What the Data for This Category Looks Like
The core data for this category comes from search ad backends, search term reports, bidding ranking data, and landing page data. Most updates occur on a daily or weekly basis. Commonly used documents include structured keyword lists, negative keyword tables, bid records, and search term reports. Fields include keyword text, matching mode, bid value, search volume, conversion-related data, and more. Data is mostly tabular and structured, with common export formats of CSV or Excel.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Multi-source heterogeneous ad data sources require retrieval to support cross-format, cross-field fused recall. This prevents missing relevant entries due to data format differences. Daily or weekly update frequencies require the knowledge base to be configured with scheduled incremental sync tasks. This ensures data timeliness. Structured documents such as keywords and negative keyword tables require retrieval to support both exact matching and semantic expansion. This balances precise rule-based queries and flexible adaptation to user questions. The most time-consuming keyword expansion and negative keyword screening steps require limiting the number of recall results to a reasonable range. This avoids exceeding the model's context processing capacity. It also requires prioritizing entries strongly related to the user's technical support questions, to prevent invalid results from interfering with decision-making.
Configuration Recommendations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75-0.85 | Balances semantic relevance and recall coverage for search ad keywords. Avoids filtering valid expanded keywords with too high a threshold, and avoids introducing irrelevant results with too low a threshold. |
limit | Top 10-15 entries | Adapts to business scenarios of negative keyword screening and keyword expansion. Controls the number of returned results to avoid exceeding the model context window. |
searchMode | Semantic + exact hybrid | Supports exact matching of core keywords and negative keyword table entries, as well as semantic expansion of relevant search term report data. Covers different query needs for pre-sales and technical support. |
usingReRank | Enabled | Re-ranks multi-source recall results. Prioritizes entries with higher matching degrees to user questions. For example, displays corresponding negative keyword table content first when a user asks about negative keyword rules. |
datasetSearchUsingExtensionQuery | Enabled | Performs query expansion on users' colloquial technical support questions. Covers more implicitly relevant keyword documents. Adapts to unstructured user queries. |
collectionIds | Bind dedicated search ad keyword management knowledge base collection ID | Limits retrieval scope, avoids interference from cross-category documents, and improves retrieval accuracy. |
Three Common Mistakes
- Issue: An abnormal billing prompt is returned when calling the knowledge base retrieval interface. Cause: A custom OpenAPI token was not used. Using the platform's default token leads to a mismatch in billing rules.
- Issue: The retrieval results only return a small number of structured keyword entries, and do not cover expanded search term-related content. Cause: The
similaritythreshold was set too high, filtering semantically relevant expanded results. - Issue: An error "No available channel for model gpt-4o-mini under current group default" pops up when launching a QA search test. Cause: No available channel for gpt-4o-mini was configured under the corresponding group, or the group was bound incorrectly.
How to Verify Successful Configuration
- Upload test keyword lists, negative keyword tables, and search term reports to the corresponding knowledge base collection. Launch test queries including "How to screen negative keywords" and "What are the expanded keywords". Verify that returned results include matching document entries.
- Check the FastGPT v4.17.0 interface call logs. Confirm that the retrieval request uses a custom OpenAPI token, and that there are no abnormal billing-related log entries.
- Review the
collectionIdsparameter in the configuration. Confirm that it is bound to the dedicated search ad keyword management knowledge base collection, to avoid retrieving documents from unrelated categories. - Adjust the
similaritythreshold to 0.7 and 0.8. Verify that the number of recall results changes as expected, confirming that the threshold configuration takes effect.
The values provided are common starting points, and should be measured against sample data from the relevant 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.