What this category of data looks like
This type of data typically originates from the customer’s ITSM system, monitoring platform, asset inventory, and duty logs. Daily generated documents include on-site daily reports, service desk tickets, inspection reports, and SLA monthly reports. Data update frequencies fall into two categories: daily updates and monthly updates. Most document structures rely on structured fields. For example, on-site daily reports include inspection time windows, handled tasks, and feedback outcomes. SLA monthly reports include service compliance rate statistics. There is no unified mandatory standard format for fields, so documents must be organized per the delivery requirements of outsourcing services.
What constraints do these characteristics impose on knowledge base retrieval and recall
Heterogeneous data from multiple sources creates a need for retrieval scope adaptation, requiring compatibility with non-standard format content exported from different systems. Daily updated tickets and duty logs require retrieval response speeds to match the timeliness of on-site processing. Monthly updated SLA reports allow slightly longer recall delays. Retrieval in on-site scenarios must cover on-site engineers’ on-site operation references and service managers’ contract delivery basis. It must also avoid recalling irrelevant internal sensitive customer data. At the same time, errors directly impact service contract performance, so the accuracy and relevance of retrieval results must be strictly controlled.
How to configure the settings
This document adapts to the knowledge base retrieval configuration logic of FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the scenario requirements of on-site tickets, balances recall relevance and coverage, and avoids missing key processing steps |
limit | Top 8 entries | Adapts to quick review needs during on-duty shifts. The number of single returned entries is sufficient to cover common operations scenarios, while controlling retrieval time |
searchMode | semantic | Targets unstructured content in on-site daily reports and inspection reports, uses semantic matching to recall relevant processing experience |
usingReRank | Enabled | Optimizes the ranking of recall results, prioritizes returning the most matching processing solutions for current tickets, and reduces on-site decision-making costs |
datasetSearchUsingExtensionQuery | Disabled | The knowledge base for on-site services is a fixed set of delivery documents, prevents the model from generating additional content that deviates from service requirements |
collectionIds | Specify the exclusive collection ID for on-site services | Isolates other non-outsourcing related internal customer data, ensures the retrieval scope only covers compliant service documents |
Three common configuration mistakes
- Insufficient number of retrieval return entries, only 1 result is returned. The cause is that the
limitconfiguration value is set too low, failing to cover reference needs for multi-step processing in on-site scenarios. - Retrieval results mix documents from non-exclusive collections. The cause is failing to specify
collectionIdsas the exclusive collection for on-site services, leading to recall of irrelevant internal sensitive customer data. - Professional operations and maintenance terms in the knowledge base are incorrectly split. This happens when no split prohibition rule is configured for corresponding fields, leading to term splitting that reduces retrieval accuracy.
How to verify successful configuration
- Import a sample on-site daily report, perform a retrieval test, and check whether the returned results include relevant content from this document.
- View the FastGPT retrieval log, confirm that the
limitandsimilarityparameters take effect, and the returned results meet the configured value range. - Attempt to retrieve documents from non-exclusive collections, confirm they are not recalled, and verify that the data isolation configuration works.
- Simulate an on-site ticket query, check that the recall results only include original text fragments from the knowledge base, and no additional content generated by the model appears.
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 2026-10-19.