What the data for this category looks like
This category’s data primarily comes from search ad backends, search term reports, bidding ranking data, and landing page data. Most updates follow a daily or weekly schedule. Daily processed documents include keyword lists, negative keyword tables, bid records, and search term reports. Core fields include keyword text, match type, single bid, and historical search volume. Some fields have defined units: bids use yuan, search volume uses counts. Data is stored as structured tables with clearly segmented fields. It covers all core information across the full keyword deployment lifecycle.
How Data Characteristics Impact Reranking and Hybrid Retrieval
This category’s data has high structural integrity and clearly segmented fields. Hybrid retrieval must cover multiple dimensional fields including keyword text, match type, and bid. This avoids missing key related information during single-dimensional retrieval. The daily or weekly update frequency means retrieval indexes must sync on the corresponding cycle. This ensures the timeliness of retrieval results. Daily processed documents are mostly bulk keyword-related data. The reranking step must first match keywords mentioned in tickets against existing stored data. It must also adjust sorting weights using historical search volume. This meets the need for rapid anomaly location during operations duty scenarios.
How to Set Configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
usingReRank | Enabled | Adapts to the need for quickly filtering highly relevant tickets and keyword data during operations duty scenarios |
similarity | 0.65–0.75 | Matching text similarity threshold for search ad keywords. Balances recall precision and coverage |
limit | Top 10 entries | Limits the number of returned results. Adapts to operations scenarios where users quickly browse core information |
searchMode | Multi-field hybrid retrieval | Covers multiple dimensional fields including keyword text, match type, and bid. Aligns with the data characteristics of this category |
| Rerank model access address | Privately deployed bge-rerank-base image address | Meets private access requirements for FastGPT v4.17.0 |
Three Common Configuration Errors
- Symptom: Rerank model calls return 401 unauthorized errors, while configurations on other platforms run normally. Cause: Access permissions for the privately deployed rerank model are not correctly configured. This includes failing to fill in a valid API key or failing to add the FastGPT service IP to the allowlist.
- Symptom: When reranking is not enabled, the relevance of retrieval results cannot adapt to the search ad keyword management business scenario. Cause: Basic hybrid retrieval only relies on vector similarity sorting. It does not adjust sorting logic using business fields such as keyword match type and historical search volume. It cannot accurately match the keyword requirements in operations tickets.
- Symptom: The number of retrieval results after configuration does not match the set
limitparameter. Cause: The recall upper limit for hybrid retrieval and the return upper limit after reranking are not correctly set. This causes the result count to exceed expectations.
How to Verify Successful Configuration
- Navigate to the FastGPT model test page. Enter ticket keywords related to search ads. Verify that the rerank model returns sorted results correctly.
- View retrieval logs. Confirm that the
similaritythreshold takes effect. It filters out low-relevance results below the set value. - Check index sync status. Confirm that keyword data index updates complete on a daily or weekly cycle.
- Verify the number of returned results for the
limitparameter. Confirm it matches the set value.
Final Note
The values provided here are common starting points. They should be evaluated against the reader’s own sample data.
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.