Reference only; verify against the target FastGPT release and deployment documentation.
What this category of data looks like
The data sources for urban public transit business processing mainly come from the public transit group operation scheduling system, citizen service hotline ticket database, and official website service guide section. Data updates are synced in real time alongside business changes. Updates are completed on the same day when new routes launch, stops are adjusted, or processing procedures are modified.
The document structure of each individual data entry is fixed, including fields such as the name of the processing matter, corresponding route number, processing location, required materials, time limit for completion, and consultation phone number. Field units are mostly common public service units such as working days, person-times, and kilometers, with no complex nested structures.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The characteristics of scattered data sources and real-time updates require the knowledge base retrieval link to support incremental synchronization of multi-source datasets, to avoid update delays caused by full indexing.
The clearly defined but widely varying fields of business processing procedures require the retrieval link to support precise matching by fields, to avoid irrelevant route information being introduced by broad semantic recall.
The short length but highly scenario-specific single data entries require reasonable control of the context length of recall results, to avoid redundant information interfering with the LLM's process organization.
In addition, user queries may contain colloquial expressions. This requires the retrieval link to support query expansion to match standardized processing guide documents.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.78 | The keyword precision requirement for public transit business processing is high. A threshold that is too low will introduce irrelevant route information. |
limit | Top 8 entries | The steps for a single business processing procedure usually do not exceed 6. Recalling too many entries will cause context overload. |
searchMode | vector+keyword | Public transit data has both structured fields (route numbers) and unstructured text (processing procedures). The hybrid mode balances precision and recall. |
usingReRank | Enabled | The ranking of multi-source recall results needs to be recalibrated based on the semantic relevance to the user's query. |
datasetSearchUsingExtensionQuery | Enabled | Users may use colloquial expressions such as "how to tap to pay for the bus". Query expansion can match standardized processing guide documents. |
collectionIds | Bind the dedicated dataset ID for public transit business processing | Isolate knowledge base data from other categories to avoid retrieval interference. |
Three common mistakes
- Phenomenon: The number of retrieval results returned is far lower than the set
limitvalue. Cause: Multi-source dataset associated recall is not enabled, and only documents from a single data source are retrieved. - Phenomenon: The AI conversation does not reference the retrieved public transit processing procedure documents. Cause: The context association between the knowledge base retrieval node and the LLM call node is not established in the workflow, so the LLM cannot obtain the recall results.
- Phenomenon: The retrieval request returns empty results. Cause:
datasetSearchUsingExtensionQueryis not enabled, so it cannot match standardized processing guide documents corresponding to the user's colloquial query.
How to confirm the configuration is correct
- Upload a test document for the public transit IC card reissue business, initiate the query "what materials are needed for public transit IC card reissue", and verify that the core content of the document is included in the returned results.
- Adjust the
similaritythreshold to 0.65, observe that irrelevant route information appears in the retrieval results, and confirm that the threshold adjustment takes effect. - After enabling
usingReRank, compare the result rankings before and after reranking, and confirm that results with higher semantic relevance are ranked first. - Check whether the bound
collectionIdsonly include datasets related to public transit business processing, to avoid introducing irrelevant content across datasets.
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.