What data falls under this category
This category of data originates primarily from code repositories, documentation site build logs, user feedback, and version release records. Daily produced documents include installation guides, API documentation, migration tutorials, and FAQs. These are organized by functional modules, with version tags and code sample blocks attached. Update cadence follows weekly iterations or syncs with version releases. Fields include document version number, last update time, and associated code branch. There are no fixed units; values adjust based on document type.
Constraints on reranking and hybrid retrieval
Documentation for site maintenance consists mostly of technical explanatory material. This includes code snippets, version difference notes, and structured FAQs. Hybrid retrieval must differentiate weightings for different content types. This avoids confusing recall of code samples and text explanations. Data updates occur frequently. Retrieval indexes must sync latest version content quickly, to prevent recall of expired documents. Documents are organized by functional modules. The reranking step must prioritize matching functional modules tied to user questions. It must focus on specific technical explanatory content. High-frequency questions added via user feedback must be integrated into the retrieval pool. This improves accurate matching rates.
How to set configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.62–0.78 | Filters non-relevant documents with too low match scores. Meets the precise retrieval requirements for technical documents, and avoids recalling expired or irrelevant version notes |
limit | Top 6–10 results | Controls the number of initial recall results. Prevents context from exceeding model processing limits, and adapts to the long-text characteristics of documentation sites |
searchMode | hybrid | Enables hybrid retrieval mode. Integrates semantic matching and keyword matching to cover users with different query methods, such as direct searches for API names or functional scenarios |
usingReRank | Enabled | Performs secondary sorting on initial recall results. Prioritizes returning technical content most closely aligned with user questions, and improves response quality for pre-sales Q&A and client technical support |
| Rerank model timeout threshold | 25–35 seconds | Adapts to batch indexing and real-time Q&A scenarios for documentation site maintenance. Prevents request blocking that causes index interruptions or response delays |
Three common configuration mistakes
- Symptom: After enabling
usingReRank, online recall tests do not return reranked results. Cause: Reranking configuration was not enabled during the knowledge base indexing step, or no valid rerank model interface was bound. - Symptom: The rerank model returns an error message when called. Cause: Correct rerank model address and key were not configured, or the model version does not meet FastGPT v4.17.0 compatibility requirements.
- Symptom: Calling the rerank model displays
Invalid URL (POST /v1/rerank). Cause: The rerank model interface path was configured incorrectly, or the proxy service did not properly forward rerank requests.
How to confirm configurations are set correctly
- Navigate to the knowledge base retrieval test page in FastGPT v4.17.0. Enter keywords for the target document, and check whether initial recall results and reranked results show differences.
- View the knowledge base indexing logs. Confirm that the rerank model call status is successful, with no timeout or error records.
- Adjust the value range of the
similarityparameter. Verify that retrieval result matching rates change in line with configuration adjustments. - Manually trigger a full indexing task. Confirm that the reranking configuration takes effect synchronously with the indexing task.
The values provided here are common starting points. They should be evaluated 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.