What this type of data looks like
Data is sourced from logging systems, distributed tracing platforms, metrics monitoring systems, and business application SDKs. Update frequency follows daily cycles or version iteration cycles. Regular associated documents include logging collection specifications, tracing instruction documentation, metric dictionaries, and dashboard configuration records. Data fields include log level, timestamp, trace ID, metric name and corresponding value, and other fields. There are no unified mandatory units; confirm details based on specific collection configurations.
What constraints do these characteristics impose on the re-ranking and hybrid retrieval workflow
Multi-source data origins create differences in field formats. Hybrid retrieval must accommodate the log, tracing, and metric field structures of different systems to prevent format confusion in retrieval results. The daily or version-tied update rhythm requires the retrieval knowledge base to sync the latest collection specifications and dashboard configurations regularly, ensuring reference data used in the re-ranking workflow complies with current business standards. Diverse associated document types require the re-ranking workflow to prioritize matching high-frequency questions from pre-sales Q&A or technical support. The length of individual log and tracing data varies widely. The re-ranking workflow must adapt to context content of different lengths, avoiding exceeding model input limits.
How to configure settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Log and tracing data have strong contextual relevance. Re-ranking filters low-relevance recall results and improves the accuracy of pre-sales Q&A |
limit | Top 10 entries | Logging platforms typically return a large volume of recall results per retrieval round. Limiting to the top 10 entries balances retrieval efficiency and coverage |
similarity | 0.72–0.80 | Semantic similarity of log fields must align with the business scenario. Too low a value introduces irrelevant logs; too high a value filters valid troubleshooting clues |
searchMode | Hybrid retrieval | Log text, tracing records, and metric data must be recalled simultaneously. Hybrid retrieval integrates results from multiple data sources |
| Re-ranking model timeout | 300 seconds | Retrieval and re-ranking of log data require processing long texts. 300 seconds covers processing time for most scenarios |
| Re-ranking model type | bge-rerank-base or re-ranking model deployed via Ollama | Adapts to the semantic characteristics of log and tracing data, meeting retrieval requirements for technical support scenarios |
Three common configuration errors
- Phenomenon: Re-ranking model calls return an Invalid URL (POST /v1/rerank) error. Cause: The re-ranking model API address is not configured correctly, or the third-party interface re-ranking path does not match FastGPT's call format.
- Phenomenon: A 401 unauthorized error occurs when testing access to the bge-rerank-base image. Cause: The re-ranking model access key is not configured correctly, or the key does not have permission to access the corresponding interface.
- Phenomenon: The number of results after re-ranking does not match the set
limitvalue. Cause: The actual valid number of recall results is less than the set limit, or the re-ranking model filters too many low-relevance results, resulting in insufficient remaining entries.
How to confirm successful configuration
- On the knowledge base configuration page of FastGPT v4.17.0, verify that the re-ranking switch is enabled and the
usingReRankparameter is active. - Initiate a simulated technical support query, and check if the number of retrieval results matches the set
limitvalue. - Compare retrieval results with and without re-ranking enabled, and confirm that re-ranked results are better aligned with the query's business scenario.
- Review re-ranking model call logs, and confirm there are no errors and normal interface returns.
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.