What this category of data looks like
Typical users of this category include ad optimizers, growth analysts, and agency operations staff. Daily produced and used documents include ad delivery daily and weekly reports, real-time price adjustment records, material A/B test data tables, and delivery effect optimization review documents. Data sources include ad platform backend data, attribution system conversion data, CRM customer lead data, and landing page visit data. Most updates occur via daily synchronization or real-time push. Document fields include delivery-related metrics such as impressions, clicks, conversions, and customer acquisition cost. Common units include counts, yuan, and percentages.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The data characteristics of this category impose multiple constraints on the knowledge base retrieval and recall link. Most data is updated in real time or daily, so the knowledge base must support incremental synchronization to ensure retrieval results match the latest delivery data. Daily documents include structured data tables and unstructured review text. The retrieval system must balance precise field matching and semantic association recall to cover business documents in different formats. Data from multiple sources has inconsistent field definitions, so the retrieval link must support basic field alignment processing to avoid confusion from cross-system data. The core scenarios for delivery effect optimization include abnormal metric positioning and budget allocation. The retrieval system must return highly relevant results quickly to support real-time decision-making. Incorrect retrieval results may directly increase customer acquisition costs or exhaust budgets ahead of schedule, which raises higher requirements for recall accuracy.
How to set configurations
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the basic relevance requirements of delivery business for retrieval results, filters low-value historical data |
limit | Top 8–12 entries | Balances retrieval response speed and result information volume, adapts to the quick viewing scenario for abnormal metric positioning |
searchMode | hybrid | Balances semantic retrieval and keyword retrieval, covers structured data tables and unstructured review documents |
usingReRank | Enabled | Performs secondary ranking on initial recall results, improves display priority of highly relevant documents |
datasetSearchUsingExtensionQuery | Enabled | Generates extended query terms based on the user's original question, adapts to fuzzy query needs in delivery scenarios |
collectionIds | Configured by delivery business domain grouping | Limits the retrieval scope to only knowledge base collections related to delivery effect optimization, avoids interference from irrelevant documents |
Three common mistakes
- Symptom: Unable to obtain a valid knowledge base ID when calling the knowledge base node in a workflow. Cause: Failed to correctly create the knowledge base in the FastGPT backend and copy its unique identifier, or failed to authorize the knowledge base for use by the corresponding workflow.
- Symptom: No search results returned after passing the
collectionIdsparameter. Cause: The configured knowledge base collection has not synchronized the latest documents for the delivery business, or the knowledge base ID entered in the parameter does not match the actual deployed collection ID. - Symptom: A Query read timeout error is returned when calling the knowledge base retrieval interface. Cause: The timeout configuration was not adjusted for the high-frequency retrieval scenario of the delivery business, causing a large number of concurrent retrieval requests to exceed the system response threshold.
How to verify correct configuration
- Submit a test query containing delivery business keywords, check the relevance of returned results, and adjust the similarity threshold to a range that matches business needs.
- Pass the specified
collectionIdsparameter to initiate a retrieval, verify that returned results only include documents from the corresponding business domain, with no irrelevant content included. - After enabling the
usingReRankconfiguration, compare the order of initial recall results and reranked results, confirm that highly relevant documents are displayed first. - Simulate a single retrieval request, check if the interface response meets business real-time requirements, adjust the retrieval limit parameter to balance speed and result information volume.
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.