What the data for this category looks like
Data for this category primarily comes from government and enterprise communication operation manuals, service process specifications, work order archive documents, and policy implementation rules. Updates follow a monthly baseline, with temporary additions during major policy adjustments. Most documents are structured long texts, containing fields such as chapter numbering, effective dates, applicable scenarios, operation steps, and quantitative indicators. Quantitative indicators often use units like Mbps, hours, and number of units. Some documents include locally stored topology diagrams and configuration screenshots.
What constraints do these characteristics impose on knowledge base retrieval and recall
Structured long texts have uneven text block spans due to chapter boundaries, which can cause semantic breaks if split across chapters, harming retrieval accuracy. Monthly update frequency requires the knowledge base sync cycle to align with the update rhythm, avoiding data lag or redundancy. Locally stored non-online images attached to documents mean image content cannot be directly associated during retrieval, requiring additional configuration of image parsing paths. Quantitative indicators have dedicated units, so unit matching logic must be retained during retrieval to avoid result deviations from fuzzy matching.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Meets the strictness requirements of government and enterprise documents, avoids low-relevance results |
limit | Top 8 entries | Adapts to the need for multiple reference items in government and enterprise document drafting, balances recall volume and context length |
searchMode | semantic_search | Matches the semantic association needs of structured documents, adapts to the retrieval logic for professional terminology |
usingReRank | Enabled | Performs secondary sorting on initial recall results, corrects local deviations from semantic matching |
datasetSearchUsingExtensionQuery | Enabled | Supplements derivative queries associated with policies, covers extended information needs for document drafting |
collectionIds | Array of dataset IDs specified by business domain | Limits retrieval scope to corresponding domains such as communication operation and work order service, avoids interference from cross-domain documents |
Three common mistakes
- Phenomenon: Local image links in retrieval results fail to load normally, with broken icons displayed on the interface. Cause: Local image hosting paths are not configured, and only text content was imported into the knowledge base without associating locally stored image files.
- Phenomenon: The exported knowledge base file size is only hundreds of KB, with most document content missing. Cause: The full export option is not enabled, and only a small amount of recently updated incremental data was exported.
- Phenomenon: After using the API to upload PDF files or add chunked content, the new content is not retrieved. Cause: The correct
collectionIdsparameter was not specified, or the PDF enhanced parsing switch was not enabled, so the new content was not associated with the corresponding dataset or not fully parsed.
How to confirm the configuration is complete
Reference only; verify against the target FastGPT release and deployment documentation.
- Upload a Markdown document containing local images, perform a retrieval test, and check if the results include normally accessible image association information.
- Perform a full export operation, verify that the size of the exported file matches the total volume of actual documents.
- Call the API to upload PDF files or add chunked content, perform a retrieval test, and confirm that the new content is normally recalled.
- Adjust the similarity threshold range, verify that the relevance of retrieval results meets the requirements of the business scenario.
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 2026-09-28.