What the Data for This Category Looks Like
Data sources for this category include software asset management systems, procurement systems, terminal scan files, and vendor authorization documents. Core documents for daily maintenance include license ledgers, software usage lists, procurement contracts, and compliance audit reports. Data is updated monthly or quarterly. Fields include authorization ID, installed node count, authorization validity period, vendor name, and more. Units are mostly sets, devices, and days. Document structures include both structured tables and unstructured text. A single data entry often links to multiple terminal installation nodes.
What Constraints These Characteristics Impose on the "Re-Ranking and Hybrid Retrieval" Link
Decentralized data sources that include both structured and unstructured content require hybrid retrieval to support both vector semantic matching and precise keyword matching, to avoid missing compliance information in structured fields. Longer update cycles mean retrieval must prioritize recently updated authorization ledger data. The re-ranking recall range must be limited to the current valid dataset, to avoid including expired authorization entries. Most fields are precise IT asset attributes. Re-ranking must prioritize the relevance of fields such as authorization ID and expiration date, while also accounting for general semantic similarity. A single data entry links to multiple terminal nodes. Retrieval results must be sorted by authorization compliance, to prevent over-installed or expired entries from appearing first.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
searchMode | Hybrid Retrieval (Semantic + Keyword) | Adapts to the characteristics of license data that includes both structured ledgers and unstructured contracts, balancing semantic understanding and precise keyword matching |
usingReRank | Enabled | License data has high compliance requirements. Re-ranking filters low-match over-installed or expired results, improving retrieval accuracy |
similarity | 0.75–0.85 | Balances retrieval recall rate and accuracy, avoiding missing compliance-related results or introducing irrelevant entries |
limit | Top 10 entries | Adapts to the information volume of a single ticket during operations duty, preventing information overload for on-duty personnel |
Three Common Mistakes
- Issue:
usingReRankis enabled but has no effect, and retrieval results are not sorted by compliance priority. Cause: The re-ranking model node is not bound in the workflow, or the association rules between the knowledge base and the re-ranking link are not configured. - Issue: When multiple variables are output after an HTTP request, a single variable cannot be bound to the software license knowledge base. Cause: Variable mapping rules are not configured in the re-ranking link, and the knowledge base subset related only to software authorizations is not specified.
- Issue: The re-ranking model image fails to load when deployed on an ARM architecture. Cause: A re-ranking model image adapted to ARM architecture is not used; the default image only supports AMD architecture.
How to Confirm Proper Configuration
- Navigate to the knowledge base configuration page of FastGPT v4.17.0, check if
searchModeis set to hybrid retrieval and ifusingReRankis enabled. - Submit a test ticket, enter "Query currently over-installed software authorizations", and verify that retrieval results are sorted by authorization compliance, with priority given to field relevance.
- View workflow run logs to confirm that the re-ranking node was called successfully, with no model loading failures or timeout errors.
- Switch the retrieval mode to pure semantic retrieval, compare the differences between the two retrieval results, and confirm that hybrid retrieval covers both keyword matching and semantic matching results.
The values provided here are common starting points, and should be evaluated against the samples specific to the deployment environment.
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.