What the data for this category looks like
Typical users of this category are growth operations specialists, user operations specialists, and data analysts. Documents created and used daily include growth experiment plans, user segmentation tables, outreach copy, and retention analysis reports. Data sources include user behavior logs, push systems, experiment platforms, and CRM systems. Data is updated daily or weekly. Document structures include structured user tag fields, experiment variable configurations, outreach timing rules, and other content. Some documents are unstructured analysis reports and copy assets.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Multi-source and heterogeneous data sources require knowledge base retrieval to support integrated recall across systems, avoiding limiting retrieval scope to a single document type. Daily or weekly update frequency requires the retrieval pipeline to support incremental updates, ensuring recalled content is synchronized with the latest user behavior and experiment data. Documents include structured segmentation rules and short-text outreach copy, requiring retrieval matching to balance structured fields and natural language semantics. The most time-consuming user segmentation and outreach rhythm design processes require retrieval response speed to be controlled within a reasonable range, avoiding impact on operational decision efficiency. Errors may lead to user dissatisfaction and reduced retention, so the accuracy and relevance of recall results must be strictly controlled.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the matching accuracy of short-text and structured documents such as outreach copy and user segmentation tables, avoiding recalling irrelevant experiment plans or analysis reports |
limit | 8 | Matches the core reference content related to segmentation and outreach required for daily growth operations, avoiding excessive redundant information interfering with decision-making |
usingReRank | Enabled | For multi-source and heterogeneous user behavior logs and experiment data, re-ranking can filter low-correlation recall results and improve the matching accuracy of outreach copy |
datasetSearchUsingExtensionQuery | Enabled | Adapts to scenario-based questions from growth operations, such as "how to design new user 7-day retention outreach", optimized questions can match knowledge base content more accurately |
collectionIds | Bound to growth operations exclusive dataset ID | Limits the retrieval scope to exclusive documents such as growth experiment plans and user segmentation tables, avoiding recalling irrelevant technical documents or operational content from other categories |
Three common mistakes
- Phenomenon: The number of document chunks returned by retrieval exceeds expectations, or complete long document content cannot be recalled. Cause: Files are not split for scenarios where a single document exceeds 3000 chunks, causing the system to fail to load full document content.
- Phenomenon: The knowledge base selection option does not appear in the tool call popup. Cause: The knowledge base is not bound to the tool call permission scope, or the knowledge base retrieval configuration for tool calls is not enabled.
- Phenomenon: The re-ranking function passes tests, but the actual retrieval returned results have a re-ranking mark of false. Cause: When deploying the re-ranking model via TEI on a compute server, the input and output format of the model is not correctly configured, causing FastGPT to fail to call the re-ranking interface properly.
How to confirm the configuration is complete
- Enter the knowledge base retrieval test page of FastGPT v4.17.0, input questions related to growth operations, and check whether the number of returned results matches the
limitconfiguration value. - View the re-ranking mark of retrieval results, confirm that after enabling the
usingReRankparameter, results are sorted by relevance, with no abnormal false marks. - Check the retrieval scope, confirm that only documents within the growth operations exclusive dataset are recalled, with no irrelevant technical documents or operational content from other categories appearing.
- Upload a single document exceeding 3000 chunks, test whether the required document chunk content can be fully recalled, with no truncation or missing content.
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.