What this category of data looks like
The data sources for response template recommendation mainly include historical resolved tickets from government hotlines and state-owned enterprise service hotlines, officially published standard response template documents, and real-time pending tickets in circulation. Data updates follow full daily synchronization or real-time incremental update after a ticket is resolved. The document structure of a single data entry includes fields such as question category, original user query, standard response content, accepting department, and resolution time. Most field types are strings; some fields like resolution time use date format. Response content length ranges from short text to long paragraphs. Reference only; verify against the target FastGPT release and deployment documentation.
Constraints for Knowledge Base Retrieval and Recall
Since data sources include colloquial original user queries and standardized responses, retrieval must balance semantic matching and keyword accuracy. Do not rely solely on a single matching pattern, as this can lead to missed recalls or incorrect matches. Data updates occur frequently, so the retrieval link must access the latest incremental data in real time. This prevents use of expired response templates. Documents include a question category field, which can filter and narrow the retrieval scope to reduce interference from irrelevant data. Maintain consistency of the category field. Inconsistent category labeling can cause retrieval bias. Single response content lengths vary widely. Configure a reasonable recall upper limit to ensure the total context length of retrieval results fits the large model’s input limits.
Configuration Settings
| Configuration Item | Recommended Value/Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Hotline questions mostly fall into similar scenarios; this range balances recall completeness and relevance |
limit | 10 | Ticket questions have high repetition rates; 10 results cover most common response templates |
searchMode | vector+keyword | Adapts to colloquial hotline questions, balancing semantic matching and keyword accuracy |
usingReRank | Enabled | Filters low-relevance recall results to improve the accuracy of recommended templates |
datasetSearchUsingExtensionQuery | Enabled | Performs semantic expansion on colloquial questions to improve matching accuracy |
collectionIds | Enter the exclusive hotline knowledge base ID | Defines retrieval boundaries to avoid interference from irrelevant data in search results |
Three Common Configuration Mistakes
- Symptom: A knowledge base retrieval node is configured in the workflow, but retrieval does not trigger during runtime, and the process proceeds directly to pure AI chat. Cause: The retrieval node’s output is not correctly connected to the large model’s context input parameter, or the node trigger conditions are set incorrectly.
- Symptom: Unexpected markdown heading 1 and heading 2 formatting remains in knowledge base content returned by retrieval. Cause: Original documents imported to the knowledge base were not pre-cleaned of formatting, or format filtering configuration in the chunking stage was not enabled.
- Symptom: The large model’s output response includes the phrase "Citation: [1]". Cause: No requirement to remove citation identifiers was added to the large model prompt, or the retrieval results’ citation field was not correctly stripped.
How to Confirm Correct Configuration
- Manually enter a common hotline question, and check if retrieval results return matching standard response templates.
- Review workflow run logs to confirm the knowledge base retrieval node triggered normally and returned the expected number of results.
- Check the large model output content to confirm no unexpected formatting markers or citation identifiers are present.
- Adjust the
similarityorlimitparameter to verify whether retrieval result relevance or quantity matches the expected adjustment direction.
The values provided are common starting points and should be evaluated using local test 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.