What the data for this category looks like
Internal office data for urban public transit comes primarily from enterprise internal operation management manuals, station safety regulations, employee job handbooks, and regularly updated shift schedules, ticket policy documents, and dispatch logs. Data update cycles vary significantly: operation manuals are revised quarterly, shift schedules are updated weekly, and dispatch logs are generated and archived daily. Documents fall into two structure types: long-form standard texts and short-item process documents. Fields include route number, station ID, job title, service hours, and more. Common units are route numbers, hours, and passenger trips.
What constraints these characteristics impose on knowledge base retrieval and recall
Long-form texts such as operation manuals and safety regulations can cause semantic fragmentation after chunking. Retain complete job operation logic during chunking. Frequently updated shift schedules and dispatch logs require a knowledge base synchronization cycle of no more than 7 days. Otherwise, retrieval results will lag behind actual operational arrangements. Short-item process documents exist in large volumes, which can lead to mixed irrelevant results. Strengthen precise matching rules. Documents with structured fields such as route numbers and station IDs must support targeted recall by business identifiers to avoid invalid cross-route results.
How to Configure Settings
The following configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.78 | Semantic similarity threshold for matching urban public transit short process documents and long standard texts, to avoid recalling irrelevant cross-route documents |
limit | Top 8 entries | Controls the number of results retrieved per query, adapting to the information reception capacity of transit employees during single queries |
searchMode | hybrid | Combines keyword and semantic retrieval, covering structured fields such as route numbers and unstructured content in process descriptions |
usingReRank | Enabled | Re-ranks initial retrieval results to filter redundant results with mismatched station IDs |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands colloquial employee queries, such as associating "night shift dispatch" with corresponding shift rule documents |
collectionIds | Bound by business group | Independently associates separate knowledge base collections such as operation manuals, shift schedules, and employee handbooks, to avoid interference from cross-category data |
Three Common Misconfigurations
- Phenomenon: Retrieval results include non-target route station dispatch documents, and the station ID field in the documents does not match the route number from the query. Cause:
collectionIdstargeted binding to corresponding business collections is not configured, resulting in recall of cross-category data. - Phenomenon: Old documents remain in retrieval results after the knowledge base is updated, and results lag behind actual operational arrangements. Cause: No synchronization cycle matching the data update frequency is set, and full knowledge base retraining is not triggered.
- Phenomenon: Local images in imported Markdown files fail to display properly. Cause: Local images are not uploaded to an online image bed, original links are not replaced, and the knowledge base image resource path is not configured as required.
How to Verify Successful Configuration
- Submit a query that includes a route number, and confirm the station ID of the recalled results matches the queried route.
- Trigger a full knowledge base retraining, wait for synchronization to complete, and verify that updated shift schedule documents can be retrieved normally.
- Upload a Markdown file with local image links, replace the links with online image bed links, and confirm that images load correctly in retrieval results.
- Adjust the
similarityparameter, observe changes in the number and relevance of recalled results, and confirm the threshold meets business 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.