What Data for This Category Looks Like
Data for this category comes primarily from product help centers, customer organizational structures, training recording videos, and ticket systems. Daily generated documents include training slides, user manuals, training attendance sheets, and problem Q&A lists. Document structures usually include fields such as chapter numbers, operation steps, applicable roles, and version numbers. Data updates occur once per order or product version iteration. Single update content centers on function adjustments for the current version, or training requirement adjustments for individual customers.
Constraints on Citation Sources and Traceability
Dispersed data sources and inconsistent update frequencies require citations to match the latest version of document content. This avoids citing outdated training slides or older user manuals. Differences in fields across documents require accurate extraction of identifying information for each corresponding document during traceability. Examples include version numbers for user manuals, and recording times for training videos. Q&A content from ticket systems is mostly scattered single questions. Traceability requires linking these questions to corresponding training segments or manual chapters. It also requires deduplication of multi-source data to avoid repeated citations of the same customer question. Additionally, single content entries for this category of documents have relatively controlled length. Reasonable citation range settings ensure answer completeness and accuracy.
How to Configure Settings
| Configuration Item | Recommended Approach | Basis for This Approach |
|---|---|---|
quoteMaxToken | 800–1200 token | Matches the single-chapter content length of training slides and user manuals. Prevents citation content from exceeding context limits. Based on the context processing logic of FastGPT v4.17.0 |
| Similarity Threshold | 0.5–0.7 | Adapts to scattered ticket Q&A content. Filters low-relevance unrelated documents while retaining sufficient candidate sources |
responseData | Enable source field return | Used to extract identifying information such as document version numbers and update times. Ensures traceability information is accurate and traceable |
| Single Search Limit | Top 3 results | Adapts to the relatively concentrated quantity of documents in this category. Avoids excessive redundant citations interfering with user reading |
detail | Enable detailed traceability | Displays specific chapters or question entries of documents. Helps users quickly locate corresponding training or manual content |
Three Common Configuration Mistakes
- Symptom: The AI returns only plain text with no source information. Cause: The
responseDataparameter is not enabled, or the source return switch is not configured. This causes the API call to not return traceability fields for the corpus. - Symptom: Cited content has very low relevance to the query, and even unrelated training attendance sheet content appears. Cause: The similarity threshold is set too low, such as below 0.4. This leads to a large number of unrelated documents being retrieved, and matching rules are not adjusted for the document structure of this category.
- Symptom: After re-uploading abnormally trained files, the knowledge base still shows repeated citations of the same content. Cause: Knowledge base deduplication configuration is not enabled, or original abnormal entries are not overwritten during upload. This leaves duplicate data uncleaned.
How to Confirm Configuration Is Complete
- Call the API to send a test request. Check if traceability-related fields appear in the returned results. Confirm source information returns normally.
- Input a query known to correspond to a specific training slide chapter. Check if the cited documents in the returned results match that chapter. Verify that matching accuracy meets expectations.
- Upload a batch of documents of the same type. Trigger training and check the training status. Confirm no abnormal errors occur. Also verify that repeated recall of the same content does not happen during citation.
- Adjust the
quoteMaxTokenparameter. Check if the length of the returned cited content falls within the set range. Verify that the parameter takes effect.
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.