What the Data for This Category Looks Like
The data for this category mainly comes from target systems, public vulnerability databases, automated scanner results, and manual attack verification records. Daily produced documents include test authorization letters, penetration test reports, vulnerability reproduction records, rectification suggestions, and other materials. A single document usually contains fields such as project number, vulnerability severity level, reproduction steps, authorization scope, and more. Data updates are carried out on a per-project basis, and knowledge base content is synchronized when each project is completed or when a major vulnerability version is updated.
Constraints on the Knowledge Base Retrieval and Recall Link
The characteristics of this category of data impose multiple constraints on the retrieval and recall link. First, the data includes structured scan results and unstructured manual verification records, requiring precise recall adapted to multi-format content. Second, data is updated on a per-project basis, so support for incremental synchronization of the knowledge base by project dimension is needed to avoid redundant data interfering with retrieval. In addition, documents contain test authorization scope and vulnerability severity level fields. Retrieval must associate and filter content from unauthorized projects to prevent legal authorization disputes. Meanwhile, vulnerability reproduction steps have high detail density, so the length of recalled segments must be limited to avoid context overload.
Configuration Recommendations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Penetration test documents contain a large number of professional security terms. This range can filter low-similarity irrelevant content and avoid recalling unrelated vulnerability reports |
limit | Top 8–12 entries | Penetration test reports have many detailed content dimensions. An appropriate number of recalled entries can cover reference information for different vulnerability types while controlling the total context length |
usingReRank | Enabled | Retrieval results for penetration test scenarios are easily disturbed by term synonyms. Reranking can improve the sorting priority of highly relevant content |
datasetSearchUsingExtensionQuery | Enabled | Customer pre-sales questions often contain vague vulnerability descriptions. Extended queries can supplement relevant security terms to improve recall coverage |
collectionIds | Configure corresponding collection IDs grouped by project | Penetration test data is stored isolated by project. Retrieval by collection ID can prevent sensitive authorized data leaks across projects |
| Chunk length | 800–1200 characters | Vulnerability reproduction steps in penetration test documents have many details. This range can retain complete step information and avoid truncating critical content |
| Retrieval timeout | 60–120 seconds | Penetration test documents have long content. This range can ensure that the complete retrieval process is completed |
Three Common Misconfigurations
- Uploading a DOCX document containing images triggers an error indicating incorrect file format. This occurs if the image processing logic of the document parsing plugin is not configured correctly, or the uploaded file exceeds the 100 MB single-file size limit supported by the knowledge base.
- Enabling question answering splitting mode does not enable question and answer recall for image content. This happens when the image OCR text extraction configuration is not enabled, or the retrieval link does not associate the text content parsed from the image.
- Calling the external program interface for uploading HTML titles and content fails to correctly associate the content with the specified knowledge base. This occurs if the
titleandcontentfields are not passed as required by the interface, or the field formats do not meet the interface specifications.
How to Verify Proper Configuration
- Verify that the AI Proxy environment variables
AIPROXY_API_ENDPOINT(root address of the AI Proxy service, do not append path suffixes such as/v1) andAIPROXY_API_TOKEN(AI Proxy administrator token, must match the AI Proxy server configuration and be non-empty) on the FastGPT v4.17.0 version App/Admin server are correctly configured. - Upload a DOCX penetration test report containing images to verify that parsing succeeds and no format error prompts appear.
- Initiate a pre-sales consultation containing a vague vulnerability description, and confirm that the number of retrieval results matches the range of the configured
limitparameter, and that the sorting meets expectations. - Call the external upload interface to pass HTML titles and content, and verify that the content is successfully synchronized to the specified knowledge base collection.
The values provided here 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.