What does the data for this category look like
This category of data mainly comes from code repositories, documentation site build logs, user feedback, and version release records. Daily produced documents include installation guides, API references, migration tutorials, and frequently asked questions. Content is divided into fields by version and module, including update time, document path, version identifier, and main body content. The data update rhythm is mostly weekly synchronization or triggered by version releases, with a single update covering document content for a single version or module.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Documentation site data is divided by version and module. When performing retrieval, version identifiers must be associated to avoid confusion between cross-version content, and document content for the corresponding module must be targeted matched. The high-frequency update rhythm requires the retrieval link to support incremental synchronization configuration, to avoid excessive resource occupation from full refreshes. Technical documents contain professional terminology and version difference content. Recall must prioritize exact keyword matching, while expired content must be filtered. In addition, most document content is structured technical instructions. Retrieval results must be sorted by document chapter or module, to allow operations staff to quickly locate required content.
How to configure settings
The following configuration items apply to FastGPT v4.17.0 and later versions. Corresponding settings must be completed in the FastGPT backend and server environment.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirements for technical documents, filters low-relevance non-target content |
limit | Top 8–12 entries | Balances recall coverage and context length, adapts to the scenario of operations staff quickly browsing content |
searchMode | embedding | For structured technical documents, vector retrieval is more suitable for professional term matching than keyword retrieval |
usingReRank | Enabled | Optimizes the sorting of recall results, prioritizes returning document chapters most relevant to the query |
datasetSearchUsingExtensionQuery | Enabled | Automatically optimizes user queries to adapt to professional terminology expressions in technical documents, improving recall accuracy |
collectionIds | Collection ID of the exclusive documentation site dataset | Targets recall of content from the exclusive dataset maintained for the documentation site, avoiding mixing in content from other business knowledge bases |
Three common configuration errors
- Phenomenon: The number of retrieval results is far lower than the configured
limitvalue, and expected document content is not recalled. Cause:collectionIdsis not configured correctly, causing the retrieval scope to cover non-document site datasets, or an incorrect collection ID is configured. - Phenomenon: Retrieval response time is too long, exceeding expected interaction thresholds. Cause: The
similaritythreshold is not set reasonably, resulting in recall of too many low-relevance content, which increases subsequent context splicing and model call time. OrusingReRankis not enabled, leading to excessive redundant results. - Phenomenon: In external access scenarios, streaming segmented content is not returned as expected, and full results are output directly. Cause: The streaming return configuration on the application side is not enabled correctly, or the streaming response link of AI Proxy is not connected correctly, resulting in interaction experience that does not meet requirements.
How to confirm configurations are correctly applied
- Confirm that the App/Admin server has correctly configured the
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKENenvironment variables, and such parameters are not repeatedly set in the FastGPT backend interface, workflow global variables, or application API configuration entry. - Initiate a test retrieval for documentation site content, input queries such as "installation steps" or "API interface description", and verify that the source of recalled results is the target documentation site dataset, and the number of returned results matches the configured
limitrange. - Compare retrieval response times when
usingReRankis enabled and disabled, to confirm that the impact of the configuration on result sorting and response speed meets expectations. - Test the streaming reply function in external access scenarios, to confirm that retrieval results are returned step-by-step in segmented form, with no lag or one-time full output.
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.