new model.
- Symptom: The number of asset entries returned by a single retrieval is far lower than expected, failing to cover all associated data required for ledger summary. Cause: The
limitparameter was not adjusted to a value suitable for batch summary, or the hybrid retrieval mode was not enabled, leading to overly strict keyword filtering.
How to Verify Proper Configuration
- Import a single property rights registration form, run a keyword search, and check whether the returned results include key fields such as asset number and ownership subject, to confirm that the retrieval mode and similarity threshold settings are effective.
- Initiate multiple consecutive queries, check that the matching accuracy of returned results for each round does not decrease significantly, to confirm that no abnormal drift occurs in the context cache or vector space.
- After replacing the embedding model, run a batch document index rebuild task, and verify the vector generation speed of the new model and the recall consistency of existing data.
- Compare the returned results of vector retrieval and hybrid retrieval, to confirm that the ranking priority of highly relevant documents meets expectations after
usingReRankis enabled.
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.