What the data for this category looks like
Data sources for this category include raw datasets, ETL operation logs, annotation platform records, and sampling quality inspection results. Daily associated documents include cleaning rule documents, quality inspection sampling tables, abnormal data lists, and delivery reports. Data is updated daily or per batch task cycle. Document fields include data ID, anomaly type, processing status, sampling ratio, delivery batch number, etc. There are no unified mandatory units; some fields use units such as items or GB. In operations duty and ticket scenarios, frequently retrieved content focuses on three categories: anomaly type matching, cleaning step references, and historical processing records.
What constraints these characteristics impose on the "reranking and hybrid retrieval" link
Dispersed data sources mean retrieval must cover multiple data types, so hybrid retrieval is a necessary configuration to recall all relevant content. High update frequency requires adapting to real-time or batch update index rhythms to avoid retrieving expired cleaning rules or abnormal data. Diverse document types include structured tables, unstructured logs, and rule text. The reranking link must distinguish content priority, and prioritize matching anomaly types and processing status fields mentioned in operation tickets. Single batch data volume is large, so the number of retrieved results must be controlled to avoid exceeding the context window and reducing ticket processing efficiency.
How to set the configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.70–0.82 | Matches keyword relevance for cleaning rules and abnormal data, adapting to text features of multi-source heterogeneous documents |
limit | Top 6–10 entries | Controls the scale of retrieval results, adapting to context window limits for operation tickets |
searchMode | hybrid | Covers multiple data sources including raw datasets, ETL logs, and annotation platform records, improving recall completeness |
usingReRank | Enabled | Reranks hybrid recall results by anomaly type and processing priority mentioned in tickets, prioritizing highly relevant content |
| Rerank model response timeout | 25–35 seconds | Adapts to real-time response requirements for operations duty, avoiding excessive wait times that reduce ticket processing efficiency |
Three common mistakes
- Phenomenon: Reranking model is enabled during online recall testing, but returned results do not apply reranking logic. Cause: The reranking association option was not enabled synchronously in the knowledge base index configuration, resulting in vector data adapted for reranking not being generated during the indexing phase.
- Phenomenon: An Invalid URL (POST /v1/rerank) error is returned when calling the reranking model. Cause: The API endpoint configuration for the reranking model does not match the path of the actual deployed service, and the full request address of the target service was not filled in correctly.
- Phenomenon: Retrieved results do not prioritize displaying abnormal data lists or cleaning rules. Cause: A reasonable similarity threshold was not set, causing low-relevance content to be recalled first, and priority was not adjusted through the reranking link.
How to confirm the configuration is correct
- Enter the knowledge base retrieval test page, enable the reranking model switch, enter test ticket text containing anomaly type keywords, and check whether returned results prioritize displaying cleaning rules or abnormal data lists.
- View the reranking model call logs, confirm that normal POST requests are sent to the configured API address, and there are no connection timeout or URL error records.
- Adjust the value range of the similarity threshold, verify whether the relevance of retrieved results changes as expected with threshold adjustment.
- Disable the reranking model switch, compare the sorting differences between the two retrieval results, confirming that the reranking link is indeed active.
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.