What the data for this category looks like
Daily documents for this category include delivery plans, targeting settings, budget scheduling, conversion tracking configurations, and more. Data sources include advertising platform backends, user profile databases, budget management spreadsheets, and attribution analysis tools. Each individual document typically contains fields such as unique plan identifiers, targeting audience package parameters, daily budget amounts, conversion event configuration items, and others. Units involved include yuan, thousand impressions, conversion event names, and more. Most data updates occur via daily synchronization or are triggered by delivery activity changes.
What constraints do these characteristics impose on the "knowledge base retrieval and recall" link?
The scattered nature of data sources requires the retrieval link to support associated recall across multiple datasets, rather than pulling information only from a single advertising platform. The large number of document fields with dedicated units requires retrieval to match field precision, to avoid recalling irrelevant configuration items. The high-frequency update characteristic requires regular refreshing of dataset indexes, to ensure recalled content is synchronized with the latest delivery configurations. The accuracy of delivery data directly impacts business outcomes, so relevance validation for recalled results must meet the strict standards of the business scenario. Additionally, the structured nature of individual documents requires retaining field integrity during chunking, to avoid splitting critical configuration information.
How to configure the settings
Under FastGPT v4.17.0, the following configuration items can meet the knowledge base retrieval needs for the delivery plan setup scenario:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the field precision requirements for delivery configurations, to avoid recalling irrelevant low-relevance content |
limit | Top 8 entries | Adapts to the number of fields in a single delivery document, to avoid context overload that affects subsequent processing |
searchMode | vector_search | Prioritizes matching semantic and field characteristics for vector retrieval needs of structured delivery data |
usingReRank | Enabled | Performs secondary sorting on recalled results, to optimize the display priority of results related to delivery configurations |
datasetSearchUsingExtensionQuery | Enabled | Expands the retrieval scope to cover associated configuration data across advertising platforms and budget spreadsheets |
collectionIds | Bind the IDs of corresponding datasets such as delivery plans and targeting settings | Precise limitation of the retrieval scope to avoid interference from irrelevant datasets |
Three common mistakes
- A log error
try reducing the size of the batcappears during knowledge base creation. This occurs because chunking parameters are not adjusted for structured delivery documents, and the total length of batched processed documents exceeds the system default threshold. - After retrieval results are passed to an HTTP workflow, AI conversations cannot correctly associate knowledge base content. This occurs because the returned retrieval results do not carry structured field identifiers and source information, only retaining plain text fragments.
- The total token count of retrieved recalled content exceeds the context limit, causing subsequent AI processing to fail. This occurs because the number of recalled entries from the
limitparameter is not controlled, and chunking and recall quantities are not adjusted based on the average length of delivery documents.
How to confirm the configuration is correct
- Manually trigger a single retrieval, and check whether the fields of the returned results match those of the target delivery configuration document.
- View the retrieval logs, and confirm that the number of recalled results falls within the range set by the
limitparameter. - Check the format of retrieval results passed in the workflow, and confirm that metadata and original content fields are included.
- After simulating a delivery configuration change, trigger a dataset refresh, and verify whether the retrieval results are synchronized to the latest content.
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.