What this category of data looks like
This category of data comes from enterprise contract management systems, compliance risk databases, law and regulation libraries, case management systems, and reporting platforms. The data updates monthly. Daily documents include compliance risk lists, contract legal review opinions, major issue legal opinions, compliance review forms, and violation report investigation results. Most documents have three core fields: issue-related basis, review opinion, and processing result. There are no unified unit-type fields. Document formats vary across different sources.
What constraints these characteristics impose on knowledge base retrieval and recall
The monthly update frequency requires the knowledge base synchronization cycle to match the update rhythm. This avoids frequent synchronization consuming system resources. Multi-source, multi-format document characteristics require the retrieval system to adapt to different document field structures. This prevents recall bias caused by format differences. The strict nature of compliance content requires strict control over retrieval result relevance. This stops low-relevance content from interfering with compliance judgments. Multi-system data sources require limiting the retrieval scope to compliance-specific datasets. This prevents unrelated data from mixing into retrieval results.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the strict requirements of legal compliance documents, balances recall coverage and result relevance |
limit | Top 10 results | Adapts to the volume of monthly updated compliance documents, balances retrieval efficiency and result coverage |
searchMode | Precise matching | Prioritizes exact keyword matching for clause-based compliance basis content |
usingReRank | Enabled | Improves the ranking priority of highly relevant compliance documents, enhances retrieval result strictness |
datasetSearchUsingExtensionQuery | Enabled | Adds synonym expansion for policies and regulations, covers compliance basis expressed in different terms |
collectionIds | Configured for compliance-specific datasets | Limits retrieval scope to knowledge base collections related to legal compliance, avoids interference from unrelated data |
Three common mistakes
- Imported PDF compliance documents cannot be opened in retrieval results. No exclusive adaptation rules for document parsing are configured, or associated parsing plugin configurations are not synchronized during backup and recovery.
- The reranking model continues to occupy video memory after retrieval completes. No logic to automatically unload the reranking model after the retrieval process ends is set, leading to long-term video memory resource occupation.
- Core compliance basis clauses are not included in retrieval results. The
datasetSearchUsingExtensionQueryextended query is not enabled, so synonymous expressions of policies and regulations cannot be covered, resulting in missed recall of key basis.
How to confirm the configuration is correct
- Enter the FastGPT knowledge base retrieval test page, input typical questions related to compliance risks, and check if returned result document types all belong to compliance-related documents.
- View the retrieval configuration panel, confirm the switch states of
usingReRankanddatasetSearchUsingExtensionQuerymatch the preset configuration. - Simulate the monthly data update process, verify if the knowledge base synchronization task can trigger normally and complete the update.
- Check the ranking of retrieval results, confirm highly relevant compliance basis documents appear at the top, meeting preset matching requirements.
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.