What the data for this category 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. Update frequency is daily or weekly. Data fields include user segmentation tags, outreach response rates, experiment group identifiers, and more. There are no unified fixed units. Field meanings must be matched based on specific document types. The most time-consuming daily tasks are user segmentation, outreach rhythm design, and experiment result attribution.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Data sources for this category are scattered and heterogeneous. Hybrid retrieval requires integrating multi-source data such as user behavior logs and CRM systems, so cross-source association retrieval rules must be configured. Different documents have different update frequencies. Daily updated data requires a more appropriate retrieval refresh interval, while weekly updated documents can use a longer cycle. Document types include segmentation tables, outreach copy, and more. Reranking requires setting differentiated weights for different document types. Since the most time-consuming daily tasks are user segmentation and experiment attribution, the response duration of the retrieval link must be controlled to avoid slowing down the overall workflow. Errors will lead to user dissatisfaction and reduced retention, so the relevance check of retrieval results must be more strict.
How to set the configuration
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the semantic matching accuracy required for user segmentation tags and outreach response data, balancing recall scope and result relevance |
limit | Top 10–15 results | Matches the common result viewing volume used in daily retention analysis and experiment attribution, avoiding too many results slowing down workflows |
searchMode | Hybrid Retrieval | Requires integrating multi-source data such as user behavior logs and CRM customer tags; single semantic retrieval cannot cover cross-source associated information |
usingReRank | Enabled | For multi-source heterogeneous retrieval results, reranking optimizes the relevance of outreach copy matching and user segmentation recommendations |
Three common mistakes
- Symptom: After enabling
usingReRank, result reranking always returns false, and returned results are not reordered as expected. Cause: The reranking model's runtime environment is not configured correctly, or the reranking execution node is not bound in the retrieval workflow. - Symptom: Hybrid retrieval returns results with weaker relevance than single semantic retrieval. Cause: Differentiated weights are not set for multi-source heterogeneous data, leading to low-value data sources interfering with overall retrieval accuracy.
- Symptom: When an HTTP request outputs multiple variables in a workflow, specified variables cannot be associated with the corresponding knowledge base. Cause: Mapping rules between variables and knowledge base retrieval entrances are not configured in the workflow, causing retrieval to fail to match target data sources.
How to confirm configuration is complete
- Review system retrieval logs to confirm that returned results include associated data from both user behavior logs and CRM systems, verifying that hybrid retrieval functionality is active.
- Manually adjust the
similarityparameter value, observe changes in retrieval result relevance, and confirm that parameter configuration takes effect normally. - Compare result sorting differences between enabling and disabling
usingReRank, verifying that reranking functionality operates correctly. - Review system version information to confirm that the current running version is FastGPT v4.17.0, which matches the parameter requirements of this configuration.
The values provided here are common starting points. They should be evaluated 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.