What the data for this category looks like
The entities involved in this category are mostly emergency management bureaus, water affairs bureaus, meteorological bureaus, sub-districts, towns and townships, and emergency command centers. Daily produced documents include duty logs, early warning notices, consultation minutes, rapid disaster reports, emergency response plans, material allocation slips, and more. Data mainly comes from meteorological early warning systems, hydrological monitoring systems, emergency command platforms, grassroots disaster reporting systems, and video surveillance, with a daily update frequency. Most documents contain precise time, location, and numerical fields, with units including millimeters, cubic meters, number of people, and others. Most document structures are divided according to event processes or departmental division of labor.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The data for this category is mostly real-time updated structured and semi-structured content, and contains a large number of precise numerical values and exclusive units. This requires the retrieval link to accurately match fields and unit information. The complexity of metadata brought by multi-source data requires that retrieval results must simultaneously associate the original metadata of the corresponding system to avoid information loss. The structure of documents divided by event processes or departmental division of labor requires that the hierarchical subordinate relationship between content be retained during retrieval to avoid logical disconnection. The real-time requirements of emergency response also put clear constraints on the response speed of retrieval, ensuring that results can be quickly transmitted to the drafting link.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category of documents contains a large number of precise numerical values and scene descriptions. This interval filters low-match irrelevant content and ensures the relevance of retrieval results |
limit | Top 8 entries | Emergency document drafting needs to cover multi-dimensional rainfall and water condition information. Too many results increase the burden of context processing, while too few results fail to cover all key information |
searchMode | Hybrid vector retrieval + keyword retrieval | This category of data contains structured numerical values and semi-structured text. The hybrid mode matches both semantics and precise fields |
usingReRank | Enabled | In emergency scenarios, content most relevant to the current drafting task needs to be recalled first. Reranking optimizes the result sorting logic |
datasetSearchUsingExtensionQuery | Enabled | Queries during document drafting are often scattered scene descriptions. Optimization expands query semantics and covers more relevant documents |
collectionIds | Bind dedicated collections for emergency duty, meteorology and hydrology | Avoid retrieving documents from unrelated businesses and ensure the accuracy of results |
Three common mistakes
- Phenomenon: Retrieval results only return text fragments without corresponding metadata fields. Cause: Metadata association logic is not configured, only vector fragments are recalled, and information such as the system source and timestamp of the original data is not bound.
- Phenomenon: The title hierarchy and subordinate relationship are lost after markdown format documents are imported into the knowledge base. Cause: The knowledge base chunking process does not retain title hierarchy identifiers, and only segments according to plain text, leading to disconnection of the logical subordinate relationship between content.
- Phenomenon: Extra spaces appear between numbers and text in retrieval results. Cause: Format residues of original documents are not processed during data import, or the connection between numerical values and text is not standardized during chunking.
How to confirm the configuration is correct
- Initiate a test query containing precise numerical values and locations, and check whether the returned results carry corresponding metadata fields such as system source and time.
- Upload a markdown test document containing multi-level titles, and check whether the retrieval results retain the subordinate hierarchical relationship between titles.
- Import test data containing connections between numbers and text, and verify that there are no extra spaces between numbers and text in the retrieval results.
- Simulate high-frequency queries in emergency scenarios, and check whether the retrieval response time meets the requirements of real-time disposal.
The values provided here are common starting points and should be evaluated 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.