What this category of data looks like
Daily documents for this category include growth experiment plans, user segmentation tables, outreach copy, and retention analysis reports. Data sources are user behavior logs, push systems, experiment platforms, and CRM systems. Most updates occur daily or weekly. Document fields typically include user segmentation tags, experiment variables, outreach channels, retention metrics, and other related fields. Some documents require association with real-time user behavior data. The content structure focuses on business process records and analysis result summaries.
What constraints these characteristics impose on the "knowledge base retrieval and recall" link
Data for this category comes from multiple business systems, with most updates on a daily or weekly basis. Documents include segmented business fields such as user segmentation tags and experiment variables. This requires knowledge base retrieval to support multi-source data associated recall, while adapting to the chunking logic for long-text business documents to avoid splitting that breaks business context. Since errors will directly lead to user dissatisfaction, reduced retention, or wasted channel budgets, the retrieval link must ensure stability, avoid timeouts or result deviations, and support filtering recall scope by business fields.
How to set the configurations
The following configurations are based on FastGPT v4.17.0, adapted for the user growth and retention operation business scenario:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Match the semantic matching accuracy of business documents, balance recall coverage and result relevance |
limit | Top 8 entries | Match the conventional analysis dimension count of growth operation documents, avoid context overload |
usingReRank | Enabled | Optimize ranking for long-text business documents, improve precise recall effectiveness |
searchMode | hybrid | Combine keyword and semantic retrieval, adapt to retrieval needs for multi-field business documents |
datasetSearchUsingExtensionQuery | Enabled | Expand queries based on business document fields, cover segmented retrieval scenarios such as user segmentation and experiment variables |
collectionIds | Specified by business domain | Limit retrieval scope to growth operation-exclusive knowledge base collections, avoid interference from irrelevant documents |
Three common mistakes
- Phenomenon: Retrieval takes more than 30 seconds, and cannot quickly return results for documents such as quotation sheets or experiment plans. Cause: Failed to adjust the
limitand re-ranking configuration for business documents, or failed to limitcollectionIds, resulting in an overly large retrieval scope. - Phenomenon: After linking two knowledge bases, the same question only returns results from a single knowledge base. Cause: Failed to enable the multi-knowledge base result merging logic, or failed to correctly set the
limitparameter to restrict the number of returned entries. - Phenomenon: Retrieval results include documents unrelated to growth and retention operations. Cause: Failed to limit the retrieval scope to the exclusive business knowledge base collection via
collectionIds, or thesimilaritythreshold was set incorrectly.
How to confirm the configuration is correct
- Manually enter typical business questions, such as "Q3 user segmentation plan", check whether the returned results include exclusive documents such as user segmentation tables and growth experiment plans.
- Simulate a multi-knowledge base linking scenario, verify that the same question returns relevant results from multiple knowledge bases.
- Test retrieval latency to confirm it meets business response requirements, adjust the
similarityorlimitparameters to optimize speed based on actual conditions. - Check the relevance of retrieval results, adjust the value range of the
similaritythreshold according to business needs.
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.