What data looks like for this category
Data for this category comes from asset management systems, procurement systems, financial fixed asset modules, and terminal management platforms. Daily documents include asset ledgers, inventory sheets, asset issuance and return forms, and scrap applications. Data update frequency is monthly, quarterly, or annually. Document fields include asset number, name, specification/model, user, storage location, purchase date, status, and others. Units are mostly physical units such as units, sets, and pieces. The core focus is on asset ownership, location, and user.
What constraints do these characteristics impose on the re-ranking and hybrid retrieval link
This category has low data update frequency, so high-frequency incremental indexing is not required. A fixed-cycle full-update retrieval configuration can be used. Most documents have structured attributes. Hybrid retrieval needs to balance semantic matching and precise field matching. Asset number, user, and storage location are high-frequency query dimensions. The re-ranking link should prioritize associating these high-value fields. Each document has clear, focused core asset attributes. The field completeness of retrieval results must be controlled to avoid irrelevant information interfering with fast queries in operations duty scenarios.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.7–0.85 | Most documents in this category have structured asset attributes. This setting balances semantic matching and precise recall, avoiding introducing irrelevant results if set too low, and missing valid asset information if set too high |
limit | Top 10–15 results | Operations duty scenarios require fast location of asset information. Too many results increase screening costs, while too few fail to cover all relevant assets |
searchMode | Hybrid Retrieval | This category needs to match both asset semantic descriptions and precise fields such as asset number. Hybrid retrieval balances recall breadth and matching accuracy |
usingReRank | Enabled | There are many similar fields in asset information such as devices of the same model. Re-ranking sorts based on semantic relevance, prioritizing results that best match the query intent |
| Private re-ranking model configuration | Calibrated based on actual testing | This category has low data update frequency. Private models can be deployed as needed, avoiding delays from relying on external interfaces |
Three common misconfigurations
- Symptom: Retrieval results are empty or reference fields are empty after enabling
usingReRank. Cause: The deployment path and environment variables of the private re-ranking model are not configured correctly, or thesimilaritythreshold is set too high, resulting in no valid content after the original retrieval results are filtered by re-ranking. - Symptom:
searchModeis set to pure semantic retrieval mode, and content related to precise fields such as asset number and user cannot be recalled. Cause: The characteristics of structured data in this category are not considered, and hybrid retrieval mode is not enabled. Only relying on semantic matching cannot cover precise query requirements. - Symptom:
limitis set to more than top 20 results, leading to retrieval timeout or excessive computing resource usage in the re-ranking link. Cause: The low-latency requirements of operations duty scenarios are not considered, and the number of initial retrieval results is not controlled, increasing the computing burden on the re-ranking link.
How to confirm the configuration is correct
- Execute test queries that include fields such as asset number, user, and storage location, and check whether the retrieval results contain valid content for the corresponding fields.
- View retrieval logs to confirm that
usingReRankis enabled and the re-ranking model call status is normal. - Adjust the
similaritythreshold and compare the number of retrieval results under different thresholds to confirm that recall rate and relevance are balanced. - Test the private re-ranking model call flow to confirm there are no interface errors or timeout issues.
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.