What the data for this category looks like
Data for this category primarily originates from tracking systems, data warehouses, CRM systems, and work order systems. Update frequencies are mostly daily or weekly. Daily associated documents include product usage analysis reports, tracking plans, indicator caliber tables, and dashboard requirement forms. Common data fields include user ID, function module, operation duration, trigger time, customer ID, work order priority, and more. Units include counts, minutes, days, and others. There are some differences in data field formats across different sources.
Constraints on re-ranking and hybrid retrieval
Scattered data sources and varying formats require hybrid retrieval to cover both semantically similar analysis documents and precisely matched field data, avoiding limitations of single retrieval modes. Data updates occur daily or weekly, so retrieval datasets must be synchronized regularly, and re-ranking thresholds must adapt to the distribution of updated datasets. Unifying indicator calibers is a core time-consuming step. Re-ranking results must prioritize semantic alignment with indicator calibers, which restricts weight settings for retrieval and re-ranking to ensure caliber consistency in analysis results.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches semantic association strength of product usage data, filters low-correlation tracking records and documents |
limit | Top 15–20 entries | Balances retrieval efficiency and sample coverage required for analysis |
searchMode | hybrid | Recalls both semantically similar analysis reports and precisely matched indicator caliber fields, adapting to multi-source data retrieval needs |
usingReRank | Enabled | Performs secondary sorting on hybrid retrieval results to correct semantic bias from basic retrieval |
| Re-ranking model image | Arm architecture compatible image | Adapts to arm architecture deployment environments, avoiding runtime compatibility errors |
Three Common Mistakes
- Issue: The number of hybrid retrieval results does not match the configured
limit, and some knowledge base data is not recalled. Cause: Corresponding retrieval variables are not bound to each knowledge base in the workflow, leading to failed cross-source data association. - Issue:
usingReRankis set to Enabled but results are not re-ranked as expected. Troubleshooting: The required environment variablesAIPROXY_API_ENDPOINT(root address of the AI Proxy service, e.g., http://aiproxy:3000, do not append /v1 or /chat/completions) and non-emptyAIPROXY_API_TOKEN(AI Proxy admin token, consistent with AI Proxy server configuration) for FastGPT v4.17.0 and later App/Admin server are not correctly configured. Missing configuration will cause startup verification failure or functional abnormalities. - Issue: Re-ranking function reports errors after deployment, with logs indicating architecture incompatibility. Cause: An amd architecture re-ranking image was used, and the arm architecture deployment environment was not adapted.
How to Confirm Configurations Are Correct
- Check the App/Admin server environment variables for FastGPT v4.17.0 and later versions, confirm that
AIPROXY_API_ENDPOINT(root address of the AI Proxy service, e.g., http://aiproxy:3000, do not append /v1 or /chat/completions) and non-emptyAIPROXY_API_TOKEN(AI Proxy admin token, consistent with AI Proxy server configuration) are correctly configured and not empty. - Initiate a test retrieval, review the sorting logic of returned results, and confirm that the re-ranking function is active.
- Verify the recall scope of hybrid retrieval, confirm that multi-source data including tracking logs and indicator caliber tables are covered.
- Check the architecture compatibility of the deployment environment, confirm that the re-ranking model image matches the runtime environment.
The values provided are common starting points and 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.