What does the data for this category look like?
Data for this category comes from contracts, work breakdown structure (WBS), client schedules, and project management systems. Daily documents include project plans, milestone lists, weekly reports, and risk issue lists. Data updates occur weekly or after a milestone is completed. Most document structures include fields such as task nodes, responsible persons, time nodes, and risk levels. No unified fixed units apply. Some fields require matching corresponding values based on specific project scenarios.
What constraints do these characteristics impose on knowledge base retrieval and recall?
Scattered data sources across multiple systems require cross-source recall to associate document fragments from different systems. Weekly or milestone-based update cycles do not require real-time incremental synchronization, but they require recall logic that supports periodic batch updates. Diverse fields in daily documents require multi-dimensional semantic and keyword matching to cover different business dimensions such as tasks, time, and risks. Operations duty and work order scenarios require quick location of project schedules and risk handling solutions. Clear requirements exist for the timeliness and accuracy of recall results, and irrelevant project content must be excluded from retrieval results.
How to configure the settings?
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Project documents typically include structured fields and clear business keywords. This range balances recall accuracy and coverage |
limit | Top 8–12 results | Operations duty and work order scenarios require fast access to critical information. Excessive results increase screening costs for frontline staff |
searchMode | vector+keyword | Project documents have both semantic descriptions and structured fields. Hybrid search balances semantic matching and precise keyword recall |
usingReRank | Enabled | Multi-source recall results may have redundancy. Reranking improves the relevance of top results, adapting to business needs for quick location |
datasetSearchUsingExtensionQuery | Enabled | User queries in project scenarios often omit context such as project numbers and client names. Extended queries supplement implicit information |
collectionIds | Bind the exclusive collection ID of the corresponding project | Documents from different projects must be isolated to avoid interference from cross-project content in retrieval results |
Three common configuration errors
- Calling the knowledge base content addition interface returns a missing parameter error. The cause is that required fields for this category of documents such as project plans and risk lists are not passed, and only general text content is provided.
- Structured project document retrieval results do not match structured fields. The cause is that multi-vector support for table data is not enabled in knowledge base settings, and only a single text vector is generated.
- Retrieval results include irrelevant cross-project documents. The cause is that the
collectionIdsparameter is not specified, so content from all knowledge base collections under the current tenant is recalled.
How to confirm the configuration is properly set up?
- Users log in to the FastGPT v4.17.0 knowledge base management page, view the settings for the target table document, and confirm that the multi-vector support switch is enabled.
- Users pass the specified
collectionIdsparameter when calling the retrieval interface, and verify that the returned results only include documents from the corresponding project. - Users submit a test query containing keywords such as project nodes and risk levels, and check whether the retrieval results include exclusive documents of this category.
- Users enable the
usingReRankfunction, compare the top results before and after reranking, and confirm that the relevance of the top 3 results meets business expectations.
The configuration values provided here are common starting points, and should be evaluated against specific project samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.