What the data for this category looks like
Daily on-site deployment documents include deployment manuals, installation records, configuration lists, and on-site check-in sheets. Data sources include customer data center environments, equipment lists, software packages, and network topologies. Data updates are adjusted with each project order or phase progress. Associated entities are implementation engineers, system engineers, and customer IT staff. Core fields include equipment model, software version, configuration parameters, on-site check-in time, and other details. The overall data consists of scattered heterogeneous on-site documents without a unified standardized format.
What constraints these characteristics impose on reranking and hybrid retrieval
Data sources are scattered and consist of heterogeneous customer site documents. Hybrid retrieval must support deployment manuals, installation records, and other data sources with different structures. It cannot rely on a single retrieval logic. Individual data volumes are small, but updates change with each order. Retrieval links must be lightweight to support rapid on-site troubleshooting. Equipment and software information linked to on-site work orders is scattered across multiple documents. The reranking link must prioritize matching the current customer’s exclusive configuration fields. This avoids retrieving generic documents. It must also control computing resource consumption to adapt to the limited hardware environment of on-site deployments.
How to set configurations
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | On-site deployment work orders require precise matching of the current customer’s equipment and documents. Reranking filters irrelevant generic retrieval results |
similarity | 0.72–0.80 | On-site deployment documents are mostly customer-specific configurations. This threshold filters low-similarity non-current customer documents |
limit | Top 6–10 entries | The number of on-site documents associated with a single work order is limited. Excessive retrieval results increase GPU video memory usage and computation time |
searchMode | Hybrid retrieval | Must retrieve heterogeneous data sources such as deployment manuals, installation records, and configuration lists simultaneously. It integrates multiple types of on-site information |
Three common misconfigurations
- Phenomenon: GPU video memory usage grows rapidly to 6–7 GB after calling the reranking model. Multiple GPUs become occupied. Cause: The number of retrieved documents per call is not limited. During concurrent calls, model video memory usage accumulates beyond the single-card threshold.
- Phenomenon: An offline reranker model returns
rerank error fastgpt | { message: { detail: ' (truncated error). Cause: The model loading path is not configured correctly. The document chunk length exceeds the maximum input length supported by the model. - Phenomenon: A large number of generic documents not from the current customer’s on-site environment appear in hybrid retrieval results. Cause: The similarity threshold is set too low. It fails to filter generic deployment documents unrelated to the current work order.
How to confirm configurations are properly set
- Check the GPU video memory usage of the reranking model. Confirm single-card usage does not exceed the on-site server’s preset threshold.
- Submit a retrieval request for an on-site deployment work order. Verify returned results prioritize matching the current customer’s equipment and configuration information.
- Switch the retrieval mode to hybrid retrieval. Confirm multiple types of documents such as deployment manuals and installation records are returned simultaneously.
- Adjust the similarity threshold. Verify the number of retrieval results changes as expected with the threshold adjustment.
The values given 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.