What the data for this category looks like
Data for this category originates primarily from enterprise existing IT system inventories, official domestic software and hardware directories, project test environment records, and publicly available vendor materials. Daily produced documents include software and hardware adaptation lists, compatibility test reports, system migration plans, and operations and maintenance issue ledgers. Data update cycles follow individual independent innovation replacement projects, or synchronize the latest domestic software and hardware directories quarterly. Document structures mostly use software and hardware model, version number, compatibility status, and test conclusion as core fields, with no unified fixed format, adjusted based on project requirements.
What constraints these characteristics impose on the reranking and hybrid retrieval link
Data sources are scattered and formats are inconsistent. Hybrid retrieval must accommodate field differences across multi-source heterogeneous documents to prevent chaotic search results. Update frequencies align with individual project cycles or quarterly intervals. Reranking models must support dynamic adaptation to newly added domestic software and hardware entries, and cannot rely on fixed knowledge base snapshots. Core documents are mostly technical test content, with fields such as model and version designed for precise matching. Hybrid retrieval must prioritize associating precisely matched hardware models and compatibility conclusions. Single-project data volumes are relatively controllable but scenario-focused. Reranking thresholds must align with the short-text characteristics of technical documents, avoiding redundant results that disrupt rapid decision-making during operations on-duty shifts.
How to set configurations
This page applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled (set to true) | Documents for independent innovation replacement adaptation mostly involve precise technical matching. Enabling reranking filters low-relevance generic results, focusing on compatibility status and model matching items |
similarity | 0.75–0.85 | Adaptation documents mostly rely on precise model matching. A threshold that is too low will introduce irrelevant hardware entries, while a threshold that is too high may miss compatibility test conclusions |
limit | Top 8 entries | Single-project operation and maintenance data volume is limited. Retaining the top 8 entries covers major compatibility scenarios, avoiding excessive redundant content for on-duty personnel to review |
searchMode | Hybrid retrieval mode | Simultaneous matching of system model, version number, and compatibility conclusion is required. Hybrid retrieval balances keyword matching and semantic association |
Three common errors
- Symptom: After enabling
usingReRank, the application returns no search results, but the knowledge base test returns results normally. Cause: The reranking model’s API key or interface address is not configured, or the reranking model’s input length limit is lower than the chunked document length of the knowledge base. - Symptom: Post-reranking results are sorted incorrectly, inconsistent with expected compatibility priority. Cause: The
similaritythreshold is set too high, filtering some relevant compatibility test reports, or thelimitvalue is too low, failing to retain enough candidate results for the reranking model to sort. - Symptom: Search results include a large number of general IT documents unrelated to independent innovation adaptation. Cause: Hybrid retrieval mode for
searchModeis not enabled, only pure semantic retrieval is used, and precise matching of hardware model fields is not associated.
How to confirm configurations are correctly set
- Enter the knowledge base test page of the FastGPT application, enable the reranking function, enter an operation and maintenance question containing a specific hardware model, and check whether the returned results first display adaptation lists and compatibility test reports.
- View the FastGPT application’s retrieval logs to confirm that the
usingReRankparameter is active, and the reranking model’s call request returns a normal status code. - Adjust the test range of the
similaritythreshold, observe changes in the relevance of search results, and confirm that the threshold setting matches the document characteristics of the current project. - Verify the hybrid retrieval effect of multi-source documents, confirm that results matching both hardware models and compatibility conclusions are displayed first.
The values provided are common starting points and should be evaluated against sample datasets relevant to the specific deployment.
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.