What data for this category looks like
Data sources for this category include advertising platform backends, customer qualification libraries, contract systems, and publicly available regulatory rules. Daily processed documents fall into four categories: ad review forms, qualification attachments, delivery records, and rectification notices. Data update frequency is adjusted per single ad campaign or natural month. Core document information covers compliance markers, delivery periods, qualification validity periods, violation types, and other dimensions. There is no unified structured format. Some documents are unstructured scanned materials or text records.
What constraints do these characteristics impose on the reranking and hybrid retrieval link
Data sources for this category are scattered, and format differences are significant. Hybrid retrieval must support cross-database recall of multi-source data, while adapting to metadata weight configurations for different documents. The update rhythm per campaign or monthly means retrieval scope must be limited to valid documents within the corresponding period. This avoids redundant historical data interfering with retrieval results. Errors in compliance record-keeping are directly linked to regulatory inspections, customer liability, and account penalties. The reranking step must strictly filter low-relevance results. It must ensure recalled content is core documents required for compliance reviews and evidence retention, while reducing manual screening time.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
usingReRank | Enabled | This category requires strict filtering of low-relevance evidence. Enabling reranking improves result relevance and prevents irrelevant documents from being included in compliance retrieval results |
similarity | 0.75–0.85 | Compliance documents must highly match ticket questions such as violation types and delivery periods. This range filters irrelevant historical delivery records and qualification files |
limit | Top 10 entries | Compliance record-keeping requires sufficient coverage of evidence chains. Too many results increase manual organization time. 10 entries balances recall completeness and screening efficiency |
searchMode | Hybrid retrieval mode | Data for this category comes from multiple systems. Hybrid retrieval can recall both structured and unstructured documents, adapting to retrieval needs for multiple data types |
Three common errors
- Phenomenon: Reranking function does not activate. Retrieval results are not sorted by relevance, and remain in the random order of basic recall. Cause: The
usingReRankconfiguration is not enabled, or a reranking model suitable for compliance scenarios is not bound in FastGPT v4.17.0. - Phenomenon: Retrieval results contain a large number of low-relevance documents, leading to excessive manual screening time. Cause: The
similarityvalue is set too low, failing to filter qualifications or delivery records unrelated to the ticket question. - Phenomenon: The reranking step returns empty results or throws an error. Cause: The
limitvalue is set too large, causing the reranking model to process more documents than the system's carrying capacity, or the retrieval time range is not limited, recalling expired invalid documents.
How to confirm the configuration is complete
- Log in to the FastGPT v4.17.0 scenario configuration page, and check if the
usingReRankswitch is in the enabled state. - Enter a test ticket question such as "Recent advertising compliance review records", and verify that the
similarityscores of retrieval results fall within the preset range. - Check that
searchModeis set to hybrid retrieval mode, and confirm that both qualification attachments and delivery records can be recalled. - Check that the number of retrieval results matches the value set in the
limitconfiguration, and verify that the number of results from the reranking step meets expectations.
The values provided here 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.